Merge pull request #1858 from clonea1/contrib/gate-experiment-findings

docs: measured comparison of three CSI frame-selection gates, and the harness
Merge upstream/main, resolve README.md table-row conflict
2026-10-20 05:40:14 -05:00 · 2026-09-09 13:10:02 +00:00 · 2026-09-09 08:40:14 -04:00 · 2026-09-09 12:38:26 +00:00 · 2026-09-09 12:38:23 +00:00 · 2026-09-09 12:38:18 +00:00
45 changed files with 7130 additions and 351 deletions
@@ -98,6 +98,8 @@ jobs:
        run: |
          make test_adr110
          ./test_adr110
+          make test_c6_antenna
+          ./test_c6_antenna

      - name: Verify binary size budget
        working-directory: firmware/esp32-csi-node
@@ -184,3 +186,22 @@ jobs:
          name: esp32-csi-node-firmware-${{ matrix.variant }}
          path: firmware/esp32-csi-node/release-staging/
          retention-days: 90
+
+  build-c6-idf6:
+    name: Build XIAO C6 external antenna (ESP-IDF 6.0.2)
+    runs-on: ubuntu-latest
+    container:
+      image: espressif/idf:v6.0.2
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262
+        with:
+          submodules: recursive
+
+      - name: Build the IDF 6 canary image
+        working-directory: firmware/esp32-csi-node
+        run: |
+          . $IDF_PATH/export.sh
+          idf.py -D SDKCONFIG_DEFAULTS="sdkconfig.defaults;sdkconfig.defaults.esp32c6;sdkconfig.defaults.xiao-external" set-target esp32c6
+          idf.py build
+          test "$(stat -c%s build/esp32-csi-node.bin)" -le $((1152 * 1024))
+          strings build/esp32-csi-node.bin | grep -F "external U.FL"
@@ -17,11 +17,16 @@ firmware/esp32-csi-node/sdkconfig.defaults.bak
 firmware/esp32-hello-world/sdkconfig.old

 # Host-built firmware test binaries (compiled from test/*.c, not source)
-firmware/esp32-csi-node/test/test_adr110
-firmware/esp32-csi-node/test/test_vitals
-firmware/esp32-csi-node/test/fuzz_serialize
-firmware/esp32-csi-node/test/fuzz_edge
-firmware/esp32-csi-node/test/fuzz_nvs
+# Compiled host-test and fuzz binaries. The Makefile writes them beside their
+# sources, so they land in a tracked directory. Ignore them by SHAPE rather
+# than by name: every source here carries an extension and no binary does.
+#
+# This replaced a hand-maintained list of five targets, which had gone stale --
+# test_csi_sanitize, test_mmwave_detect and test_thermal were all built by
+# `make host_tests`, ignored by nothing, and got committed by accident.
+firmware/esp32-csi-node/test/test_*
+firmware/esp32-csi-node/test/fuzz_*
+!firmware/esp32-csi-node/test/*.*
 firmware/esp32-csi-node/test/*.exe
 firmware/esp32-csi-node/test/*.obj

@@ -719,6 +719,8 @@ Start with the user, build, and calibration guides; expand for the full referenc
 | [Semantic Primitives — Precision/Recall](docs/integrations/semantic-primitives-metrics.md) | Per-primitive F1 on the held-out paired-capture set: someone-sleeping, possible-distress, room-active, elderly-inactivity-anomaly, meeting, bathroom, fall-risk, bed-exit, no-movement, multi-room. |
 | [Claude Code / Codex Plugin](plugins/ruview/README.md) | The `ruview` plugin + marketplace — skills, `/ruview-*` commands, agents, and the Codex prompt mirror |
 | [Portable harness — `npx @ruvnet/ruview`](harness/ruview/README.md) | MetaHarness-minted, host-portable RuView operator harness — `ruview.*` MCP tools + the MEASURED-vs-CLAIMED honesty guardrail enforced in code ([ADR-182](docs/adr/ADR-182-npx-ruview-harness-via-metaharness.md)). A lighter, multi-host companion to the in-repo plugin. |
+| [CSI Frame Selection — Measured](docs/csi-frame-selection-measured.md) | Which rate gate actually buys cross-node pairing, measured on a nine-node ESP32-C6 fleet as a 2x2 factorial in three Latin-square blocks. Selecting frames by 802.11 `rx_seq` roughly doubled pairing (+31.7 pp) while mesh-time alignment bought +0.5 pp, inside noise. Includes the trade it makes, a prediction that turned out wrong, the limits of the measurement, and `run_arms.py` so the comparison can be re-run rather than taken on trust. |
+| [ESP32-C6 Firmware Runbook](firmware/esp32-csi-node/RUNBOOK.md) | What to run, in what order, and what to check: the three-file build that keeps OTA rollback compiled in, a five-key pre-flash gate, which flash regions OTA can and cannot reach, a single-board diagnostic flash with a revert path, and the tunables whose right value depends on your house rather than on the chip. |
 | [Architecture Decisions](docs/adr/README.md) | 205 ADRs — why each technical choice was made, organized by domain (hardware, signal processing, ML, platform, infrastructure) |
 | [Domain Models](docs/ddd/README.md) | 8 DDD models (RuvSense, Signal Processing, Training Pipeline, Hardware Platform, Sensing Server, WiFi-Mat, CHCI, rvCSI) — bounded contexts, aggregates, domain events, and ubiquitous language |
 | [rvCSI — edge RF sensing runtime](https://github.com/ruvnet/rvcsi) | Rust-first / TypeScript-accessible / hardware-abstracted CSI runtime: multi-source ingestion (incl. real nexmon_csi `.pcap` from a **Raspberry Pi 5** / Pi 4 / Pi 3B+ — CYW43455 / BCM43455c0) → validation → DSP → typed events → RuVector RF memory ([ADR-095](docs/adr/ADR-095-rvcsi-edge-rf-sensing-platform.md), [ADR-096](docs/adr/ADR-096-rvcsi-ffi-crate-layout.md), [domain model](docs/ddd/rvcsi-domain-model.md)). Now its own repo — [`ruvnet/rvcsi`](https://github.com/ruvnet/rvcsi) — vendored here under `vendor/rvcsi`; 9 `rvcsi-*` crates on crates.io, `@ruv/rvcsi` on npm, plus a Claude Code plugin. |
@@ -0,0 +1,193 @@
+# ADR-345: Per-link CSI attribution and node-to-node ranging
+
+- **Status**: Proposed — measurement layer implemented and validated on hardware; localization not built
+- **Date**: 2026-08-28
+- **Deciders**: Joe
+- **Tags**: csi, wire-format, esp32-c6, localization, adr-018, adr-060, adr-110
+
+## Context
+
+Every attempt at device-free position estimation in this repository has failed
+against the same wall, and the wall has now been measured twice from different
+directions.
+
+**Phase is unusable.** The bistatic tier needed a signed per-link radial
+velocity, which requires the common-mode component of CSI phase. Measured
+2026-08-28 across all three ESP32-C6 nodes over ~40,000 frames, the
+frame-to-frame change in common-mode phase is statistically indistinguishable
+from uniform random — resultant length `R` = 0.011–0.032 against a 0.000
+reference, standard deviation 1.78–1.81 against `pi/sqrt(3)` = 1.814, and no
+lag-1 autocorrelation. A second, unwrap-free estimator agreed. The mechanism is
+packet-detection timing quantization: common-mode phase is `-2*pi*f_c*tau`, and
+one ADC sample of jitter at 20 MHz is `2*pi * 2.412e9 * 50e-9` ~= 758 rad, about
+120 full wraps, re-randomized every packet. Not calibratable. The bistatic tier
+is default-off as a result (`bistatic_tier_enabled`).
+
+**Amplitude works but had no parallax.** The three AP -> node links leave a
+single distant transmitter within a 36.5 degree fan and converge on the node
+cluster. Bearings computed from `room_config.json`: 198.5, 221.2, 235.0 degrees.
+Scalar per-node energy weighting was closed with a clean negative result on
+2026-08-27 — magnitude responds to proximity, direction does not.
+
+**The receivers were hearing far more than we recorded.** A node in promiscuous
+MGMT+DATA mode with no `filter_mac` produces CSI for every transmitter on the
+channel. Measured on node 2: the AP accounted for only **34%** of frames; the
+two peer nodes supplied 39% and 27%. Those peer frames are valid measurements
+of *node-to-node* channels — but ADR-018's wire format carries no transmitter
+address, so the sink could not attribute them. They interleaved into one
+`frame_history` with mixed geometry, and the only available remedy
+(`filter_mac`, ADR-060) fixes the mixing by discarding ~75% of the data.
+
+The node-to-node links between the same three boards cross the room from three
+sides at **108.7 degrees** of angular spread (bearings 0.0, 73.2, 108.7) versus
+36.5 for the AP links — roughly 3x the parallax, with both endpoints known,
+fixed, and under our control rather than wherever the router happens to live.
+Link count also grows as `N*(N-1)/2`: 3 nodes give 3 links, 6 give 15, 13 give
+78.
+
+## Decision
+
+Carry the transmitter MAC on the CSI wire and model CSI per **link**
+(receiver, transmitter) rather than per **node**, as a layer alongside the
+existing per-node model.
+
+### 1. Wire format v2
+
+`CSI_MAGIC_V2 = 0xC5110008`, header 26 bytes: the v1 header unchanged through
+byte 19, then the transmitter address (`info->mac`, 802.11 addr2) at bytes
+20..25, with I/Q from byte 26.
+
+`0xC5110002`..`0xC5110007` were already claimed by the vitals, feature and other
+edge packets, so v2 takes the next free value rather than an adjacent one — see
+`issue_928_magic_collision_tests` for why that matters. The parser accepts both
+magics, so a mixed fleet keeps working and v1 nodes simply report no
+transmitter.
+
+### 2. Per-link state (`links.rs`)
+
+`LinkTable` keyed by `LinkId { rx_node, tx_mac }`, each link holding its own
+amplitude history, RSSI EMA, learned resting baseline and motion metric.
+
+The metric is the mean over subcarriers of the temporal standard deviation of
+**AGC-normalized** amplitude. Removing each frame's own mean across subcarriers
+is essential rather than cosmetic: the ESP32 applies automatic gain control per
+packet, so absolute amplitude jumps between packets for reasons unrelated to the
+channel, and any statistic built on the raw level measures the radio's gain
+control instead of the room.
+
+Amplitude only, deliberately — nothing here can depend on phase coherence.
+
+### 3. Beacon rate
+
+`BEACON_PERIOD_MS` 100 -> 20. At 10 Hz a peer samples a node-to-node link with a
+5 Hz Nyquist while motion Doppler lives at 1–20 Hz; it aliases before it starts.
+20 ms matches both the receiver's own `CSI_MIN_PROCESS_INTERVAL_US` ceiling and
+the gateway self-ping already used for the AP link.
+
+### 4. `GET /api/v1/links` and a live UI panel
+
+The link table was initially reachable only through a 30-second log line, which
+is the wrong medium for a quantity whose value is watching it respond while you
+walk across the room. `GET /api/v1/links` returns one row per link — `motion`,
+`raw_motion`, `rssi_dbm`, `frames`, `kind` — and `LinkMeshPanel` renders them in
+the Sensing tab at 1 Hz. `raw_motion` is shown beside `motion` deliberately: a
+link whose baseline has not settled reports a healthy raw value and a near-zero
+motion, and with only the one number that is indistinguishable from a dead link.
+
+Node identity is **inferred, not measured**. Nothing in the system records a
+node's own MAC — a node reports the addresses it hears, never its own. But a
+radio does not receive its own transmissions, so given the set of receivers
+currently reporting, a transmitter heard by all of them except exactly one is
+that one receiver's own board; a transmitter heard by every receiver is external
+infrastructure. `links::infer_transmitting_node` declines (returns `None`) with
+fewer than three receivers, where "heard by all but one" and "heard by exactly
+one" are the same observation, and whenever more than one receiver is missing,
+which a peer out of range produces indistinguishably. The field is named
+`tx_node_inferred` and the UI draws the peer id with a trailing `?`.
+
+Checked against the known MAC map on 2026-08-28 the inference labelled all four
+transmitters correctly — nodes 0, 1 and 2 by address, and the AP as external —
+with no access to that map. A wrong answer here mislabels a display row; it
+never reaches a metric.
+
+## Consequences
+
+### Validated on hardware, 2026-08-28
+
+All three nodes on wire v2 at 20 ms produced the full mesh: **9 links** (3 AP +
+6 directional node-to-node).
+
+Reciprocity — each physical channel measured independently from both ends:
+
+| channel | direction A | direction B | ratio |
+|---|---|---|---|
+| 0 <-> 1 | 2.229 | 1.930 | 1.15 |
+| 0 <-> 2 | 1.353 | 1.000 | 1.35 |
+| 1 <-> 2 | 1.398 | 1.486 | 1.06 |
+
+Nothing in the implementation enforces this. Three independent radio pairs
+agreeing within 6–35% is the strongest available evidence that the per-link
+separation is real rather than an artifact of frame bucketing.
+
+Link quality is also far more uniform than the AP links: node-to-node raw motion
+spans 1.0–2.2 across all six, versus 3.8–17.8 for the three AP links (which
+track RSSI: -67, -72, -80). Uniform link quality is desirable for tomographic
+localization — no single link dominates the solution.
+
+### Known limitations
+
+- **Peer links are ~10x weaker.** A 16-byte ESP-NOW beacon offers far less
+  energy to estimate a channel response from than a full data frame. Whether
+  1.0–2.2 perturbation is sufficient to localize with is **unknown and untested**.
+- **Three links is below threshold.** Tomographic localization needs many
+  crossing links; 3 will not localize regardless of quality. This ADR delivers
+  the measurement layer, not localization.
+- **Localization is not implemented.** No position tier consumes `LinkTable`.
+  The live position estimate remains `doppler_centroid`.
+- **Per-node path regression.** Removing `filter_mac` to admit peer frames
+  returned the *per-node* `frame_history` to a mixed-transmitter stream, since
+  it still consumes every frame regardless of source. Node 2's classification
+  signal is noisier as a result. The fix is to select a link to feed the
+  per-node path — software filtering that keeps the other links — and it is not
+  yet done.
+
+### Scaling limits to settle before more nodes
+
+- **Airtime.** Every node broadcasts, so N nodes put N/period frames per second
+  on the channel being sensed. 3 nodes at 20 ms is 150 pps; 13 would be 650 pps.
+  `BEACON_PERIOD_MS` should become an NVS setting scaled to fleet size.
+- **`MAX_LINKS` is 64.** 13 nodes give 78 receiver-transmitter pairs plus the AP
+  plus neighbours, which overflows. Raise deliberately, with the memory implied.
+- **Placement.** Links must *cross* the target area. Nodes spread through a
+  house give roughly one link per room, each attenuated by walls and perturbed
+  by anyone anywhere along it — whole-house coarse presence, not position.
+  Perimeter placement around a single room is the geometry that works.
+
+## Related bugs found and fixed while implementing
+
+- **ADR-060 MAC filtering cost ~75% of capture rate.** The 50 Hz rate gate ran
+  *before* the MAC filter and stamped `s_last_process_us` before the check, so a
+  frame from any other transmitter consumed the slot and was then discarded.
+  Enabling `filter_mac` dropped yield from ~42 pps to 6–13. Filtering first
+  restores it (measured 31–37 pps). A 6-byte `memcmp` is negligible ISR work next
+  to the CSI processing the gate guards.
+- **The entire edge pipeline was dead on ESP32-C6.** `EDGE_MAX_SUBCARRIERS` was
+  128, an S3-era assumption; a C6 on an HE-capable AP delivers HE20 frames with
+  256 bins, and `process_frame` rejected every one. ADR-110 onboarded the C6 to
+  the capture path but not this one. Vitals, presence, fall detection and
+  per-slot counting had never run on these boards. Now target-conditional on
+  `CONFIG_SOC_WIFI_HE_SUPPORT`, with the four subcarrier-sized arrays hoisted to
+  static so peak stack is *lower* than it was at 128.
+- **A target-conditional constant that failed silently.** `edge_processing.h`
+  had no `#include` at all, so `CONFIG_SOC_WIFI_HE_SUPPORT` was undefined and
+  `#if` evaluated it as 0 — selecting the wrong size with no warning. An
+  undefined identifier in `#if` is 0 in C. This is the same failure mode as the
+  bug it was fixing, one layer up.
+
+## References
+
+- `v2/crates/wifi-densepose-sensing-server/src/links.rs` — per-link state, 8 tests
+- `v2/crates/wifi-densepose-sensing-server/src/phase_diag.rs` — the phase measurement
+- `firmware/esp32-csi-node/main/csi_collector.c` — wire v2 serialization, filter ordering
+- `firmware/esp32-csi-node/main/edge_processing.h` — `EDGE_MAX_SUBCARRIERS`
+- ADR-018 (wire format), ADR-060 (MAC filtering), ADR-110 (C6 onboarding)
@@ -0,0 +1,134 @@
+# ADR 355: Local empty room bootstrap prior
+
+Status: Accepted
+
+Date: 2026-09-04
+
+## Context
+
+Most installations are occupied when RuView first starts. An uncalibrated
+signal heuristic can interpret HVAC, clock error, neighboring motion, or a
+short covariance window as a person. It can also surface plausible numeric
+vitals before a room model establishes occupant attribution.
+
+One empty home is not representative training data for all homes and contains
+no positive human examples. It is useful only as a recent negative prior for
+the same installation. It cannot define a universal human signal model.
+
+The field model was memory only. The server also used a relative `data/`
+directory, which is unreliable for embedded application hosts.
+
+## Decision
+
+RuView may persist a completed empty room field model as a local bootstrap
+prior only after two server controlled stages:
+
+1. Normal calibration reaches both 1,000 accepted frames and 600 seconds.
+   The frame target represents twenty complete fifty frame runtime windows.
+   A caller may select one currently live ESP32 source. The server checks its
+   freshness before replacing an active bootstrap prior, binds collection to
+   that source, and rejects frames from other radios. When no source is
+   selected, the first accepted ESP32 source preserves legacy behavior.
+2. Promotion observes 12 new samples spaced by at least one second. Each
+   sample has a four second bounded acquisition timeout. At least 10 must
+   resolve empty, all ticks must be fresh, and none may publish numeric vital
+   signs.
+
+The stored image contains aggregate means, environmental modes, mode energies,
+noise statistics, sorted scalar residual references, configuration, one source
+node ID, and a server generated model ID. It excludes raw CSI, waveforms,
+device addresses, room names, pose, identity, credentials, and positive human
+labels. Runtime matching compares a fifty frame mean with the learned residual
+distribution. Its score is empirical background conformance, not an absence
+probability.
+
+The image is bound to a SHA 256 digest of the stable installation ID, carries
+an integrity digest over the exact stored payload bytes, is capped at 1 MiB,
+and inherits the field model expiry.
+Files with an invalid schema, dimensions, values, node set, installation
+binding, digest, size, or lifetime fail closed.
+
+Restoring an image creates a field model in `bootstrap_only` mode. It may
+reduce an uncalibrated false person count, but it cannot authorize calibrated
+evidence or numeric heart and breathing rates. Explicit calibration replaces
+it with fresh statistics. Cancelling an unfinished capture restores a valid
+prior. Administrator scoped reset removes it.
+
+A restored image may receive one operator confirmed held out empty refinement.
+The server measures 12 fresh, source bound runtime residuals at one second
+spacing and requires zero stale samples and zero numeric vital samples. It
+persists only the scalar residuals. The new boundary is the larger of the
+existing threshold and 110 percent of the held out 95th percentile, capped at
+125 percent of the prior threshold. The one use marker persists in the model,
+so repeated requests cannot ratchet the threshold upward.
+
+The status endpoint reports the learned runtime window, residual threshold,
+reference count, current background conformance, and the negative only
+authority boundary. A restored model never resumes collection and never
+persists raw calibration frames.
+
+The server accepts `--data-dir` or `RUVIEW_DATA_DIR`, allowing an application
+host to provide a protected state directory.
+
+## Consequences
+
+Startup can use recent measured background structure instead of only generic
+heuristics. No accuracy lift is claimed until occupied and empty held out data
+from multiple installations is recorded.
+
+The largest failure mode is overfitting to a temporarily quiet home and
+suppressing a weak occupant. Controls are low authority, 24 hour expiry,
+installation binding, vital abstention, explicit recalibration, and reset.
+
+A human signal model still requires positive, consented, leakage separated
+occupied sequences. Empty data is never relabeled as human training.
+
+## Implementation
+
+1. `wifi-densepose-signal::ruvsense::field_model` exports and validates
+   `FieldModelSnapshotV1`.
+2. `wifi-densepose-sensing-server::bootstrap_baseline` owns schema, integrity,
+   storage, lifetime, and the held out gate.
+3. `POST /api/v1/calibration/bootstrap/promote` measures the validation window
+   and stores the image only on pass.
+4. `GET /api/v1/calibration/status` reports bootstrap authority and explicitly
+   denies calibrated evidence and numeric vital authority.
+5. Numeric vitals fail closed unless an explicit fresh calibration resolves
+   exactly one occupant at sufficient signal quality and confidence.
+6. The legacy `edge_vitals` WebSocket message applies the same publication
+   gate as `sensing_update`. When the bootstrap background matches, it emits
+   absent state, zero people, null numeric rates, and an explicit abstention
+   reason instead of forwarding raw edge candidates.
+7. `POST /api/v1/calibration/start?source_node_id=N` selects a live source for
+   deterministic single link capture. A missing or stale selection fails
+   without disabling the active bootstrap prior.
+8. `POST /api/v1/calibration/bootstrap/refine?confirmed_empty=true&source_node_id=N`
+   performs the bounded one time held out refinement. It requires administrator
+   authority when transport authentication is enabled.
+
+## Measured installation evidence
+
+One physical empty home run on source Node 5 collected 1,762 accepted frames
+over 655 seconds and produced 35 privacy reduced reference windows. Promotion
+observed 12 of 12 fresh empty samples, zero stale samples, and zero numeric
+vital samples. After process restart, 12 of 12 fresh samples matched the empty
+background and contained no numeric vital signs. Background conformance scores
+ranged from 50.0 to 64.3 percent with a median of 54.3 percent.
+
+This is measured negative evidence for one installation. No occupied positive
+holdout was recorded, so it does not prove person recall, person counting, or
+cross home accuracy improvement.
+
+## Acceptance test
+
+With a physically empty room, complete calibration on stable nodes, promote
+the single source model, and restart the server with the same installation ID
+and data directory. Status must report `binding_mode=bootstrap_only`, zero
+collection frames, and false authority for calibrated evidence and numeric
+vitals. Twelve fresh empty samples must contain no numeric vital signs.
+Starting a new room calibration must replace the prior and begin at frame zero.
+When several radios are live, selecting Node 5 must leave Nodes 3 and 7 out of
+the source set for the entire capture.
+After restart, a confirmed empty holdout may refine the restored boundary once.
+The stored snapshot must reject a second refinement and must continue to deny
+calibrated evidence and numeric vital authority.
@@ -0,0 +1,151 @@
+# ADR 356: Rate truthful CSI and reliability gated room adaptation
+
+Status: Accepted
+
+Date: 2026-09-05
+
+## Context
+
+RuView mixed three different notions of time. ESP32 firmware requested a packet
+rate, the server observed a delivered rate, and several server paths still used
+fixed 2 Hz, 10 Hz, or 20 Hz assumptions. One feature named
+`dominant_freq_hz` was not a temporal frequency at all. It was the index of the
+strongest subcarrier multiplied by 0.05. A second defect compared a newly
+inserted frame with itself, forcing that motion delta to zero.
+
+These errors can move a spectral peak into the breathing or heartbeat band and
+produce plausible but false values. They also make models sensitive to WiFi
+traffic and processor scheduling rather than physical motion.
+
+ESP-IDF documents that `wifi_csi_info_t.first_word_invalid` marks the first four
+CSI bytes invalid because of a hardware limitation. It also documents that the
+CSI receive callback runs in the WiFi task and should hand work to a lower
+priority task rather than perform lengthy processing. RuView serialized the
+invalid bytes unchanged and fed the original buffer to edge processing.
+
+Recent work suggests four useful directions with different maturity:
+
+1. OpenCSI represents each link as a dimensionless residual against its own
+   quiet period temporal standard deviation and exposes baseline maturity. Its
+   reported transfer results are limited to binary presence.
+2. MORIC derives motion centered delay and Doppler views to reduce dependence
+   on static environment specific CSI structure.
+3. Parameter efficient CSI crowd counting combines self supervised pretraining,
+   small room adapters, and a stateful counting machine.
+4. WiFi JEPA masks complete radio links and predicts their latent embeddings,
+   using real and ray traced CSI for pose representation learning.
+
+The papers report promising external results. None is measured RuView evidence,
+and none authorizes deployment without leakage free occupied and empty held out
+evaluation.
+
+## Decision
+
+### Firmware ingestion
+
+1. When ESP-IDF marks the first CSI word invalid, zero the first four serialized
+   CSI bytes while retaining the fixed packet shape.
+2. Set ADR-018 byte 19 bit 5 to record that sanitation occurred.
+3. Feed edge DSP from the sanitized serialized payload, never the original
+   invalid driver buffer.
+4. Keep callback work bounded to sanitation, serialization, and the existing
+   queue handoff. Do not add transforms to the WiFi task.
+
+### Server time model
+
+1. Estimate delivered CSI rate from accepted monotonic frame timestamps after
+   five samples. Use 20 Hz only during startup.
+2. Cap the physical input rate at the firmware contract of 50 Hz so a burst or
+   timestamp artifact cannot retune downstream filters above their supported
+   range.
+3. Reconfigure the vital detector only after a 20 percent rate change. Clear
+   temporal and smoothing state whenever its clock changes. Samples calculated
+   with different clocks must never be blended.
+4. Compute temporal dominant frequency from the recent mean amplitude series
+   with demeaning, a Hann window, and bounded direct Fourier bins from 0.1 Hz to
+   the smaller of 4 Hz or 45 percent of sample rate.
+5. Return zero for short, flat, nonfinite, or spectrally ambiguous windows.
+6. Compare the current frame with the prior frame, not itself, for motion delta.
+7. Export zero sample rate until the estimate is mature. Never present a startup
+   fallback as a measured node property.
+
+### Model compatibility and room reliability
+
+1. Preserve the historical fifth input feature for version 1 adaptive models at
+   the model boundary. Version 2 and later models consume the true temporal
+   frequency. This prevents an existing model from silently receiving a
+   different input distribution after a server update.
+2. A background model is unreliable until it contains at least ten finite,
+   nonnegative residual references. An unreliable model cannot suppress a
+   runtime change and exposes no conformance score.
+3. Expose background maturity, reliability, and a robust residual Z score based
+   on the median and median absolute deviation.
+4. Keep the persisted empty home prior negative only. It cannot authorize
+   calibrated presence, person count, pose, or numeric vitals.
+
+### Model roadmap
+
+The following work remains proposal only:
+
+1. Evaluate an OpenCSI style per link quiet residual normalization for binary
+   presence without replacing absolute magnitude features used for motion.
+2. Evaluate delay and Doppler views after phase sanitation for activity.
+3. Pretrain a frozen multilink encoder and adapt small room specific modules for
+   counting.
+4. Evaluate link masked self supervision for pose only when at least three
+   spatially distinct links and synchronized positive pose labels exist.
+
+Every candidate uses immutable manifests and splits by room, subject,
+recording, and time. It is compared with the frozen baseline under identical
+seeds. Counting includes confusion matrices and mean absolute error. Pose
+includes PCK at 20 centimetres, MPJPE, and the mean pose baseline. A shuffled
+label control must fail.
+
+Promotion requires at least 10 percent held out lift, less than 10 milliseconds
+added p95 inference latency, no more than 20 percent peak memory overhead, zero
+leakage or secret findings, frozen anchor retention, and no regression in empty
+rejection or occupied recall. Promotion remains an explicit maintainer action.
+
+## Consequences
+
+The firmware no longer preserves a known invalid prefix as sensor evidence. The
+server uses an observed physical clock for spectral features and vital filters.
+Existing version 1 adaptive models remain compatible. Sparse background samples
+fail open to observation and fail closed for suppression.
+
+The direct Fourier estimator is intentionally small and bounded to 128 temporal
+samples. It is portable and deterministic, but it is not a replacement for a
+phase sanitized delay and Doppler pipeline.
+
+The largest remaining uncertainty is human recall. Empty home evidence can
+measure false presence but cannot show whether weak occupied signals are being
+suppressed. The fix path is a consented, leakage separated occupied holdout with
+multiple spatial links and synchronized reference labels.
+
+## Research basis
+
+1. Espressif ESP-IDF WiFi CSI programming guide:
+   https://docs.espressif.com/projects/esp-idf/en/v5.5.4/esp32/api-guides/wifi.html
+2. Espressif ESP CSI reference implementation:
+   https://github.com/espressif/esp-csi
+3. OpenCSI:
+   https://arxiv.org/abs/2607.26665
+4. MORIC:
+   https://arxiv.org/abs/2506.12997
+5. Parameter efficient CSI crowd counting:
+   https://arxiv.org/abs/2601.02203
+6. WiFi JEPA:
+   https://arxiv.org/abs/2607.11064
+
+## Acceptance test
+
+Run the firmware serializer host tests and the focused hardware, signal, and
+sensing server Rust suites. The deterministic serializer test must prove the
+first four bytes are zero only when ESP-IDF marks them invalid and bit 5 must
+round trip through the Rust parser. The temporal estimator must recover a
+1.2 Hz synthetic signal within 0.21 Hz, abstain on short and flat inputs, and
+remain below 10 milliseconds p95 on the target Mac. Existing version 1 model
+input must retain the legacy proxy. A background model with nine references
+must never report a match or suppress a runtime change. Physical firmware is
+not qualified until an S3 and C6 boot and stream the sanitized flag without
+watchdog, memory, or transport regressions during a five minute burn in.
@@ -0,0 +1,143 @@
+# ADR 357: Raw CSI calibration sample integrity
+
+Status: Accepted
+
+Date: 2026-09-06
+
+## Context
+
+The empty room field model requires 1,000 accepted observations over at least
+600 seconds. The sensing server previously invoked calibration from both the
+raw CSI path and the one second edge vital path. Edge vital packets did not add
+a CSI observation, so that path repeatedly fed the existing history tail. A
+counter could therefore satisfy the frame gate without representing 1,000
+independent CSI packets.
+
+Node liveness also combined edge vital and CSI traffic. A source emitting only
+edge vital summaries could appear live enough to start calibration even though
+it had no fresh raw CSI. Finally, cloning the complete per node history before
+each feed copied as much as 204,800 bytes for a 100 frame, 256 subcarrier
+window even though calibration consumes only the current observation.
+
+The earlier single installation result recorded in ADR 355 predates this
+admission boundary. Its frame count is not accepted as evidence of independent
+CSI coverage. It remains useful only as diagnosis of the failure mode and must
+not be promoted as a production bootstrap prior.
+
+## Decision
+
+1. Only a packet on the frozen, model specific CSI grid may advance field
+   calibration. Edge vital packets update their own health and presentation
+   state but never feed the field model. The independent live inference grid
+   remains unchanged.
+2. Each calibration session stores the last admitted wrapping sequence for
+   every contributing node. Duplicate sequences are ignored. Normal forward
+   progress and the `u32::MAX` to zero wrap are accepted.
+3. A packet one through sixteen sequence values behind the current high water mark
+   is bounded UDP reordering. It is recorded and dropped without feeding the
+   model or moving the high water mark. A backward depth above sixteen, or an
+   ambiguous half range sequence, latches that node as faulted for the remainder
+   of the session. Status exposes the policy, reorder window, per node reorder
+   count, maximum observed depth, and `sequence_fault_node_ids`. Finalization
+   returns `calibration_sequence_discontinuity` after a fault. The operator must
+   cancel and restart the empty room capture.
+4. A requested source may start calibration only after a header only, bounded
+   observation window identifies one stable grid. The selected grid needs at
+   least ten seconds of evidence, a measured rate of at least 2 Hz, a latest
+   sample newer than five seconds, and no observed gap of five seconds or more.
+   Calibration status reports the immutable grid, selection evidence, current
+   grid age, and `last_sequence_by_node` without exposing CSI values.
+5. Start, cancel, reset, and process initialization clear sequence admission
+   state. A failed feed does not advance the sequence cursor, so the same
+   packet may be retried after a transient model input failure.
+6. The field bridge receives the current amplitude slice rather than a cloned
+   history. A separate bounded runtime history retains only that same frozen
+   grid, so the model is never trained on one symbol layout and evaluated on a
+   different layout. Canonical grid normalization remains unchanged.
+7. The exact source grid is stored inside the digest protected bootstrap v2
+   payload. A restored model rebuilds only its matching runtime history. An
+   unbound legacy image cannot become an active startup prior.
+8. Successful finalization clears the selected grid runtime history. Bootstrap
+   promotion waits up to 35 seconds for a complete fresh runtime window before
+   scoring twelve additional samples. This covers a 50 frame window at the 2 Hz
+   admission floor plus ten seconds of bounded jitter. Calibration observations
+   therefore cannot enter the held out decision window.
+9. Stop and promotion require the frozen grid to remain fresh and reject any
+   latched sequence fault immediately before persistence. Bootstrap v2 accepts
+   exactly one source node because the current runtime supports one grid bound
+   field model.
+10. The legacy `--calibrate` startup switch fails closed before opening ports.
+    Grid evidence does not exist at process start, so operators start the server,
+    wait ten seconds, then call the explicit single source calibration API.
+
+## Consequences
+
+Calibration frame count now means unique, forward raw CSI observations within
+one uninterrupted sequence epoch. A vitals only node cannot produce a
+bootstrap model. Late packets never become calibration evidence.
+
+One physical Node 5 capture observed sequence 4,119,566 followed by 4,119,563,
+4,119,564, and 4,119,565 in one receive batch. A second independent 90 second
+capture observed bounded late packets at depths 9, 10, and 11. The window of
+sixteen covers the measured maximum with five positions of margin. At the
+firmware acceptance ceiling of 50 Hz, the window represents no more than 320
+milliseconds. Late packets are always dropped. Larger regressions remain
+failed closed.
+
+The same 90 second capture measured three interleaved HT grids on Node 5. The
+64 subcarrier stream delivered 436 frames at 4.84 Hz with a 1.56 second maximum
+gap. The 128 stream delivered 100 frames at 1.11 Hz with a 5.05 second maximum
+gap. The 192 stream delivered 49 frames at 0.54 Hz with an 11.53 second maximum
+gap. Only the 64 grid exceeded the 2 Hz admission floor and stayed inside the
+five second freshness boundary, so the final field model binds to that grid.
+
+The current wire format still has no authenticated boot epoch or stable device
+identity. A restart very early in a session or a same endpoint adversary cannot
+be distinguished cryptographically. A future ADR 018 revision must carry an
+authenticated device identifier and a random per boot identifier before the
+server claims exact restart or identity collision detection.
+
+No new raw CSI is persisted. Sequence cursors and fault markers remain bounded
+in memory to active calibration sources. The bootstrap schema advances to v2
+to protect the grid identity while its negative only authority remains
+unchanged.
+
+## Implementation
+
+1. `NodeState` records header only recent grid evidence and a separate bounded
+   field model history.
+2. `AppStateInner::maybe_feed_calibration_frame` owns source binding, sequence
+   ordering, fault latching, model feed, and cursor commit.
+3. `field_bridge::maybe_feed_calibration` accepts exactly one current amplitude
+   slice and rejects an empty observation.
+4. Calibration start, status, stop, cancel, reset, promotion, refinement, and
+   restart enforce the frozen source and grid contract.
+5. `bootstrap_baseline` persists the source grid inside the verified v2 image.
+6. Finalization clears training history, and promotion blocks until the model
+   has a complete fresh scoring window plus twelve separately spaced samples.
+7. Startup calibration without measured grid evidence exits with a typed
+   operator instruction instead of creating a model that can never receive a
+   frame.
+
+## Acceptance test
+
+Unit tests must prove stable grid selection, rate and gap rejection, first
+packet acceptance, forward progress, duplicate rejection, source and grid
+isolation, maximum counter wrap, empty input retry, measured depth eleven
+reordering without a model feed, depth sixteen acceptance, depth seventeen
+latching, half range rejection, session clear, receipt visibility, protected
+grid persistence, and typed stop refusal.
+Tests must also prove startup flag refusal, missing and stale grid refusal,
+complete holdout isolation, and sequence fault rejection during holdout.
+
+On physical hardware, begin a Node 5 empty room capture and observe a fresh
+sixty second interval. Calibration frame growth must match unique 64 subcarrier
+Node 5 sequences while the independent global inference grid remains unchanged.
+Edge vital packets and other grids must produce zero growth. The bound grid
+must remain active, sequence faults must remain empty, and numeric vital output
+must remain absent. Any bounded late packet must appear in the reorder receipts
+while leaving the model count unchanged. Complete both the 1,000 frame and 600
+second gates, then run the twelve sample server controlled bootstrap holdout.
+Restart and prove that only the persisted grid rebuilds the runtime history.
+Do not claim human detection improvement until a separate consented occupied
+holdout demonstrates retained recall.
@@ -92,10 +92,13 @@ Statuses: **Proposed** (under discussion), **Accepted** (approved and/or impleme
 | [ADR-150](ADR-150-rf-foundation-encoder.md) | RF Foundation Encoder: pose-preserving, subject/room/device-invariant CSI embedding | Proposed |
 | [ADR-151](ADR-151-room-calibration-specialist-training.md) | Per-Room Calibration & Specialized Model Training (room-first → bank of small ruVector specialists) | Proposed |
 | [ADR-152](ADR-152-wifi-pose-sota-2026-intake.md) | WiFi-Pose SOTA 2026 Intake: geometry-conditioned calibration, external benchmarks, foundation-encoder recipe | Proposed |
+| [ADR-345](ADR-345-per-link-csi-and-node-to-node-ranging.md) | Per-link CSI attribution and node-to-node ranging — why phase-based position fails on this hardware (measured), and the coordination-free `(tx_mac, rx_seq)` join key that does not need it | Proposed — measurement layer implemented and validated on hardware; localization not built |
 | [ADR-348](ADR-348-independent-rust-multivariate-forecasting.md) | Independent Rust multivariate forecasting for RuView | Proposed |
 | [ADR-349](ADR-349-governed-local-and-fal-forecast-training.md) | Governed local and fal.ai forecast training | Proposed |
 | [ADR-350](ADR-350-ruvector-predictive-memory-and-ruvllm-boundary.md) | RuVector predictive memory and RuVLLM authority boundary | Proposed |
 | [ADR-354](ADR-354-forecast-governance-enforcement-in-ruforecast-submodule.md) | fal.ai forecast governance and spend enforcement now real, in the vendored RuForecast submodule | Accepted |
+| [ADR-355](ADR-355-local-empty-room-bootstrap-prior.md) | Local empty room bootstrap prior | Accepted |
+| [ADR-356](ADR-356-rate-truthful-csi-and-reliability-gated-adaptation.md) | Rate truthful CSI and reliability gated room adaptation | Accepted |

 ### Platform and UI

@@ -0,0 +1,165 @@
+# Which CSI frames a node keeps decides whether cross-node fusion works
+
+A measured comparison of three frame-selection gates on a nine-node ESP32-C6
+fleet, and a method for repeating it in your own environment.
+
+**This is one house.** Nine boards, one floor plan, one RF environment, one
+transmitter population. The numbers below are what we measured here; the part
+worth reusing is the method, not our figures. `run_arms.py` is included so you
+can measure yours.
+
+---
+
+## The problem
+
+Cross-node fusion pairs frames by transmission identity: two nodes are fusing
+usefully only when they hold measurements of *the same packet in the air*. The
+metric is `paired_fraction` — of the transmissions the fleet reported, what
+share were captured by two or more nodes.
+
+Every node also rate-limits which CSI callbacks it keeps, because processing
+every callback at 100–500 Hz destabilises the WiFi ISR. **How that limiting is
+done turns out to matter far more than we expected**, because it decides
+whether two nodes keep the *same* frames or merely the *same number* of frames.
+
+Three mechanisms, all valid rate limits:
+
+| gate | rule |
+|---|---|
+| `elapsed` | at least 20 ms since **this node's** last accepted frame |
+| `mesh` | first frame heard in each shared 20 ms window of a mesh-aligned clock |
+| `seq` | keep frames where the 802.11 `rx_seq` is divisible by N |
+
+`elapsed` gives every node an independent phase. `mesh` makes nodes agree on
+the *window*. `seq` makes them agree on the *frame*, using a number the
+transmitter assigns — so it needs no clock, no leader and no synchronisation.
+
+## Method
+
+Four arms — the three gates plus `mesh`+`seq` stacked — run as a 2x2 factorial.
+Three deliberate choices, each of which we got wrong first:
+
+**Whole fleet per arm, never a subset.** An earlier attempt compared four gated
+nodes against four ungated ones. Nodes differ enormously in how much traffic
+they hear, so that comparison could not be read at all.
+
+**Verify the gate is actually running.** Each switch is read back from every
+node before sampling. "Config was stored" is not "gate is live", and an earlier
+run could not distinguish a null result from a change that never applied.
+
+**Frame rate is NOT a valid check that a gate is working.** A bucket gate needs
+only one survivor per window, so at ~14 candidate frames per bucket even a
+perfect 1-in-4 gate moves frame rate about 2%. Use instead
+
+    R_window = (delta early_drop + delta accepted) / delta accepted
+
+computed from **differenced** counters. Never divide lifetime totals: they
+carry a large post-boot burst. One node read R = 1126 moments after a reboot,
+26.5 an hour later, and 19.2 windowed.
+
+**Establish the noise floor first.** A null run — three hours, no changes —
+gave drift of sd **0.45 pp** across 20-minute windows, about 2x the ±0.2 pp
+binomial counting error. Without that number none of the results below can be
+called significant.
+
+Arms are 20 minutes in Latin-square blocks (`U M S MS / S MS M U / MS U S M`),
+so slow drift lands on every arm equally rather than on whichever ran first.
+
+## Results
+
+Three blocks, twelve arms, every arm verified live on all nine nodes.
+
+| arm | block 0 | block 1 | block 2 | mean |
+|---|---|---|---|---|
+| `elapsed` | 0.2881 | 0.2409 | 0.2012 | **0.2434** |
+| `mesh` | 0.2692 | 0.2580 | 0.2167 | **0.2480** |
+| `seq` (period 8) | 0.5819 | 0.5573 | 0.5422 | **0.5605** |
+| `mesh`+`seq` | 0.5639 | 0.5623 | 0.5093 | 0.5452 |
+
+**`seq` roughly doubles pairing: +31.7 pp, 2.30x.** That gap is ~70x the
+measured drift floor and reproduces in all three blocks; within-block contrasts
+sit at +0.153 to +0.175 every time.
+
+**Mesh-time alignment bought essentially nothing here: +0.5 pp**, about one
+standard deviation of the noise floor. We built that gate and expected it to be
+the answer; it is indistinguishable from doing nothing. Stacking it on top of
+`seq` is marginally *worse* than `seq` alone.
+
+### It is a trade, not a free win
+
+At period 8 the seq arms carried ~75–95k transmissions per 20 minutes against
+~120–160k ungated. You are buying pairing with raw frame rate. Absolute paired
+transmissions per minute still went **up** (2629 vs 2524), because the frames
+it discards are disproportionately ones only one node would have kept — but
+whether that trade is right depends on what else consumes the stream.
+
+One measurement that made the trade easier here: on-device edge processing was
+**unaffected**. Frames actually processed stayed ~1900 per node per 4 minutes
+in both modes, because that pipeline is separately rate-limited. The cost falls
+on the uplink, not on local sensing.
+
+### The mesh arm was partly an elapsed arm
+
+The mesh gate uses mesh-time buckets only while `c6_sync_espnow_is_valid()`. A
+node without valid sync falls back to the elapsed gate rather than gating on a
+meaningless epoch -- correct behaviour, but it weakens this comparison.
+
+Three of the nine nodes never synchronised during these runs. For those three,
+the `mesh` and `mesh`+`seq` arms silently ran the elapsed gate. So **+0.5 pp is
+not a clean fleet-wide measurement of window alignment**; it is "mesh where sync
+held, elapsed where it did not", and the true effect of alignment on a fully
+synchronised fleet is not measured here.
+
+We are reporting the number as observed rather than reconstructing it, because
+the arms were assigned before we understood the fallback, and a post-hoc
+per-node reanalysis would not be the experiment we ran.
+
+The cause is worth naming because it is environmental, not a firmware defect.
+Time sync is leader-elected and hub-and-spoke: one node broadcasts and the rest
+follow. Whether every node hears that leader is a property of the building. A
+larger or more divided house fragments into sync domains, and any option that
+depends on shared time then degrades quietly for the nodes that fell out --
+which is precisely the situation where someone would most want to trust this
+measurement.
+
+### A prediction we got wrong
+
+Our mesh had split into two leader domains with three nodes never synchronising
+at all. We predicted `seq` would win *most* on node pairs spanning that split.
+Per-pair yield says otherwise:
+
+| pair group | pairs | seq / mesh |
+|---|---|---|
+| same sync domain | 7 | 5.82x |
+| spans the two domains | 8 | 6.24x |
+| involves a never-synced node | 21 | **5.16x** |
+
+Cross-domain is marginally best, but the never-synced nodes are the *lowest* —
+the opposite of the prediction — and within-group spread (3.6x–8.1x) dwarfs
+every between-group difference.
+
+So `seq` is **not** a workaround for broken time sync. It wins roughly
+uniformly, including where sync is healthy. The useful consequence is negative:
+repairing our mesh split would probably not have closed the gap, and we would
+have spent that effort for nothing.
+
+## What we suggest
+
+Not "use `seq`". **Make it selectable and measure your own environment.** The
+right answer depends on your transmitter population, your node placement and
+what consumes your CSI stream — none of which we can see from here.
+
+`run_arms.py` runs the factorial unattended: it applies each arm fleet-wide,
+verifies every node adopted it, samples `/api/v1/fusion`, and writes one row
+per arm. Run the null pass first; without your own drift floor the contrasts
+cannot be judged.
+
+## Limits
+
+- One environment, nine nodes. Treat the numbers as an existence proof, not a
+  constant.
+- Only period 8 was characterised. 4 and 16 are unmeasured, so the
+  yield/pairing curve has a single point on it.
+- `paired_fraction` measures **frame pairing**, which is a precondition for
+  cross-receiver work. Nothing here measures sensing accuracy, and none of it
+  should be read as an accuracy claim.
@@ -0,0 +1,63 @@
+# WiFlow Reproduction Protocol
+
+Status: Proposed research protocol
+
+Tracks issue #1786.
+
+## Objective
+
+Determine whether a CSI optical flow style representation improves reusable camera free motion intelligence in RuView rather than only improving a proxy flow metric.
+
+## Required capture metadata
+
+Every run records hardware family, firmware, antenna count and geometry, channel width, center frequency, packet rate, timestamp source, synchronization method, room identity, participant identity class, label source, calibration, packet loss, missing subcarriers, preprocessing version, model commit, accelerator, software versions, power mode, and random seeds.
+
+## Baselines
+
+1. current RuView CSI preprocessing plus strongest task specific sensing head
+2. WiFlow style motion field representation
+3. simpler temporal derivative or phase and amplitude motion baseline
+
+All conditions use identical train, validation, and held out environment splits.
+
+## Metrics
+
+Primary intermediate metrics:
+
+* endpoint error or equivalent flow error
+* directional consistency
+* temporal stability
+* uncertainty calibration
+
+Required downstream metrics:
+
+* tracking quality
+* occupancy or presence quality
+* pose or activity metric where available
+* spatial state consistency in RuField or WorldGraph
+
+Systems metrics:
+
+* p50 and p95 inference latency
+* peak memory
+* CPU, GPU, or accelerator utilization
+* energy per inference where measurable
+* bytes per observation and sustained bandwidth
+
+## Stress tests
+
+Test unseen rooms, unseen participants, changed antenna geometry, a second RF hardware family, packet loss, burst loss, missing subcarriers, lower packet rate, clock jitter, and partial calibration failure.
+
+## Promotion gate
+
+The representation graduates only when the same learned representation improves at least two downstream RuView tasks on held out environments, or reaches equivalent downstream quality with materially lower edge cost.
+
+A better optical flow proxy score alone is insufficient.
+
+## Security and privacy
+
+Visual or LiDAR labels may be used as privileged training evidence but must not become a runtime dependency for camera free deployment claims. Dataset provenance and consent requirements remain explicit.
+
+## Rollback
+
+The experiment is isolated behind a representation interface. Existing RuView sensing heads remain the production fallback.
@@ -115,6 +115,30 @@ python -m esptool --chip esp32s3 --port COM7 --baud 460800 \
  0x20000 firmware/esp32-csi-node/build/esp32-csi-node.bin
 ```

+#### Seeed XIAO ESP32-C6 external antenna
+
+The XIAO ESP32-C6 has a hardware RF switch between its internal ceramic
+antenna and U.FL connector. Generic C6 boards do not share this wiring, so the
+firmware leaves GPIO3 and GPIO14 untouched by default. For an XIAO with an
+external antenna already attached while powered off, build with one of the
+opt-in overlays:
+
+```bash
+# Controlled A condition: internal antenna, RF path stated in the boot log.
+idf.py -D SDKCONFIG_DEFAULTS="sdkconfig.defaults;sdkconfig.defaults.esp32c6;sdkconfig.defaults.xiao-internal" \
+  set-target esp32c6
+idf.py build
+
+# Controlled B condition: external U.FL antenna.
+idf.py -D SDKCONFIG_DEFAULTS="sdkconfig.defaults;sdkconfig.defaults.esp32c6;sdkconfig.defaults.xiao-external" \
+  set-target esp32c6
+idf.py build
+```
+
+The boot log must report the intended `internal ceramic` or `external U.FL`
+path before that measurement is labelled. The standard C6 build keeps the
+board default and emits no antenna-path claim.
+
 ### 3. Provision WiFi credentials (no reflash needed)

 ```bash
@@ -257,10 +281,17 @@ Offset  Size  Field
 12      4     Sequence number (LE u32)
 16      1     RSSI (i8)
 17      1     Noise floor (i8)
-18      2     Reserved
+18      1     PPDU type (ADR-110; zero when tagging is disabled)
+19      1     Flags: bit0=40 MHz, bit2=STBC, bit4=sync valid,
+              bit5=invalid first CSI word zeroed by firmware
 20      N*2   I/Q pairs (n_antennas * n_subcarriers * 2 bytes)
 ```

+When ESP-IDF marks `first_word_invalid`, the firmware preserves packet geometry,
+zeros the first four CSI bytes, and sets byte 19 bit 5. The same sanitized bytes
+feed the edge DSP path. This prevents a documented hardware artifact from being
+learned as motion while keeping older readers wire compatible.
+
 ### Vitals Packet (32 bytes)

 ```
@@ -0,0 +1,334 @@
+# Firmware runbook — ESP32-C6 CSI node
+
+Operational procedure. `README.md` explains the project; this says exactly what
+to run, in what order, and what to check. Follow it top to bottom.
+
+If you are an agent: **read this file before touching firmware.** Everything
+below was learned by getting it wrong at least once.
+
+---
+
+## 1. Build
+
+**One command. Copy it.** From the **repository root**, not this directory:
+
+```bash
+MSYS_NO_PATHCONV=1 docker run --rm \
+  -v "$(pwd)/firmware/esp32-csi-node:/project" -w /project \
+  espressif/idf:v5.4 bash -c \
+  "cat sdkconfig.defaults sdkconfig.defaults.16mb sdkconfig.defaults.esp32c6 \
+     > sdkconfig.defaults.build && \
+   SDKCONFIG_DEFAULTS='sdkconfig.defaults.build' idf.py set-target esp32c6 && \
+   idf.py build"
+```
+
+Takes ~3 minutes cold, well under a minute incremental.
+
+### The three defaults files are NOT all automatic
+
+IDF picks up `sdkconfig.defaults` and `sdkconfig.defaults.<target>` on its own.
+It does **not** pick up `sdkconfig.defaults.16mb`, and that file is where
+`CONFIG_BOOTLOADER_APP_ROLLBACK_ENABLE=y` lives.
+
+Build without it and you get an image that is correct in every visible respect
+and **cannot roll back a bad OTA** — while `ota_rollback_boot_check()` sits in
+the app expecting the capability. Silent, and only discoverable by pushing a
+deliberately bad image.
+
+That is why the command above concatenates all three explicitly rather than
+relying on discovery.
+
+### Never `rm -rf sdkconfig` without a backup
+
+The README's headline example opens with `rm -rf build sdkconfig`. `sdkconfig`
+is **untracked and gitignored**. `partitions_16mb.csv` was lost to exactly this,
+and `CLAUDE.md` warns about it *even when a README documents it*.
+
+```bash
+cp sdkconfig "sdkconfig.backup-$(date +%Y%m%d-%H%M%S)"   # then regenerate
+```
+
+A stale `sdkconfig` is a real hazard, not a hypothetical: one found on
+2026-09-03 said `FLASHSIZE=4MB`, `partitions_4mb.csv`,
+`DYNAMIC_TX_BUFFER_NUM=64` — against a fleet running 16MB and 128. Anything
+built from it would silently not be the configuration under test.
+
+### Verify before you flash anything
+
+```bash
+grep -E "^CONFIG_(IDF_TARGET|ESPTOOLPY_FLASHSIZE|PARTITION_TABLE_CUSTOM_FILENAME|\
+ESP_WIFI_DYNAMIC_TX_BUFFER_NUM|BOOTLOADER_APP_ROLLBACK_ENABLE)=" sdkconfig
+```
+
+All five must read:
+
+| key | expected |
+|---|---|
+| `CONFIG_IDF_TARGET` | `"esp32c6"` |
+| `CONFIG_ESPTOOLPY_FLASHSIZE` | `"16MB"` |
+| `CONFIG_PARTITION_TABLE_CUSTOM_FILENAME` | `"partitions_16mb.csv"` |
+| `CONFIG_ESP_WIFI_DYNAMIC_TX_BUFFER_NUM` | `128` |
+| `CONFIG_BOOTLOADER_APP_ROLLBACK_ENABLE` | `y` |
+
+**Do not use the native `C:\Espressif\...esp-idf-v5.5.5`.** Wrong version, its
+venv does not match the default python, and a container-built `build/` refuses
+a native build outright. Docker only.
+
+---
+
+## 2. What OTA can and cannot reach
+
+This is the distinction that decides whether a change needs a cable.
+
+| region | offset | reachable by OTA? |
+|---|---|---|
+| bootloader | `0x0` | **NO — USB only** |
+| partition table | `0x8000` | **NO — USB only** |
+| otadata | `0xF000` | rewritten by OTA |
+| ota_0 | `0x20000` | yes (4 MB) |
+| ota_1 | `0x420000` | yes (4 MB) |
+| coredump | `0x820000` | no |
+| storage (fat) | `0x830000` | no |
+| nvs | `0x9000` | never written by a flash — provisioning survives |
+
+**OTA replaces the app partition only.** Anything in the bootloader —
+`CONFIG_BOOTLOADER_*`, notably `APP_ROLLBACK_ENABLE` — or any partition-table
+change **requires USB on every board**.
+
+There is **no bootloader version on the wire**. A node's sync packet reports app
+health, not bootloader capability, so you cannot determine remotely which boards
+have a rollback-capable bootloader. Track it per board, or do a USB pass across
+all of them.
+
+Corollary: "every node reports health" proves the **app** is current and proves
+nothing about the bootloader. Do not answer a bootloader question with app
+evidence.
+
+---
+
+## 3. Deploying
+
+### Over the air — app-only changes
+
+Works end-to-end and is the default for anything that is not a bootloader or
+partition change. Roll to **one node**, soak it, then the rest.
+
+### Over USB — bootloader, partition table, or recovery
+
+```bash
+python -m esptool --chip esp32c6 -b 460800 \
+  --before default_reset --after hard_reset write_flash \
+  --flash_mode dio --flash_size 16MB --flash_freq 80m \
+  0x0     build/bootloader/bootloader.bin \
+  0x8000  build/partition_table/partition-table.bin \
+  0xf000  build/ota_data_initial.bin \
+  0x20000 build/esp32-csi-node.bin
+```
+
+`0x9000` is not in that list, so SSID, password and `node_id` survive.
+
+### Safe single-board diagnostic flash, with a revert path
+
+Nodes run from whichever slot `otadata` selects. **Parse it before flashing** —
+two 32-byte entries at offsets 0 and 0x1000; active slot is
+`(max ota_seq - 1) % 2`.
+
+1. Back up `otadata` (8 KB) **and** the live app slot with `esptool read-flash`.
+2. Flash the diagnostic build — it lands in `ota_0` and rewrites `otadata`.
+3. **Revert = write the 8 KB `otadata` backup back.** Production in `ota_1` was
+   never touched, so no rebuild is needed.
+
+Validate any backup image before trusting it: byte 0 must be `0xE9`, and the
+`esp_app_desc_t` magic at offset `0x20` must be `0xABCD5432`.
+
+**Never flash the whole fleet from a workstation, and never use the server's OTA
+endpoint for a diagnostic.** USB, one board, revert path secured first.
+
+---
+
+## 4. Serial
+
+CH343 USB-serial at 115200.
+
+To read a running board **without resetting it**, open with
+`DtrEnable=$false; RtsEnable=$false`. esptool asserts RTS and *will* reset the
+board, which loses the state you were trying to observe.
+
+---
+
+## 5. Changing config without a cable
+
+Most parameters do not need a flash at all:
+
+```bash
+export RUVIEW_OTA_PSK_FILE=<path to the psk file>       # never inline the key
+python config_push.py --node <node-ip> --get
+python config_push.py --node <node-ip> --set led_brightness=10
+```
+
+Keys marked `assoc: true` in `config_api.c` (`wifi_ssid`, `wifi_password`,
+`channel_hop_count`, `dwell_ms`, `csi_channel`) are applied as a **trial**: the
+node banks its current values, reboots, and keeps the change only if it
+re-associates. Everything else applies without that risk.
+
+`config_push.py` has **no unset**. To restore a default you must set it
+explicitly. Current defaults:
+
+| key | default | source |
+|---|---|---|
+| `vital_interval_ms` | 1000 | `Kconfig.projbuild` |
+| `swarm_heartbeat_sec` | 30 | `nvs_config.c` |
+| `swarm_ingest_sec` | 5 | `nvs_config.c` |
+| `beacon_period_ms` | 0 = derive from fleet size | `nvs_config.c` |
+
+---
+
+## 6. Reading a node remotely
+
+`GET /api/v1/mesh` returns per-node `health`: `min_heap_kib`, `die_c`,
+`thermal_state`, `tx_dbm`, `reset_reason`, and from sync proto v3 the TX-path
+counters `send_fail` / `rate_skip` / `early_drop`.
+
+`min_heap_kib` is a **minimum-since-boot watermark**, not an instantaneous
+reading. It resets on reboot and only ever falls. Sampling it within a minute of
+a restart tells you nothing — that mistake produced a wrong "no effect" verdict
+on 2026-09-03; the real answer, once it had settled, was −56 KiB.
+
+`/api/v1/nodes` `rssi_dbm` is **not** the AP link — it is whatever link landed
+most recently and swings 20+ dB with no physical change. Filter `/api/v1/links`
+by the AP's BSSIDs instead, and read `csi_fps_ema` alongside, because link count
+alone hides a starved node.
+
+---
+
+## 7. Tunables, and which ones depend on your house
+
+`idf.py menuconfig` exposes these. The defaults are what a nine-node ESP32-C6
+fleet in one house settled on; several are **not** universal, and the point of
+listing them here is that you can pick and measure rather than inherit someone
+else's building.
+
+| option | default | depends on your house? |
+|---|---|---|
+| `CSI_GATE_MESH_ALIGNED` | y | **yes** |
+| `CSI_SEQ_DIAG` | n | diagnostic only |
+| `THERMAL_MONITOR` | y | no |
+| `THERMAL_THROTTLE` | n | **yes** |
+| `THERMAL_THROTTLE_C` | 72 | **yes** |
+| `THERMAL_HYSTERESIS_C` | 8 | no |
+| `UPLINK_WATCHDOG` | y | no |
+| `UPLINK_WATCHDOG_TIMEOUT_S` | 1200 | **yes** |
+
+### The ones worth measuring yourself
+
+**`CSI_GATE_MESH_ALIGNED`** decides whether nodes keep the *same* frames or
+merely the same *number* of frames, which is what makes cross-node pairing
+possible at all. Whether alignment helps depends on how much your nodes
+actually overhear in common, and that is a property of your walls and your
+transmitters. In one measured nine-node fleet, aligning windows bought
+0.5 pp of pairing while selecting frames by `rx_seq` bought +31.7 pp -- so do
+not assume the default is doing the heavy lifting in your building.
+
+Turn on `CSI_SEQ_DIAG` to measure it. It logs `(source MAC, rx_seq)` per CSI
+callback, which is what lets you compute the fraction of frames two nodes hold
+in common instead of guessing.
+
+**`THERMAL_THROTTLE` / `THERMAL_THROTTLE_C`** depend on enclosure and placement,
+not on the chip. A bare bench board never reaches 72 C; the same board sealed in
+a printed case on a sunny wall can. Read `die_c` from `/api/v1/mesh` for a few
+days before deciding, because throttling transmit power costs range.
+
+**`UPLINK_WATCHDOG_TIMEOUT_S`** is a bet about your AP. 1200 s suits a network
+where a wedge is rare and a reboot is cheap. If your uplink drops briefly and
+often, a short timeout turns a recoverable gap into a reboot loop; if it wedges
+hard, a long one leaves a node dark for twenty minutes.
+
+### A worked example: the same two options, in one house
+
+MEASURED on a nine-node ESP32-C6 fleet in a single detached house. **These are
+our numbers, not yours.** They are here to show the shape of the decision, and
+how much a building can move it.
+
+Four arms as a 2x2 factorial, 20-minute runs, three Latin-square blocks so slow
+drift landed on every arm equally. Metric is the fraction of transmissions held
+in common by two or more nodes.
+
+| arm | mean |
+|---|---|
+| `elapsed` (per-node gate) | 0.2434 |
+| `mesh` (aligned windows) | 0.2480 |
+| `seq` (keep `rx_seq` divisible by 8) | **0.5605** |
+| `mesh` + `seq` | 0.5452 |
+
+`seq` roughly doubled pairing, +31.7 pp, about 70x the measured drift floor and
+reproduced in all three blocks. Mesh alignment bought **+0.5 pp** -- inside
+noise. Stacking them was very slightly *worse* than `seq` alone.
+
+**Why that happened here, and why yours may differ.** These nodes already
+overheard a great deal in common -- two boards side by side heard the same
+transmissions 72% of the time -- so there was a large pool of shared
+transmissions for `seq` to select from, and little left for window alignment to
+recover. A house with fewer shared transmitters, thicker interior walls, or
+nodes spread so that few pairs hear the same radios has less for `seq` to work
+with, and phase alignment may matter more rather than less. We did not measure
+such a house and do not claim to know.
+
+**It is a trade, not a free win.** At period 8 the `seq` arms carried ~75-95k
+transmissions per 20 minutes against ~120-160k ungated: pairing is bought with
+raw frame rate. Absolute paired transmissions per minute still rose (2629 vs
+2524), because the discarded frames are disproportionately ones only one node
+would have kept. Whether that trade is right depends on what consumes your
+stream. On-device edge processing was **unaffected** here -- frames processed
+stayed ~1900 per node per 4 minutes in both modes, because that pipeline is
+separately rate-limited -- so the cost fell on the uplink, not on local sensing.
+If your uplink is the scarce resource, that changes the answer.
+
+**The two options differ in kind, not just in effect.** Mesh alignment is an
+agreement between receivers: it needs a shared clock, so it inherits whatever
+the sync topology does. Selecting by `rx_seq` needs no agreement at all. The
+802.11 sequence control field is assigned by the TRANSMITTER and is identical
+at every receiver, so it is a coordination-free name for one transmission --
+two nodes that have never heard of each other, sit in different sync domains,
+or run with no mesh whatsoever still keep the same subset of frames.
+
+In the code the seq test runs before any sync check and reads only the frame;
+only the mesh branch calls `c6_sync_espnow_is_valid()`. That asymmetry is the
+reason `seq` had nothing to degrade in our runs, and it is worth weighing
+directly: if your fleet's sync is fragile, `seq` is not merely the better
+performer, it is the one with no dependency that can quietly fail.
+
+**A limitation of our own measurement, which cuts against the mesh arm.** The
+mesh gate uses mesh-time buckets only while `c6_sync_espnow_is_valid()`; a node
+with no valid sync falls back to the elapsed gate rather than gating on a
+meaningless epoch. That is the right behaviour, but it means our `mesh` arms
+were partly `elapsed` arms: three of nine nodes never synchronised, so for those
+three the setting under test was silently not in effect. Read the +0.5 pp as
+"mesh where sync held, elapsed where it did not" rather than as a clean
+fleet-wide measurement of alignment.
+
+This is itself a house-dependent property. Time sync here is leader-elected and
+hub-and-spoke -- one node broadcasts and the rest follow. Whether every node can
+hear that leader is a question about your building, not about the firmware. In a
+larger or more divided house the leader may not reach everyone, the fleet
+fragments into sync domains, and any option that depends on shared time degrades
+quietly for the nodes that fell out. Check `is_leader` and sync validity in
+`/api/v1/mesh` before concluding that an alignment setting did or did not help.
+
+**A prediction we got wrong, kept because it is the useful part.** Our mesh had
+split into two leader domains with three nodes never synchronising at all. We
+expected `seq` to win most on pairs spanning that split. Per-pair yield said
+otherwise: same-domain 5.82x, cross-domain 6.24x, involving a never-synced node
+**5.16x** -- the lowest, the opposite of the prediction -- and within-group
+spread (3.6x-8.1x) dwarfed every between-group difference.
+
+So `seq` is **not** a workaround for broken time sync; it wins roughly
+uniformly, including where sync is healthy. The consequence is negative and
+worth having: repairing the mesh split would probably not have closed the gap,
+and that effort would have been spent for nothing.
+
+### Measuring rather than guessing
+
+A gate comparison needs a fixed environment, repeated blocks, and every arm run
+on every node -- otherwise node placement is confounded with the setting. The
+harness used for the published comparison is in the repository; use it rather
+than eyeballing a frame rate, which moves for reasons unrelated to the setting.
@@ -14,6 +14,7 @@ set(SRCS
    "c6_twt.c"
    "c6_timesync.c"
    "c6_lp_core.c"
+    "c6_antenna_select.c"
    # ADR-110 D1 workaround — ESP-NOW cross-node sync (works on S3+C6)
    "c6_sync_espnow.c"
    # ADR-110 B1/B2 unblock — soft-AP HE/TWT (C6-only when enabled)
@@ -48,6 +49,12 @@ if(IDF_TARGET STREQUAL "esp32c6")
    list(APPEND REQUIRES ieee802154 ulp esp_hw_support)
 endif()

+# ESP-IDF 6 split the temperature sensor driver out of the legacy `driver`
+# component. Keep the 5.4 lane working while declaring the v6 dependency.
+if(IDF_VERSION_MAJOR GREATER_EQUAL 6)
+    list(APPEND REQUIRES esp_driver_tsens)
+endif()
+
 # ADR-061: Mock CSI generator for QEMU testing + ADR-081 mock radio binding
 if(CONFIG_CSI_MOCK_ENABLED)
    list(APPEND SRCS "mock_csi.c" "rv_radio_ops_mock.c")
@@ -1,5 +1,27 @@
 menu "CSI Node Configuration"

+    config C6_XIAO_ANTENNA_SELECT
+        bool "Control the Seeed XIAO ESP32-C6 RF switch"
+        depends on IDF_TARGET_ESP32C6
+        default n
+        help
+            Enable only for Seeed XIAO ESP32-C6 boards. The XIAO uses GPIO3
+            to enable RF-switch control and GPIO14 to choose its internal or
+            external antenna. Leave this disabled for generic ESP32-C6 boards
+            because GPIO3 and GPIO14 may have unrelated functions there.
+
+    choice C6_XIAO_ANTENNA_PATH
+        prompt "Seeed XIAO ESP32-C6 antenna path"
+        depends on C6_XIAO_ANTENNA_SELECT
+        default C6_XIAO_ANTENNA_INTERNAL
+
+        config C6_XIAO_ANTENNA_INTERNAL
+            bool "Internal ceramic antenna"
+
+        config C6_XIAO_ANTENNA_EXTERNAL
+            bool "External U.FL antenna"
+    endchoice
+
    config CSI_NODE_ID
        int "Node ID (0-255)"
        default 1
@@ -651,3 +673,20 @@ config UPLINK_WATCHDOG_TIMEOUT_S
      packet's health field before assuming the watchdog is responsible.

 endmenu
+
+config CSI_GATE_MESH_ALIGNED
+    bool "Align the CSI rate gate to mesh time (enables cross-node pairing)"
+    default y
+    help
+      The per-node elapsed gate ("20 ms since MY last accept") gives every node
+      an independent phase, so two healthy nodes at 50 Hz accept disjoint
+      frames and nothing can be paired across them. This aligns the 20 ms
+      window to mesh time so nodes choose the same windows.
+
+config CSI_SEQ_DIAG
+    bool "Log (source MAC, rx_seq) for every CSI callback (ADR-345 phase 2 probe)"
+    default n
+    help
+      Diagnostic. Emits the transmitter address and 802.11 sequence number for
+      each CSI callback so the effect of the gate above can be measured rather
+      than assumed. Verbose; leave off in normal operation.
@@ -0,0 +1,56 @@
+#include "c6_antenna_select.h"
+
+#include "sdkconfig.h"
+
+#if defined(CONFIG_IDF_TARGET_ESP32C6) && defined(CONFIG_C6_XIAO_ANTENNA_SELECT)
+#include "driver/gpio.h"
+#include "esp_log.h"
+#include "freertos/FreeRTOS.h"
+#include "freertos/task.h"
+
+#define XIAO_RF_SWITCH_POWER_GPIO GPIO_NUM_3
+#define XIAO_RF_SWITCH_SELECT_GPIO GPIO_NUM_14
+
+static const char *TAG = "c6_antenna";
+#endif
+
+esp_err_t c6_xiao_antenna_apply(void)
+{
+#if defined(CONFIG_IDF_TARGET_ESP32C6) && defined(CONFIG_C6_XIAO_ANTENNA_SELECT)
+    gpio_config_t config = {
+        .pin_bit_mask = (1ULL << XIAO_RF_SWITCH_POWER_GPIO) |
+                        (1ULL << XIAO_RF_SWITCH_SELECT_GPIO),
+        .mode = GPIO_MODE_OUTPUT,
+        .pull_up_en = GPIO_PULLUP_DISABLE,
+        .pull_down_en = GPIO_PULLDOWN_DISABLE,
+        .intr_type = GPIO_INTR_DISABLE,
+    };
+    esp_err_t err = gpio_config(&config);
+    if (err != ESP_OK) {
+        return err;
+    }
+
+    err = gpio_set_level(XIAO_RF_SWITCH_POWER_GPIO, 0);
+    if (err != ESP_OK) {
+        return err;
+    }
+    vTaskDelay(pdMS_TO_TICKS(100));
+
+#if defined(CONFIG_C6_XIAO_ANTENNA_EXTERNAL)
+    const c6_xiao_antenna_t antenna = C6_XIAO_ANTENNA_EXTERNAL;
+    const char *name = "external U.FL";
+#else
+    const c6_xiao_antenna_t antenna = C6_XIAO_ANTENNA_INTERNAL;
+    const char *name = "internal ceramic";
+#endif
+
+    err = gpio_set_level(XIAO_RF_SWITCH_SELECT_GPIO,
+                         c6_xiao_antenna_gpio14_level(antenna));
+    if (err == ESP_OK) {
+        ESP_LOGI(TAG, "XIAO ESP32-C6 antenna path: %s", name);
+    }
+    return err;
+#else
+    return ESP_OK;
+#endif
+}
@@ -0,0 +1,17 @@
+#pragma once
+
+typedef enum {
+    C6_XIAO_ANTENNA_INTERNAL = 0,
+    C6_XIAO_ANTENNA_EXTERNAL = 1,
+} c6_xiao_antenna_t;
+
+static inline int c6_xiao_antenna_gpio14_level(c6_xiao_antenna_t antenna)
+{
+    return antenna == C6_XIAO_ANTENNA_EXTERNAL ? 1 : 0;
+}
+
+#ifdef ESP_PLATFORM
+#include "esp_err.h"
+
+esp_err_t c6_xiao_antenna_apply(void);
+#endif
@@ -13,7 +13,7 @@
 *      `ulp_embed_binary(...)` in main/CMakeLists.txt from lp_core/main.c.
 *
 *   2. DISABLED — falls back to plain deep-sleep + GPIO wake-up
- *      (`esp_deep_sleep_enable_gpio_wakeup`). No debounce, no
+ *      (ESP-IDF deep-sleep GPIO wake API). No debounce, no
 *      sub-10 µA floor, but no LP toolchain dependency either.
 *      This is the path the v0.6.6 firmware shipped with.
 *
@@ -26,6 +26,7 @@
 #if defined(CONFIG_IDF_TARGET_ESP32C6) && defined(CONFIG_ULP_COPROC_TYPE_LP_CORE)

 #include "c6_lp_core.h"
+#include "esp_idf_version.h"
 #include "esp_log.h"
 #include "esp_sleep.h"
 #include "driver/rtc_io.h"
@@ -134,10 +135,17 @@ esp_err_t c6_lp_core_arm(int wake_gpio, bool active_high)
 #else
    /* --- Fallback path: plain deep-sleep GPIO wakeup (~10 µA floor) --- */
    uint64_t mask = 1ULL << wake_gpio;
+#if ESP_IDF_VERSION >= ESP_IDF_VERSION_VAL(6, 0, 0)
+    esp_sleep_gpio_wake_up_mode_t mode = active_high
+        ? ESP_GPIO_WAKEUP_GPIO_HIGH
+        : ESP_GPIO_WAKEUP_GPIO_LOW;
+    esp_err_t err = esp_sleep_enable_gpio_wakeup_on_hp_periph_powerdown(mask, mode);
+#else
    esp_deepsleep_gpio_wake_up_mode_t mode = active_high
        ? ESP_GPIO_WAKEUP_GPIO_HIGH
        : ESP_GPIO_WAKEUP_GPIO_LOW;
    esp_err_t err = esp_deep_sleep_enable_gpio_wakeup(mask, mode);
+#endif
    if (err != ESP_OK) {
        ESP_LOGE(TAG, "enable_gpio_wakeup failed: %s", esp_err_to_name(err));
        return err;
@@ -114,6 +114,9 @@ static int64_t s_last_send_us = 0;
 */
 #define CSI_MIN_PROCESS_INTERVAL_US  (20 * 1000)  /* 50 Hz */
 static int64_t s_last_process_us = 0;
+/* Mesh-epoch window index of the last accepted frame. UINT64_MAX means "none
+ * yet", which no real bucket can collide with. */
+static uint64_t s_last_gate_bucket = UINT64_MAX;
 static uint32_t s_early_drop = 0;

 /* ---- ADR-029: Channel-hop state ---- */
@@ -263,8 +266,16 @@ size_t csi_serialize_frame(const wifi_csi_info_t *info, uint8_t *buf, size_t buf
    buf[19] = 0;
 #endif

-    /* I/Q data */
+    /* I/Q data. ESP-IDF documents that the first four CSI bytes are invalid
+     * when first_word_invalid is set. Never let those hardware artifacts feed
+     * either the host or edge DSP path. Preserve the fixed ADR-018 geometry by
+     * zeroing rather than removing bytes, and mark the sanitation in bit 5. */
    memcpy(&buf[CSI_HEADER_SIZE], info->buf, iq_len);
+    if (info->first_word_invalid && iq_len > 0) {
+        size_t invalid_len = iq_len < 4 ? iq_len : 4;
+        memset(&buf[CSI_HEADER_SIZE], 0, invalid_len);
+        buf[19] |= CSI_FLAG_FIRST_WORD_SANITIZED;
+    }

    return frame_size;
 }
@@ -277,9 +288,53 @@ static void wifi_csi_callback(void *ctx, wifi_csi_info_t *info)
    (void)ctx;

    /* Early rate gate: drop excess callbacks to ~50 Hz to prevent
-     * SPI flash cache crash in WiFi ISR (wDev_ProcessFiq). */
+     * SPI flash cache crash in WiFi ISR (wDev_ProcessFiq).
+     *
+     * HOW the rate is limited decides whether cross-node fusion is possible.
+     *
+     * The original gate is "at least 20 ms since MY last accept". Each node
+     * therefore has an independent phase: node A can accept at t=0,20,40 ms
+     * while node B accepts at 7,27,47. Both are perfectly healthy and both hit
+     * 50 Hz, yet they accept disjoint frames and nothing can be paired.
+     *
+     * MEASURED 2026-08-30, two boards side by side: they HEARD the same
+     * transmissions 72% of the time but only 25% of accepted frames were
+     * common. The frames were there; the gate was throwing away the pairing.
+     *
+     * So gate on a bucket of the MESH-ALIGNED epoch instead: every node takes
+     * the first frame it hears in the same absolute 20 ms window. Two nodes
+     * that heard the same frame now both accept it. Mesh sync is good to
+     * roughly 490 us against a 20 ms bucket -- about 2.5% of a window -- so
+     * boundary disagreements are rare.
+     *
+     * Falls back to the original elapsed-time gate whenever mesh sync is not
+     * valid (no leader heard yet, or a node that just booted), because an
+     * unsynced epoch would otherwise gate on a meaningless number. A node
+     * running unsynced simply gets the old behaviour rather than no gate.
+     *
+     * The half-interval floor below keeps the crash protection honest: bucket
+     * gating alone could accept at the end of one window and the start of the
+     * next, back to back. The floor bounds that to ~100 Hz for a single pair
+     * while the average stays at 50 Hz. Do not remove it; the gate exists for
+     * a crash, not for tidiness. */
    int64_t now_us = esp_timer_get_time();
-    if ((now_us - s_last_process_us) < CSI_MIN_PROCESS_INTERVAL_US) {
+    bool take;
+#ifdef CONFIG_CSI_GATE_MESH_ALIGNED
+    if (c6_sync_espnow_is_valid()) {
+        uint64_t bucket = c6_sync_espnow_get_epoch_us()
+                        / (uint64_t)CSI_MIN_PROCESS_INTERVAL_US;
+        take = (bucket != s_last_gate_bucket)
+            && ((now_us - s_last_process_us) >= CSI_MIN_PROCESS_INTERVAL_US / 2);
+        if (take) {
+            s_last_gate_bucket = bucket;
+        }
+    } else {
+        take = (now_us - s_last_process_us) >= CSI_MIN_PROCESS_INTERVAL_US;
+    }
+#else
+    take = (now_us - s_last_process_us) >= CSI_MIN_PROCESS_INTERVAL_US;
+#endif
+    if (!take) {
        s_early_drop++;
        return;
    }
@@ -330,13 +385,16 @@ static void wifi_csi_callback(void *ctx, wifi_csi_info_t *info)
     * while the on-device Tier 1/2 pipeline receives a uniform, sustainable
     * stream. Enqueuing every burst frame overloaded the unicore C6 DSP and
     * turned 30-40 callback pps into an irregular approximately 8 Hz subset. */
-    if (info->buf && info->len > 0) {
+    if (frame_len > CSI_HEADER_SIZE) {
        if (s_next_edge_enqueue_us == 0) {
            s_next_edge_enqueue_us = now_us;
        }

        if (now_us >= s_next_edge_enqueue_us) {
-            (void)edge_enqueue_csi((const uint8_t *)info->buf, (uint16_t)info->len,
+            /* Reuse the sanitized ADR-018 payload. Feeding info->buf here
+             * would reintroduce first_word_invalid artifacts on device. */
+            (void)edge_enqueue_csi(&frame_buf[CSI_HEADER_SIZE],
+                                   (uint16_t)(frame_len - CSI_HEADER_SIZE),
                                   (int8_t)info->rx_ctrl.rssi, info->rx_ctrl.channel);

            /* Preserve the configured sample clock instead of resetting it to
@@ -17,6 +17,9 @@
 /** ADR-018 header size in bytes. */
 #define CSI_HEADER_SIZE 20

+/** ADR-018 byte 19: ESP-IDF reported and firmware sanitized invalid CSI prefix. */
+#define CSI_FLAG_FIRST_WORD_SANITIZED (1U << 5)
+
 /** Maximum frame buffer size (header + 4 antennas * 256 subcarriers * 2 bytes). */
 #define CSI_MAX_FRAME_SIZE (CSI_HEADER_SIZE + 4 * 256 * 2)

@@ -37,6 +37,7 @@
 #include "c6_twt.h"                /* ADR-110: TWT (no-op stub on S3) */
 #include "c6_timesync.h"           /* ADR-110: 802.15.4 mesh time-sync (no-op on S3) */
 #include "c6_lp_core.h"            /* ADR-110: LP-core hibernation (no-op on S3) */
+#include "c6_antenna_select.h"     /* XIAO C6 RF switch (opt-in; generic C6 no-op) */
 #include "c6_sync_espnow.h"        /* ADR-110 D1 workaround: ESP-NOW sync */
 #include "c6_softap_he.h"          /* ADR-110 B1/B2: HE/TWT soft-AP (no-op when disabled) */
 #ifdef CONFIG_CSI_MOCK_ENABLED
@@ -331,6 +332,11 @@ void app_main(void)
    ESP_LOGI(TAG, "%s CSI Node (ADR-018 / ADR-110) — v%s — Node ID: %d",
             target_name, app_desc->version, g_nvs_config.node_id);

+    /* Apply the opt-in XIAO RF path before WiFi starts. Generic C6 boards keep
+     * ownership of GPIO3 and GPIO14 because the default implementation is a
+     * no-op unless CONFIG_C6_XIAO_ANTENNA_SELECT is enabled. */
+    ESP_ERROR_CHECK(c6_xiao_antenna_apply());
+
    /* Onboard WS2812. C6 wires the LED to GPIO 8; S3 to GPIO 38 (DevKitC-1 v1.0)
     * or GPIO 48 (DevKitC-1 v1.1 / N16R8 — see #962). On S3 we drive 48 (the
     * common module). On C6, GPIO 38/48 don't exist (only 0-30) — gate by target.
@@ -0,0 +1,182 @@
+#!/usr/bin/env python3
+"""Run the gate factorial: U / M / S / MS, Latin-square blocked.
+
+Arms (NVS gate_mode, applied live -- no reboot):
+    0 U   elapsed-time gate, as upstream ships it
+    1 M   mesh-aligned bucket gate (ours, 4b54ef8e)
+    2 S   rx_seq selection + elapsed underneath
+    3 MS  both stacked
+
+Why it is built this way, and what each guard is for:
+
+* Every arm is applied to the WHOLE fleet, never to a subset. The stopped
+  2026-09-04 run compared four gated nodes against four different control
+  nodes; nodes differ, and that comparison could not be read.
+
+* After every switch the runner reads back `active.gate_mode` from EVERY node
+  and refuses to sample unless all nine agree. "Config was stored" is not
+  "gate is running" -- that gap is exactly what the previous run could not
+  distinguish from a null result.
+
+* Arms are short and blocked in a Latin square rather than run in long
+  sequential batches, so slow drift lands on all arms equally instead of being
+  aliased onto whichever arm ran first. Measured drift over 20 min windows is
+  sd 0.0045; counting noise is +/-0.002.
+
+* Metrics come from differenced cumulative counters, so the server is never
+  restarted and the fleet stays in production throughout.
+
+Output: arms_<label>.csv, one row per arm, in the fleet-baselines directory
+(outside git -- it carries presence information).
+"""
+import argparse, csv, json, os, sys, time, urllib.request
+
+PORT = 8032
+SERVER = "http://127.0.0.1:3000"
+OUTDIR = "C:/temp/ruview/fleet-baselines/gate-experiment"
+
+NODES = {0: "192.168.1.171", 1: "192.168.1.219", 2: "192.168.1.72",
+         3: "192.168.1.112", 4: "192.168.1.144", 5: "192.168.1.198",
+         6: "192.168.1.199", 7: "192.168.1.25",  8: "192.168.1.122"}
+
+ARM_NAME = {0: "U_elapsed", 1: "M_mesh", 2: "S_seq", 3: "MS_both"}
+
+# Latin square: every arm appears once per block, in a different position.
+BLOCKS = [[0, 1, 2, 3], [2, 3, 1, 0], [3, 0, 2, 1]]
+
+
+def key():
+    p = os.environ.get("RUVIEW_OTA_PSK_FILE")
+    if not p or not os.path.isfile(p):
+        sys.exit("set RUVIEW_OTA_PSK_FILE to the psk file path")
+    return open(p, encoding="utf-8").read().strip()
+
+
+def node_get(ip, path, k, timeout=10):
+    r = urllib.request.Request(f"http://{ip}:{PORT}{path}")
+    r.add_header("Authorization", "Bearer " + k)
+    with urllib.request.urlopen(r, timeout=timeout) as resp:
+        return json.load(resp)
+
+
+def node_post(ip, path, k, payload, timeout=15):
+    body = json.dumps(payload).encode()
+    r = urllib.request.Request(f"http://{ip}:{PORT}{path}", data=body)
+    r.add_header("Authorization", "Bearer " + k)
+    r.add_header("Content-Type", "application/json")
+    with urllib.request.urlopen(r, timeout=timeout) as resp:
+        return resp.status, resp.read().decode("utf-8", "replace")
+
+
+def srv(path):
+    with urllib.request.urlopen(SERVER + path, timeout=15) as r:
+        return json.load(r)
+
+
+def set_arm(mode, period, k):
+    """Apply an arm fleet-wide and PROVE every node is running it."""
+    for nid, ip in sorted(NODES.items()):
+        try:
+            node_post(ip, "/config", k, {"gate_mode": mode, "gate_seq_period": period})
+        except Exception as e:
+            print(f"  n{nid} set FAILED: {e}", file=sys.stderr)
+    time.sleep(5)
+    bad = []
+    for nid, ip in sorted(NODES.items()):
+        try:
+            act = node_get(ip, "/config", k).get("active", {})
+            if act.get("gate_mode") != mode or act.get("gate_seq_period") != period:
+                bad.append((nid, act))
+        except Exception as e:
+            bad.append((nid, str(e)))
+    return bad
+
+
+def snapshot():
+    f = srv("/api/v1/fusion")
+    m = srv("/api/v1/mesh")["nodes"]
+    nodes = {}
+    for k2, v in m.items():
+        tx = v.get("health", {}).get("tx", {})
+        nodes[k2] = {"acc": v.get("csi_fps_samples"), "ed": tx.get("early_drop"),
+                     "fps": v.get("csi_fps_ema"), "valid": v.get("is_valid")}
+    pairs = {f"{p['a']}-{p['b']}": p["common"] for p in f.get("pairs", [])}
+    return {"t": time.time(), "observations": f["observations"],
+            "transmissions": f["transmissions"], "paired": f["paired"],
+            "nodes": nodes, "pairs": pairs}
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--label", default="factorial")
+    ap.add_argument("--arm-minutes", type=float, default=20.0)
+    ap.add_argument("--settle-minutes", type=float, default=3.0)
+    ap.add_argument("--period", type=int, default=8)
+    ap.add_argument("--blocks", type=int, default=3)
+    a = ap.parse_args()
+
+    k = key()
+    os.makedirs(OUTDIR, exist_ok=True)
+    out = os.path.join(OUTDIR, f"arms_{a.label}.csv")
+    new = not os.path.exists(out)
+    fh = open(out, "a", newline="", encoding="utf-8")
+    w = csv.writer(fh)
+    if new:
+        w.writerow(["block", "arm", "arm_name", "period", "start_iso", "end_iso",
+                    "dur_s", "d_observations", "d_transmissions", "d_paired",
+                    "paired_fraction", "node_json", "pairs_json", "verified"])
+        fh.flush()
+
+    print(f"gate factorial: {a.blocks} blocks x 4 arms x {a.arm_minutes:g} min "
+          f"(+{a.settle_minutes:g} min settle), seq period {a.period}")
+    print(f"est. total {a.blocks*4*(a.arm_minutes+a.settle_minutes)/60:.1f} h -> {out}\n")
+
+    for bi in range(a.blocks):
+        order = BLOCKS[bi % len(BLOCKS)]
+        print(f"--- block {bi}: {' '.join(ARM_NAME[m] for m in order)} ---")
+        for mode in order:
+            print(f"[{time.strftime('%H:%M:%S')}] arm {ARM_NAME[mode]} ...", flush=True)
+            bad = set_arm(mode, a.period, k)
+            if bad:
+                print(f"  WARNING not all nodes on arm: {bad}", file=sys.stderr)
+            # Settle: the gate changes instantly, but csi_fps_ema and the
+            # fusion pairing window carry the previous arm for a while.
+            time.sleep(a.settle_minutes * 60)
+            s0 = snapshot()
+            time.sleep(a.arm_minutes * 60)
+            s1 = snapshot()
+
+            dtr = s1["transmissions"] - s0["transmissions"]
+            dp = s1["paired"] - s0["paired"]
+            pf = dp / dtr if dtr > 0 else float("nan")
+            nodes = {}
+            for nid in s1["nodes"]:
+                if nid in s0["nodes"]:
+                    dacc = (s1["nodes"][nid]["acc"] or 0) - (s0["nodes"][nid]["acc"] or 0)
+                    ded = (s1["nodes"][nid]["ed"] or 0) - (s0["nodes"][nid]["ed"] or 0)
+                    nodes[nid] = {"d_acc": dacc, "d_ed": ded,
+                                  "R": round((dacc + ded) / dacc, 3) if dacc > 0 else None,
+                                  "fps": round(s1["nodes"][nid]["fps"] or 0, 2),
+                                  "valid": s1["nodes"][nid]["valid"]}
+            dpairs = {kk: s1["pairs"].get(kk, 0) - s0["pairs"].get(kk, 0)
+                      for kk in s1["pairs"]}
+            w.writerow([bi, mode, ARM_NAME[mode], a.period,
+                        time.strftime("%Y-%m-%dT%H:%M:%S", time.localtime(s0["t"])),
+                        time.strftime("%Y-%m-%dT%H:%M:%S", time.localtime(s1["t"])),
+                        round(s1["t"] - s0["t"]),
+                        s1["observations"] - s0["observations"], dtr, dp,
+                        f"{pf:.6f}",
+                        json.dumps(nodes, separators=(",", ":")),
+                        json.dumps(dpairs, separators=(",", ":")),
+                        "no" if bad else "yes"])
+            fh.flush()
+            meanR = [v["R"] for v in nodes.values() if v["R"]]
+            print(f"    paired_fraction={pf:.4f}  transmissions={dtr}  "
+                  f"meanR={sum(meanR)/len(meanR):.2f}" if meanR else "", flush=True)
+
+    fh.close()
+    print("done")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,4 @@
+# Seeed XIAO ESP32-C6 external U.FL antenna test overlay.
+# Use only with the XIAO board; generic C6 boards may use GPIO3/GPIO14.
+CONFIG_C6_XIAO_ANTENNA_SELECT=y
+CONFIG_C6_XIAO_ANTENNA_EXTERNAL=y
@@ -0,0 +1,4 @@
+# Seeed XIAO ESP32-C6 internal ceramic antenna test overlay.
+# Keeps RF-switch control enabled so the boot log proves the selected path.
+CONFIG_C6_XIAO_ANTENNA_SELECT=y
+CONFIG_C6_XIAO_ANTENNA_INTERNAL=y
@@ -39,16 +39,45 @@ CFLAGS   = -fsanitize=fuzzer,address,undefined -g -O1 \
 STUBS_SRC = stubs/esp_stubs.c
 MAIN_DIR  = ../main

+# GNU ld and Apple's ld spell dead section elimination differently.
+# Keep the host sanitation test portable across the Linux CI runner and macOS.
+UNAME_S := $(shell uname -s)
+ifeq ($(UNAME_S),Darwin)
+DEAD_STRIP_LDFLAGS := -Wl,-dead_strip
+else
+DEAD_STRIP_LDFLAGS := -Wl,--gc-sections
+endif
+
 # Default fuzz duration (seconds) and jobs
 FUZZ_DURATION ?= 30
 FUZZ_JOBS     ?= 1

 .PHONY: all clean run_serialize run_edge run_nvs run_all test_adr110 run_adr110 \
        test_vitals run_vitals test_mmwave_detect run_mmwave_detect host_tests \
-        test_thermal run_thermal
+        test_thermal run_thermal test_csi_sanitize run_csi_sanitize \
+        test_c6_antenna run_c6_antenna

 all: fuzz_serialize fuzz_edge fuzz_nvs test_adr110 test_vitals test_mmwave_detect \
-     test_thermal
+     test_thermal test_csi_sanitize test_c6_antenna
+
+# --- XIAO ESP32-C6 RF switch selection ---
+# Host-side truth table for the GPIO14 antenna selector. GPIO3 is always driven
+# low by the device implementation before this level is applied.
+test_c6_antenna: test_c6_antenna_select.c $(MAIN_DIR)/c6_antenna_select.h
+	cc -std=c99 -Wall -Wextra -I$(MAIN_DIR) -o $@ $<
+
+run_c6_antenna: test_c6_antenna
+	./test_c6_antenna
+
+# --- ESP-IDF invalid first-word sanitation ---
+# Links the production serializer. Function sections and dead stripping keep
+# unrelated device-only callback code out of this host executable.
+test_csi_sanitize: test_csi_sanitize.c $(MAIN_DIR)/csi_collector.c $(STUBS_SRC)
+	cc -std=c99 -Wall -Wextra -Istubs -I$(MAIN_DIR) -ffunction-sections \
+		$^ $(DEAD_STRIP_LDFLAGS) -o $@ -lm
+
+run_csi_sanitize: test_csi_sanitize
+	./test_csi_sanitize

 # --- ADR-110 encoding unit tests ---
 # Host-side, no libFuzzer needed — plain C99 deterministic table tests
@@ -93,8 +122,9 @@ test_thermal: test_thermal_hysteresis.c $(MAIN_DIR)/thermal.h
 run_thermal: test_thermal
 	./test_thermal

-host_tests: run_adr110 run_vitals run_mmwave_detect run_thermal
-	@echo "Host tests passed (ADR-110 + vitals #998/#996 + mmwave detect #1135 + thermal)"
+host_tests: run_adr110 run_vitals run_mmwave_detect run_thermal run_csi_sanitize \
+            run_c6_antenna
+	@echo "Host tests passed (ADR-110 + CSI sanitation + vitals + mmwave + thermal + C6 antenna)"

 # --- Serialize fuzzer ---
 # Tests csi_serialize_frame() with random wifi_csi_info_t inputs.
@@ -130,5 +160,6 @@ run_nvs: fuzz_nvs
 run_all: run_serialize run_edge run_nvs

 clean:
-	rm -f fuzz_serialize fuzz_edge fuzz_nvs test_adr110 test_vitals
+	rm -f fuzz_serialize fuzz_edge fuzz_nvs test_adr110 test_vitals test_mmwave_detect \
+		test_thermal test_csi_sanitize test_c6_antenna
 	rm -rf corpus_serialize/ corpus_edge/ corpus_nvs/
@@ -90,6 +90,7 @@ int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size)
    info.rx_ctrl.sig_mode      = legacy_inputs & 0x03;
    info.rx_ctrl.cwb           = (legacy_inputs >> 2) & 0x01;
    info.rx_ctrl.stbc          = (legacy_inputs >> 3) & 0x01;
+    info.first_word_invalid    = (test_case & 0x80) != 0;

    /* Use remaining fuzz data as I/Q buffer content. */
    uint16_t iq_len;
@@ -136,6 +137,17 @@ int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size)
    if (result > out_len) {
        __builtin_trap();  /* Buffer overflow detected. */
    }
+    if (result >= CSI_HEADER_SIZE + iq_len && info.first_word_invalid && iq_len > 0) {
+        size_t invalid_len = iq_len < 4 ? iq_len : 4;
+        for (size_t i = 0; i < invalid_len; i++) {
+            if (out_buf[CSI_HEADER_SIZE + i] != 0) {
+                __builtin_trap();  /* Invalid hardware prefix escaped sanitation. */
+            }
+        }
+        if ((out_buf[19] & CSI_FLAG_FIRST_WORD_SANITIZED) == 0) {
+            __builtin_trap();  /* Sanitized payload must carry provenance. */
+        }
+    }

    /* --- Test case 1: NULL info pointer --- */
    if (test_case & 0x01) {
@@ -90,6 +90,7 @@ typedef struct {
 typedef struct {
    wifi_pkt_rx_ctrl_t rx_ctrl;
    uint8_t            mac[6];
+    bool               first_word_invalid;
    int16_t            len;     /**< Length of the I/Q buffer in bytes. */
    int8_t            *buf;     /**< Pointer to I/Q data. */
 } wifi_csi_info_t;
@@ -0,0 +1,13 @@
+#include <assert.h>
+#include <stdio.h>
+
+#include "c6_antenna_select.h"
+
+int main(void)
+{
+    assert(c6_xiao_antenna_gpio14_level(C6_XIAO_ANTENNA_INTERNAL) == 0);
+    assert(c6_xiao_antenna_gpio14_level(C6_XIAO_ANTENNA_EXTERNAL) == 1);
+
+    puts("PASS: XIAO ESP32-C6 antenna selection truth table");
+    return 0;
+}
@@ -0,0 +1,56 @@
+/** Deterministic host test for ESP-IDF first_word_invalid sanitation. */
+
+#include "esp_stubs.h"
+#include "nvs_config.h"
+#include "csi_collector.h"
+
+#include <assert.h>
+#include <stdio.h>
+#include <string.h>
+
+nvs_config_t g_nvs_config;
+
+static void make_info(wifi_csi_info_t *info, int8_t *iq, uint16_t len, bool invalid)
+{
+    memset(info, 0, sizeof(*info));
+    info->rx_ctrl.channel = 6;
+    info->rx_ctrl.rssi = -42;
+    info->rx_ctrl.noise_floor = -91;
+    info->len = (int16_t)len;
+    info->buf = iq;
+    info->first_word_invalid = invalid;
+}
+
+int main(void)
+{
+    int8_t iq[] = {1, 2, 3, 4, 5, 6, 7, 8};
+    uint8_t frame[CSI_MAX_FRAME_SIZE];
+    wifi_csi_info_t info;
+
+    make_info(&info, iq, sizeof(iq), true);
+    size_t len = csi_serialize_frame(&info, frame, sizeof(frame));
+    assert(len == CSI_HEADER_SIZE + sizeof(iq));
+    assert(frame[CSI_HEADER_SIZE] == 0);
+    assert(frame[CSI_HEADER_SIZE + 1] == 0);
+    assert(frame[CSI_HEADER_SIZE + 2] == 0);
+    assert(frame[CSI_HEADER_SIZE + 3] == 0);
+    assert(frame[CSI_HEADER_SIZE + 4] == 5);
+    assert((frame[19] & CSI_FLAG_FIRST_WORD_SANITIZED) != 0);
+
+    make_info(&info, iq, sizeof(iq), false);
+    len = csi_serialize_frame(&info, frame, sizeof(frame));
+    assert(len == CSI_HEADER_SIZE + sizeof(iq));
+    assert(memcmp(&frame[CSI_HEADER_SIZE], iq, sizeof(iq)) == 0);
+    assert((frame[19] & CSI_FLAG_FIRST_WORD_SANITIZED) == 0);
+
+    int8_t short_iq[] = {9, 10};
+    make_info(&info, short_iq, sizeof(short_iq), true);
+    len = csi_serialize_frame(&info, frame, sizeof(frame));
+    assert(len == CSI_HEADER_SIZE + sizeof(short_iq));
+    assert(frame[CSI_HEADER_SIZE] == 0);
+    assert(frame[CSI_HEADER_SIZE + 1] == 0);
+    assert((frame[19] & CSI_FLAG_FIRST_WORD_SANITIZED) != 0);
+
+    puts("CSI invalid-prefix sanitation: PASS");
+    return 0;
+}
@@ -1 +1 @@
-0.8.8
+0.8.10
@@ -145,13 +145,16 @@ impl PpduType {
 ///   bit 2   : STBC
 ///   bit 3   : LDPC (reserved — not yet populated by firmware)
 ///   bit 4   : 802.15.4 time-sync valid (C6 only)
-///   bit 5-7 : reserved
+///   bit 5   : ESP-IDF invalid first CSI word was zeroed by firmware
+///   bit 6-7 : reserved
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
 pub struct Adr018Flags {
    pub bw40: bool,
    pub stbc: bool,
    pub ldpc: bool,
    pub ieee802154_sync_valid: bool,
+    #[serde(default)]
+    pub first_word_sanitized: bool,
 }

 impl Adr018Flags {
@@ -161,6 +164,7 @@ impl Adr018Flags {
            stbc: (b & 0x04) != 0,
            ldpc: (b & 0x08) != 0,
            ieee802154_sync_valid: (b & 0x10) != 0,
+            first_word_sanitized: (b & 0x20) != 0,
        }
    }
    pub fn to_byte(self) -> u8 {
@@ -169,6 +173,7 @@ impl Adr018Flags {
        if self.stbc { b |= 0x04; }
        if self.ldpc { b |= 0x08; }
        if self.ieee802154_sync_valid { b |= 0x10; }
+        if self.first_word_sanitized { b |= 0x20; }
        b
    }
 }
@@ -413,15 +413,16 @@ mod tests {

    #[test]
    fn adr110_flags_round_trip_all_bits() {
-        // All known flag bits set: bw40 (0x01) + STBC (0x04) + LDPC (0x08) + 15.4-sync (0x10) = 0x1D
-        let data = build_test_frame_with_he(6, 1, &[(0, 0); 56], 1, 0x1D);
+        // All known flag bits set: bw40 + STBC + LDPC + sync + sanitized prefix = 0x3D.
+        let data = build_test_frame_with_he(6, 1, &[(0, 0); 56], 1, 0x3D);
        let (frame, _) = Esp32CsiParser::parse_frame(&data).unwrap();
        assert!(frame.metadata.adr018_flags.bw40);
        assert!(frame.metadata.adr018_flags.stbc);
        assert!(frame.metadata.adr018_flags.ldpc);
        assert!(frame.metadata.adr018_flags.ieee802154_sync_valid);
+        assert!(frame.metadata.adr018_flags.first_word_sanitized);
        // Round-trip the encoder
-        assert_eq!(frame.metadata.adr018_flags.to_byte(), 0x1D);
+        assert_eq!(frame.metadata.adr018_flags.to_byte(), 0x3D);
    }

    #[test]
@@ -30,7 +30,7 @@ mdns-sd = "0.11"

 # Serialization
 serde = { workspace = true }
-serde_json.workspace = true
+serde_json = { workspace = true, features = ["raw_value"] }

 # Logging
 tracing.workspace = true
@@ -43,6 +43,8 @@ per ADR-022 Phase 3.
 - **WebSocket broadcast** -- Real-time sensing updates pushed to all connected clients at
  `ws://localhost:8765/ws/sensing`.
 - **Static file serving** -- Hosts the sensing UI on port 8080 with CORS headers.
+- **Private startup baseline** -- Restores a recent, installation-bound empty-room field model
+  without storing raw CSI or authorizing numeric vital signs (ADR-355).

 ## Modules

@@ -74,6 +76,23 @@ cargo run -p wifi-densepose-sensing-server -- \
    --static-dir ./ui
 ```

+### Empty-room startup baseline
+
+Provide a stable installation ID and an application-owned private state directory:
+
+```bash
+cargo run -p wifi-densepose-sensing-server -- \
+    --installation-id installation-01 \
+    --data-dir /path/to/private/application-state
+```
+
+Complete the normal empty-room calibration, then call
+`POST /api/v1/calibration/bootstrap/promote`. The server measures a separate 12-sample holdout
+before storing an aggregate field-model snapshot. On a later start with the same installation ID,
+the snapshot is restored with `bootstrap_only` authority. It can reduce startup background false
+positives, but cannot authorize calibrated evidence or numeric heart and breathing rates. Use
+`POST /api/v1/calibration/reset` with administrator scope to remove it.
+
 ### Using as a library

 ```rust
@@ -139,6 +139,39 @@ pub fn features_from_runtime(feat: &serde_json::Value, amps: &[f64]) -> [f64; N_
    ]
 }

+/// Preserve the fifth feature's historical meaning for models trained before
+/// the server began measuring temporal frequency from a frame window.
+///
+/// Version 1 models learned a proxy equal to the strongest subcarrier index
+/// multiplied by 0.05. Version 2 and later models consume the measured
+/// temporal frequency in hertz. Keeping this translation at the model boundary
+/// prevents a server upgrade from silently changing an existing model's input
+/// distribution.
+pub fn compatible_dominant_frequency(
+    model_version: u32,
+    temporal_frequency_hz: f64,
+    amplitudes: &[f64],
+) -> f64 {
+    if model_version >= 2 {
+        return if temporal_frequency_hz.is_finite() {
+            temporal_frequency_hz.max(0.0)
+        } else {
+            0.0
+        };
+    }
+
+    amplitudes
+        .iter()
+        .enumerate()
+        .filter(|(_, amplitude)| amplitude.is_finite())
+        .max_by(|(_, left), (_, right)| {
+            left.partial_cmp(right)
+                .unwrap_or(std::cmp::Ordering::Equal)
+        })
+        .map(|(index, _)| index as f64 * 0.05)
+        .unwrap_or(0.0)
+}
+
 /// Compute statistical features from raw subcarrier amplitudes.
 fn subcarrier_stats(amps: &[f64]) -> (f64, f64, f64, f64, f64, f64, f64, f64) {
    if amps.is_empty() {
@@ -642,3 +675,21 @@ pub fn train_from_recordings(recordings_dir: &Path) -> Result<AdaptiveModel, Str
 pub fn model_path() -> PathBuf {
    PathBuf::from("data/adaptive_model.json")
 }
+
+#[cfg(test)]
+mod compatibility_tests {
+    use super::compatible_dominant_frequency;
+
+    #[test]
+    fn version_one_keeps_legacy_subcarrier_proxy() {
+        let amplitudes = [1.0, 2.0, 8.0, 3.0];
+        assert_eq!(compatible_dominant_frequency(1, 1.25, &amplitudes), 0.10);
+    }
+
+    #[test]
+    fn version_two_uses_measured_temporal_frequency() {
+        let amplitudes = [1.0, 20.0, 2.0];
+        assert_eq!(compatible_dominant_frequency(2, 1.25, &amplitudes), 1.25);
+        assert_eq!(compatible_dominant_frequency(2, f64::NAN, &amplitudes), 0.0);
+    }
+}
@@ -491,6 +491,10 @@ const READ_SAFE_MUTATIONS: &[&str] = &[
    // Capture and calibration: create data, never destroy it.
    "/api/v1/calibration/start",
    "/api/v1/calibration/stop",
+    // Promotion consumes only server measured held out evidence and stores a
+    // bounded privacy reduced prior. Cancel discards only an unfinished run.
+    "/api/v1/calibration/bootstrap/promote",
+    "/api/v1/calibration/cancel",
    "/api/v1/pose/calibrate",
    "/api/v1/recording/start",
    "/api/v1/recording/stop",
@@ -1970,6 +1974,7 @@ mod scope_gate_polarity_tests {
            (Method::DELETE, "/api/v1/models/m1"),
            (Method::DELETE, "/api/v1/recording/r1"),
            (Method::POST, "/api/v1/config/ground-truth"),
+            (Method::POST, "/api/v1/calibration/reset"),
        ] {
            assert_eq!(
                required_scope_for(&m, p),
@@ -0,0 +1,532 @@
+//! Local, privacy reduced empty room bootstrap model persistence.
+//!
+//! A bootstrap image is deliberately lower authority than a completed room
+//! calibration. It can suppress background perturbations while the server
+//! starts, but restoring it never authorizes calibrated evidence or vitals.
+
+use rand::{rngs::OsRng, RngCore};
+use serde::{Deserialize, Serialize};
+use serde_json::value::RawValue;
+use sha2::{Digest, Sha256};
+use std::fs::{self, OpenOptions};
+use std::io::{Read, Write};
+use std::path::{Path, PathBuf};
+use thiserror::Error;
+use wifi_densepose_signal::ruvsense::field_model::{
+    FieldModel, FieldModelError, FieldModelSnapshotV1,
+};
+
+pub const BOOTSTRAP_BASELINE_SCHEMA: &str = "ruview.bootstrap-empty-field-model.v2";
+pub const BOOTSTRAP_BASELINE_AUTHORITY: &str = "bootstrap_only";
+pub const BOOTSTRAP_VALIDATION_SAMPLES: usize = 12;
+pub const BOOTSTRAP_VALIDATION_MIN_EMPTY: usize = 10;
+pub const BOOTSTRAP_VALIDATION_MIN_SPACING_MS: u64 = 1_000;
+pub const BOOTSTRAP_VALIDATION_SAMPLE_TIMEOUT_MS: u64 = 4_000;
+const BOOTSTRAP_MAX_FILE_BYTES: u64 = 1_048_576;
+
+/// Exact CSI layout used to train the persisted empty room model.
+///
+/// Restoring code must require incoming frames to match this binding before
+/// the bootstrap image can influence inference.
+#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
+#[serde(deny_unknown_fields)]
+pub struct BootstrapCsiGrid {
+    pub n_subcarriers: u16,
+    pub ppdu_type: u8,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(deny_unknown_fields)]
+struct BootstrapPayloadV2 {
+    schema: String,
+    authority: String,
+    installation_binding_sha256: String,
+    source_node_ids: Vec<u8>,
+    source_grid: BootstrapCsiGrid,
+    source_model_id: String,
+    created_at_unix_ms: u64,
+    expires_at_unix_ms: u64,
+    field_model: FieldModelSnapshotV1,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(deny_unknown_fields)]
+struct BootstrapImageV2 {
+    payload: BootstrapPayloadV2,
+    content_sha256: String,
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(deny_unknown_fields)]
+struct BootstrapImageRawV2 {
+    payload: Box<RawValue>,
+    content_sha256: String,
+}
+
+#[derive(Debug, Clone, Serialize, PartialEq, Eq)]
+pub struct BootstrapBaselineMetadata {
+    pub authority: &'static str,
+    pub source_node_ids: Vec<u8>,
+    pub source_grid: BootstrapCsiGrid,
+    pub source_model_id: String,
+    pub created_at_unix_ms: u64,
+    pub expires_at_unix_ms: u64,
+    pub content_sha256: String,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct BootstrapValidationSample {
+    pub fresh_tick: bool,
+    pub calibrated_empty: bool,
+    pub vital_signs_absent: bool,
+}
+
+#[derive(Debug, Clone, Copy, Serialize, PartialEq, Eq)]
+pub struct BootstrapValidationResult {
+    pub sample_count: usize,
+    pub empty_sample_count: usize,
+    pub stale_sample_count: usize,
+    pub vital_sign_sample_count: usize,
+    pub passed: bool,
+}
+
+#[derive(Debug, Error)]
+pub enum BootstrapBaselineError {
+    #[error("a stable installation_id is required for a local bootstrap baseline")]
+    MissingInstallationId,
+    #[error("bootstrap baseline path is invalid")]
+    InvalidPath,
+    #[error("bootstrap baseline file is too large")]
+    FileTooLarge,
+    #[error("bootstrap baseline image is malformed: {0}")]
+    Malformed(String),
+    #[error("bootstrap baseline belongs to another installation")]
+    InstallationMismatch,
+    #[error("bootstrap baseline has expired")]
+    Expired,
+    #[error("bootstrap baseline digest does not verify")]
+    DigestMismatch,
+    #[error("field model snapshot is invalid: {0}")]
+    FieldModel(#[from] FieldModelError),
+    #[error("bootstrap baseline storage failed: {0}")]
+    Io(#[from] std::io::Error),
+    #[error("bootstrap baseline serialization failed: {0}")]
+    Json(#[from] serde_json::Error),
+}
+
+pub fn path_in(data_dir: &Path) -> PathBuf {
+    data_dir.join("calibration").join("bootstrap-empty-v1.json")
+}
+
+pub fn installation_binding(installation_id: &str) -> Result<String, BootstrapBaselineError> {
+    let installation_id = installation_id.trim();
+    if installation_id.is_empty() || installation_id.len() > 128 {
+        return Err(BootstrapBaselineError::MissingInstallationId);
+    }
+    Ok(hex_sha256(installation_id.as_bytes()))
+}
+
+pub fn evaluate_validation(samples: &[BootstrapValidationSample]) -> BootstrapValidationResult {
+    let empty_sample_count = samples
+        .iter()
+        .filter(|sample| sample.fresh_tick && sample.calibrated_empty)
+        .count();
+    let stale_sample_count = samples.iter().filter(|sample| !sample.fresh_tick).count();
+    let vital_sign_sample_count = samples
+        .iter()
+        .filter(|sample| !sample.vital_signs_absent)
+        .count();
+    BootstrapValidationResult {
+        sample_count: samples.len(),
+        empty_sample_count,
+        stale_sample_count,
+        vital_sign_sample_count,
+        passed: samples.len() == BOOTSTRAP_VALIDATION_SAMPLES
+            && empty_sample_count >= BOOTSTRAP_VALIDATION_MIN_EMPTY
+            && stale_sample_count == 0
+            && vital_sign_sample_count == 0,
+    }
+}
+
+pub fn store(
+    path: &Path,
+    installation_id: &str,
+    source_node_ids: &[u8],
+    source_grid: BootstrapCsiGrid,
+    source_model_id: &str,
+    created_at_unix_ms: u64,
+    field_model: &FieldModel,
+) -> Result<BootstrapBaselineMetadata, BootstrapBaselineError> {
+    let snapshot = field_model.export_snapshot()?;
+    let calibrated_at_ms = snapshot.modes.calibrated_at_us / 1_000;
+    let expiry_ms = (snapshot.config.baseline_expiry_s * 1_000.0) as u64;
+    let payload = BootstrapPayloadV2 {
+        schema: BOOTSTRAP_BASELINE_SCHEMA.to_string(),
+        authority: BOOTSTRAP_BASELINE_AUTHORITY.to_string(),
+        installation_binding_sha256: installation_binding(installation_id)?,
+        source_node_ids: source_node_ids.to_vec(),
+        source_grid,
+        source_model_id: source_model_id.to_string(),
+        created_at_unix_ms,
+        expires_at_unix_ms: calibrated_at_ms.saturating_add(expiry_ms),
+        field_model: snapshot,
+    };
+    validate_payload(&payload, installation_id, created_at_unix_ms)?;
+    let content_sha256 = payload_digest(&payload)?;
+    let image = BootstrapImageV2 {
+        payload,
+        content_sha256,
+    };
+    let bytes = serde_json::to_vec(&image)?;
+    if bytes.len() as u64 > BOOTSTRAP_MAX_FILE_BYTES {
+        return Err(BootstrapBaselineError::FileTooLarge);
+    }
+    atomic_write_private(path, &bytes)?;
+    Ok(metadata(&image))
+}
+
+pub fn load(
+    path: &Path,
+    installation_id: &str,
+    current_unix_ms: u64,
+) -> Result<(FieldModel, BootstrapBaselineMetadata), BootstrapBaselineError> {
+    let file_metadata = fs::symlink_metadata(path)?;
+    if file_metadata.file_type().is_symlink() || !file_metadata.is_file() {
+        return Err(BootstrapBaselineError::InvalidPath);
+    }
+    if file_metadata.len() > BOOTSTRAP_MAX_FILE_BYTES {
+        return Err(BootstrapBaselineError::FileTooLarge);
+    }
+    let mut bytes = Vec::with_capacity(file_metadata.len() as usize);
+    OpenOptions::new()
+        .read(true)
+        .open(path)?
+        .take(BOOTSTRAP_MAX_FILE_BYTES + 1)
+        .read_to_end(&mut bytes)?;
+    if bytes.len() as u64 > BOOTSTRAP_MAX_FILE_BYTES {
+        return Err(BootstrapBaselineError::FileTooLarge);
+    }
+    // Verify the exact payload bytes written by storage before decoding the
+    // floating point model. Equivalent JSON number spellings may serialize
+    // differently, so decode and reserialize is not an integrity contract.
+    let raw_image: BootstrapImageRawV2 = serde_json::from_slice(&bytes)?;
+    if hex_sha256(raw_image.payload.get().as_bytes()) != raw_image.content_sha256 {
+        return Err(BootstrapBaselineError::DigestMismatch);
+    }
+    let payload: BootstrapPayloadV2 = serde_json::from_str(raw_image.payload.get())?;
+    validate_payload(&payload, installation_id, current_unix_ms)?;
+    let image = BootstrapImageV2 {
+        payload,
+        content_sha256: raw_image.content_sha256,
+    };
+    let model = FieldModel::from_snapshot(
+        image.payload.field_model.clone(),
+        current_unix_ms.saturating_mul(1_000),
+    )?;
+    Ok((model, metadata(&image)))
+}
+
+pub fn remove(path: &Path) -> Result<bool, BootstrapBaselineError> {
+    match fs::symlink_metadata(path) {
+        Ok(metadata) => {
+            if metadata.file_type().is_symlink() || !metadata.is_file() {
+                return Err(BootstrapBaselineError::InvalidPath);
+            }
+            fs::remove_file(path)?;
+            Ok(true)
+        }
+        Err(error) if error.kind() == std::io::ErrorKind::NotFound => Ok(false),
+        Err(error) => Err(error.into()),
+    }
+}
+
+fn validate_payload(
+    payload: &BootstrapPayloadV2,
+    installation_id: &str,
+    current_unix_ms: u64,
+) -> Result<(), BootstrapBaselineError> {
+    if payload.schema != BOOTSTRAP_BASELINE_SCHEMA
+        || payload.authority != BOOTSTRAP_BASELINE_AUTHORITY
+        || payload.source_model_id.is_empty()
+        || payload.source_model_id.len() > 128
+        || payload.source_node_ids.len() != 1
+        || !(1..=2_048).contains(&payload.source_grid.n_subcarriers)
+        || payload.source_grid.ppdu_type > 3
+        || payload.created_at_unix_ms == 0
+        || payload.expires_at_unix_ms <= payload.created_at_unix_ms
+    {
+        return Err(BootstrapBaselineError::Malformed(
+            "invalid schema, authority, identity, node set, source grid, or lifetime".into(),
+        ));
+    }
+    if payload.installation_binding_sha256 != installation_binding(installation_id)? {
+        return Err(BootstrapBaselineError::InstallationMismatch);
+    }
+    if current_unix_ms > payload.expires_at_unix_ms {
+        return Err(BootstrapBaselineError::Expired);
+    }
+    Ok(())
+}
+
+fn payload_digest(payload: &BootstrapPayloadV2) -> Result<String, serde_json::Error> {
+    Ok(hex_sha256(&serde_json::to_vec(payload)?))
+}
+
+fn metadata(image: &BootstrapImageV2) -> BootstrapBaselineMetadata {
+    BootstrapBaselineMetadata {
+        authority: BOOTSTRAP_BASELINE_AUTHORITY,
+        source_node_ids: image.payload.source_node_ids.clone(),
+        source_grid: image.payload.source_grid,
+        source_model_id: image.payload.source_model_id.clone(),
+        created_at_unix_ms: image.payload.created_at_unix_ms,
+        expires_at_unix_ms: image.payload.expires_at_unix_ms,
+        content_sha256: image.content_sha256.clone(),
+    }
+}
+
+fn hex_sha256(bytes: &[u8]) -> String {
+    let digest = Sha256::digest(bytes);
+    digest.iter().map(|byte| format!("{byte:02x}")).collect()
+}
+
+fn atomic_write_private(path: &Path, bytes: &[u8]) -> Result<(), BootstrapBaselineError> {
+    let parent = path.parent().ok_or(BootstrapBaselineError::InvalidPath)?;
+    fs::create_dir_all(parent)?;
+    let file_name = path
+        .file_name()
+        .and_then(|name| name.to_str())
+        .ok_or(BootstrapBaselineError::InvalidPath)?;
+    let mut nonce = [0_u8; 8];
+    OsRng.fill_bytes(&mut nonce);
+    let temporary = parent.join(format!(
+        ".{file_name}.{:016x}.tmp",
+        u64::from_le_bytes(nonce)
+    ));
+
+    let mut options = OpenOptions::new();
+    options.write(true).create_new(true);
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::OpenOptionsExt;
+        options.mode(0o600);
+    }
+    let write_result = (|| -> Result<(), std::io::Error> {
+        let mut file = options.open(&temporary)?;
+        file.write_all(bytes)?;
+        file.sync_all()?;
+        fs::rename(&temporary, path)?;
+        Ok(())
+    })();
+    if write_result.is_err() {
+        let _ = fs::remove_file(&temporary);
+    }
+    write_result.map_err(Into::into)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use wifi_densepose_signal::ruvsense::field_model::FieldModelConfig;
+
+    const TEST_GRID: BootstrapCsiGrid = BootstrapCsiGrid {
+        n_subcarriers: 192,
+        ppdu_type: 2,
+    };
+
+    fn completed_model(now_us: u64) -> FieldModel {
+        let mut model = FieldModel::new(FieldModelConfig {
+            n_links: 1,
+            n_subcarriers: 4,
+            n_modes: 2,
+            min_calibration_frames: 3,
+            min_calibration_duration_s: 0.0,
+            baseline_expiry_s: 3_600.0,
+        })
+        .unwrap();
+        for index in 0..3 {
+            model
+                .feed_calibration(&[vec![1.0 + index as f64, 2.0, 3.0, 4.0]])
+                .unwrap();
+        }
+        model.finalize_calibration(now_us, 7).unwrap();
+        model
+    }
+
+    #[test]
+    fn round_trip_is_installation_bound_and_contains_no_raw_frames() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = path_in(dir.path());
+        let now_ms = 1_000_000;
+        let model = completed_model(now_ms * 1_000);
+        let stored = store(&path, "home-a", &[5], TEST_GRID, "model", now_ms, &model).unwrap();
+        let text = String::from_utf8(fs::read(&path).unwrap()).unwrap();
+        assert!(!text.contains("home-a"));
+        assert!(!text.contains("raw_csi"));
+        assert_eq!(stored.source_grid, TEST_GRID);
+
+        let (restored, loaded) = load(&path, "home-a", now_ms + 1).unwrap();
+        assert_eq!(stored, loaded);
+        assert_eq!(
+            restored.export_snapshot().unwrap(),
+            model.export_snapshot().unwrap()
+        );
+        assert!(matches!(
+            load(&path, "home-b", now_ms + 1),
+            Err(BootstrapBaselineError::InstallationMismatch)
+        ));
+    }
+
+    #[test]
+    fn tamper_and_expiry_fail_closed() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = path_in(dir.path());
+        let now_ms = 1_000_000;
+        let model = completed_model(now_ms * 1_000);
+        store(&path, "home-a", &[5], TEST_GRID, "model", now_ms, &model).unwrap();
+
+        let mut image: serde_json::Value =
+            serde_json::from_slice(&fs::read(&path).unwrap()).unwrap();
+        image["payload"]["source_grid"]["n_subcarriers"] = serde_json::json!(64);
+        fs::write(&path, serde_json::to_vec(&image).unwrap()).unwrap();
+        assert!(matches!(
+            load(&path, "home-a", now_ms + 1),
+            Err(BootstrapBaselineError::DigestMismatch)
+        ));
+
+        store(&path, "home-a", &[5], TEST_GRID, "model", now_ms, &model).unwrap();
+        assert!(matches!(
+            load(&path, "home-a", now_ms + 3_600_001),
+            Err(BootstrapBaselineError::Expired)
+                | Err(BootstrapBaselineError::FieldModel(
+                    FieldModelError::BaselineExpired { .. }
+                ))
+        ));
+    }
+
+    #[test]
+    fn store_rejects_invalid_source_grid() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = path_in(dir.path());
+        let now_ms = 1_000_000;
+        let model = completed_model(now_ms * 1_000);
+
+        for source_grid in [
+            BootstrapCsiGrid {
+                n_subcarriers: 0,
+                ppdu_type: 0,
+            },
+            BootstrapCsiGrid {
+                n_subcarriers: 2_049,
+                ppdu_type: 0,
+            },
+            BootstrapCsiGrid {
+                n_subcarriers: 192,
+                ppdu_type: 4,
+            },
+        ] {
+            assert!(matches!(
+                store(&path, "home-a", &[5], source_grid, "model", now_ms, &model,),
+                Err(BootstrapBaselineError::Malformed(_))
+            ));
+        }
+    }
+
+    #[test]
+    fn store_requires_exactly_one_source_node() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = path_in(dir.path());
+        let now_ms = 1_000_000;
+        let model = completed_model(now_ms * 1_000);
+
+        for source_node_ids in [&[][..], &[5, 7][..]] {
+            assert!(matches!(
+                store(
+                    &path,
+                    "home-a",
+                    source_node_ids,
+                    TEST_GRID,
+                    "model",
+                    now_ms,
+                    &model,
+                ),
+                Err(BootstrapBaselineError::Malformed(_))
+            ));
+        }
+    }
+
+    #[test]
+    fn source_grid_is_required_when_loading_a_v2_image() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = path_in(dir.path());
+        let now_ms = 1_000_000;
+        let model = completed_model(now_ms * 1_000);
+        store(&path, "home-a", &[5], TEST_GRID, "model", now_ms, &model).unwrap();
+
+        let mut image: serde_json::Value =
+            serde_json::from_slice(&fs::read(&path).unwrap()).unwrap();
+        image["payload"]
+            .as_object_mut()
+            .unwrap()
+            .remove("source_grid");
+        let payload = serde_json::to_vec(&image["payload"]).unwrap();
+        let rewritten = format!(
+            "{{\"payload\":{},\"content_sha256\":\"{}\"}}",
+            String::from_utf8(payload.clone()).unwrap(),
+            hex_sha256(&payload)
+        );
+        fs::write(&path, rewritten).unwrap();
+
+        assert!(matches!(
+            load(&path, "home-a", now_ms + 1),
+            Err(BootstrapBaselineError::Json(_))
+        ));
+    }
+
+    #[test]
+    fn promotion_gate_requires_twelve_fresh_samples_and_zero_vitals() {
+        assert_eq!(BOOTSTRAP_VALIDATION_MIN_SPACING_MS, 1_000);
+        assert_eq!(BOOTSTRAP_VALIDATION_SAMPLE_TIMEOUT_MS, 4_000);
+        assert!(BOOTSTRAP_VALIDATION_SAMPLE_TIMEOUT_MS >= BOOTSTRAP_VALIDATION_MIN_SPACING_MS);
+        let passing = vec![
+            BootstrapValidationSample {
+                fresh_tick: true,
+                calibrated_empty: true,
+                vital_signs_absent: true,
+            };
+            BOOTSTRAP_VALIDATION_SAMPLES
+        ];
+        assert!(evaluate_validation(&passing).passed);
+
+        let mut stale = passing.clone();
+        stale[0].fresh_tick = false;
+        assert!(!evaluate_validation(&stale).passed);
+        let mut vital = passing;
+        vital[0].vital_signs_absent = false;
+        assert!(!evaluate_validation(&vital).passed);
+    }
+
+    #[test]
+    fn restart_verifies_the_exact_stored_payload_bytes() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = path_in(dir.path());
+        let now_ms = 1_000_000;
+        let model = completed_model(now_ms * 1_000);
+        store(&path, "home-a", &[5], TEST_GRID, "model", now_ms, &model).unwrap();
+
+        let original = fs::read_to_string(&path).unwrap();
+        let image: BootstrapImageRawV2 = serde_json::from_str(&original).unwrap();
+        let rewritten_payload = image.payload.get().replacen(
+            "\"min_calibration_duration_s\":0.0",
+            "\"min_calibration_duration_s\":0.00",
+            1,
+        );
+        assert_ne!(rewritten_payload, image.payload.get());
+        let digest = hex_sha256(rewritten_payload.as_bytes());
+        let rewritten =
+            format!("{{\"payload\":{rewritten_payload},\"content_sha256\":\"{digest}\"}}");
+        fs::write(&path, rewritten).unwrap();
+
+        let (_, loaded) = load(&path, "home-a", now_ms + 1).unwrap();
+        assert_eq!(loaded.content_sha256, digest);
+    }
+}
@@ -178,7 +178,12 @@ pub struct Args {
    #[arg(long, value_name = "TYPE")]
    pub build_index: Option<String>,

-    /// Node positions for multistatic fusion (format: "x,y,z;x,y,z;...")
+    /// Node positions for multistatic fusion.
+    ///
+    /// Two forms, mixable: "x,y,z;..." positional, where each entry's identity
+    /// is its index, and "node_id:x,y,z;..." explicit. Prefer the explicit form
+    /// unless the fleet is numbered 0,1,2,... — otherwise "the third entry" and
+    /// "node 3" are different nodes and nothing in the string says which is meant.
    #[arg(long, env = "SENSING_NODE_POSITIONS")]
    pub node_positions: Option<String>,

@@ -11,6 +11,141 @@ use crate::vital_signs::VitalSigns;

 const EDGE_MAX_PERSONS: u8 = 4;

+/// Estimate a real temporal motion frequency from recent CSI amplitude
+/// history. Each frame is reduced to its mean amplitude, detrended, Hann
+/// windowed, and evaluated on exact DFT bins.
+pub fn estimate_temporal_dominant_frequency_hz(
+    frame_history: &VecDeque<Vec<f64>>,
+    sample_rate_hz: f64,
+) -> f64 {
+    if !sample_rate_hz.is_finite() || sample_rate_hz <= 0.0 {
+        return 0.0;
+    }
+
+    let recent_frames: Vec<&Vec<f64>> = frame_history
+        .iter()
+        .rev()
+        .take(128)
+        .collect();
+    if recent_frames
+        .iter()
+        .any(|frame| frame.is_empty() || frame.iter().any(|value| !value.is_finite()))
+    {
+        return 0.0;
+    }
+    let samples: Vec<f64> = recent_frames
+        .into_iter()
+        .rev()
+        .map(|frame| frame.iter().sum::<f64>() / frame.len() as f64)
+        .collect();
+    let n = samples.len();
+    if n < 16 {
+        return 0.0;
+    }
+
+    let mean = samples.iter().sum::<f64>() / n as f64;
+    let upper_hz = 4.0_f64.min(sample_rate_hz * 0.45);
+    if upper_hz < 0.1 {
+        return 0.0;
+    }
+
+    let mut powers = Vec::new();
+    for bin in 1..=(n / 2) {
+        let frequency_hz = bin as f64 * sample_rate_hz / n as f64;
+        if !(0.1..=upper_hz).contains(&frequency_hz) {
+            continue;
+        }
+        let mut real = 0.0;
+        let mut imag = 0.0;
+        let angle_step = 2.0 * std::f64::consts::PI * bin as f64 / n as f64;
+        let (step_sin, step_cos) = angle_step.sin_cos();
+        let mut angle_cos = 1.0;
+        let mut angle_sin = 0.0;
+        for (index, sample) in samples.iter().enumerate() {
+            let window =
+                0.5 * (1.0 - (2.0 * std::f64::consts::PI * index as f64 / (n - 1) as f64).cos());
+            let value = (sample - mean) * window;
+            real += value * angle_cos;
+            imag -= value * angle_sin;
+            let next_cos = angle_cos * step_cos - angle_sin * step_sin;
+            angle_sin = angle_sin * step_cos + angle_cos * step_sin;
+            angle_cos = next_cos;
+        }
+        powers.push((frequency_hz, real * real + imag * imag));
+    }
+    if powers.is_empty() {
+        return 0.0;
+    }
+    let mean_power = powers.iter().map(|(_, power)| power).sum::<f64>() / powers.len() as f64;
+    let (peak_hz, peak_power) = powers
+        .into_iter()
+        .max_by(|left, right| {
+            left.1
+                .partial_cmp(&right.1)
+                .unwrap_or(std::cmp::Ordering::Equal)
+        })
+        .unwrap_or((0.0, 0.0));
+    if mean_power <= f64::EPSILON || peak_power / mean_power < 2.0 {
+        0.0
+    } else {
+        peak_hz
+    }
+}
+
+#[cfg(test)]
+mod temporal_frequency_tests {
+    use super::*;
+
+    #[test]
+    fn recovers_temporal_frequency_in_hertz() {
+        let sample_rate_hz = 20.0;
+        let target_hz = 1.2;
+        let mut history = VecDeque::new();
+        for sample in 0..100 {
+            let t = sample as f64 / sample_rate_hz;
+            let value = 20.0 + 2.0 * (2.0 * std::f64::consts::PI * target_hz * t).sin();
+            history.push_back(vec![value; 56]);
+        }
+        let measured = estimate_temporal_dominant_frequency_hz(&history, sample_rate_hz);
+        assert!((measured - target_hz).abs() <= 0.21, "measured {measured} Hz");
+    }
+
+    #[test]
+    fn abstains_on_short_or_flat_history() {
+        let short = VecDeque::from(vec![vec![20.0; 56]; 15]);
+        assert_eq!(estimate_temporal_dominant_frequency_hz(&short, 20.0), 0.0);
+        let flat = VecDeque::from(vec![vec![20.0; 56]; 100]);
+        assert_eq!(estimate_temporal_dominant_frequency_hz(&flat, 20.0), 0.0);
+        let mut corrupt = flat;
+        corrupt[50][4] = f64::NAN;
+        assert_eq!(estimate_temporal_dominant_frequency_hz(&corrupt, 20.0), 0.0);
+    }
+
+    #[test]
+    fn temporal_frequency_p95_stays_below_server_budget() {
+        let sample_rate_hz = 50.0;
+        let mut history = VecDeque::new();
+        for sample in 0..128 {
+            let t = sample as f64 / sample_rate_hz;
+            let value = 20.0 + 2.0 * (2.0 * std::f64::consts::PI * 1.2 * t).sin();
+            history.push_back(vec![value; 256]);
+        }
+        let mut timings = Vec::with_capacity(1_000);
+        for _ in 0..1_000 {
+            let start = std::time::Instant::now();
+            std::hint::black_box(estimate_temporal_dominant_frequency_hz(
+                std::hint::black_box(&history),
+                sample_rate_hz,
+            ));
+            timings.push(start.elapsed());
+        }
+        timings.sort_unstable();
+        let p95 = timings[949];
+        eprintln!("temporal dominant frequency p95: {p95:?}");
+        assert!(p95 < std::time::Duration::from_millis(10), "p95 {p95:?}");
+    }
+}
+
 /// Person count is supporting evidence, never an independent occupancy claim.
 /// Return zero and mark invalid for contradictory or out-of-range firmware.
 fn sanitize_edge_person_count(presence: bool, raw: u8) -> (u8, bool) {
@@ -469,14 +604,7 @@ pub fn extract_features_from_frame(
        0.0
    };

-    let peak_idx = frame
-        .amplitudes
-        .iter()
-        .enumerate()
-        .max_by(|a, b| a.1.partial_cmp(b.1).unwrap_or(std::cmp::Ordering::Equal))
-        .map(|(i, _)| i)
-        .unwrap_or(0);
-    let dominant_freq_hz = peak_idx as f64 * 0.05;
+    let dominant_freq_hz = estimate_temporal_dominant_frequency_hz(frame_history, sample_rate_hz);

    let threshold = mean_amp * 1.2;
    let change_points = frame
@@ -485,7 +613,9 @@ pub fn extract_features_from_frame(
        .filter(|w| (w[0] < threshold) != (w[1] < threshold))
        .count();

-    let temporal_motion_score = if let Some(prev_frame) = frame_history.back() {
+    // The caller appends the current frame before extraction, so the true
+    // predecessor is the second item from the back.
+    let temporal_motion_score = if let Some(prev_frame) = frame_history.iter().rev().nth(1) {
        let n_cmp = n_sub.min(prev_frame.len());
        if n_cmp > 0 {
            let diff_energy: f64 = (0..n_cmp)
@@ -636,13 +766,18 @@ pub fn adaptive_override(
            .back()
            .map(|v| v.as_slice())
            .unwrap_or(&[]);
+        let dominant_frequency = adaptive_classifier::compatible_dominant_frequency(
+            model.version,
+            features.dominant_freq_hz,
+            amps,
+        );
        let feat_arr = adaptive_classifier::features_from_runtime(
            &serde_json::json!({
                "variance": features.variance,
                "motion_band_power": features.motion_band_power,
                "breathing_band_power": features.breathing_band_power,
                "spectral_power": features.spectral_power,
-                "dominant_freq_hz": features.dominant_freq_hz,
+                "dominant_freq_hz": dominant_frequency,
                "change_points": features.change_points,
                "mean_rssi": features.mean_rssi,
            }),
@@ -40,6 +40,106 @@ const ENERGY_THRESH_3: f64 = 25.0;
 /// reliable — genuine higher counts come from the multistatic fusion path.
 const MAX_SINGLE_LINK_OCCUPANCY: usize = 3;

+/// Low-authority comparison of the current source window with a restored
+/// empty-room bootstrap image. This is background conformance only and cannot
+/// authorize presence, person count, pose, identity, or vital signs.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub struct BootstrapBackgroundMatch {
+    pub matches_empty: bool,
+    pub score: Option<f64>,
+    pub normalized_residual_z: Option<f64>,
+    pub maturity: f64,
+    pub reliable: bool,
+    pub residual_energy: f64,
+    pub residual_energy_threshold: f64,
+    pub window_size: usize,
+    pub reference_window_count: usize,
+}
+
+fn mean_frame(frames: &[Vec<f64>]) -> Option<Vec<f64>> {
+    let width = frames.first()?.len();
+    if width == 0 {
+        return None;
+    }
+    let mut mean = vec![0.0_f64; width];
+    let mut count = 0_usize;
+    for frame in frames {
+        if frame.len() != width {
+            continue;
+        }
+        for (sum, value) in mean.iter_mut().zip(frame.iter()) {
+            *sum += value;
+        }
+        count += 1;
+    }
+    if count == 0 {
+        return None;
+    }
+    for value in &mut mean {
+        *value /= count as f64;
+    }
+    Some(mean)
+}
+
+fn perturbation_occupancy(field: &FieldModel, frames: &[Vec<f64>]) -> Option<usize> {
+    let adaptive_threshold = field.empty_room_residual_energy_threshold();
+    let frame = match adaptive_threshold {
+        Some(_) => mean_frame(frames)?,
+        None => frames.first()?.clone(),
+    };
+    let perturbation = field.extract_perturbation(&[frame]).ok()?;
+    let count = match adaptive_threshold {
+        Some(threshold) => {
+            let empty_threshold = threshold.max(1.0);
+            if perturbation.total_energy > empty_threshold * 25.0 {
+                3
+            } else if perturbation.total_energy > empty_threshold * 12.0 {
+                2
+            } else if perturbation.total_energy > empty_threshold {
+                1
+            } else {
+                0
+            }
+        }
+        None => {
+            if perturbation.total_energy > ENERGY_THRESH_3 {
+                3
+            } else if perturbation.total_energy > ENERGY_THRESH_2 {
+                2
+            } else if perturbation.total_energy > 1.0 {
+                1
+            } else {
+                0
+            }
+        }
+    };
+    Some(count)
+}
+
+/// Provenance for a count produced by the calibrated field model. Callers
+/// that attach calibration evidence must never substitute the score heuristic
+/// when the calibrated path cannot evaluate the observation.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum CalibratedOccupancyMethod {
+    Eigenvalue,
+    PerturbationEnergy,
+}
+
+impl CalibratedOccupancyMethod {
+    pub const fn wire_name(self) -> &'static str {
+        match self {
+            Self::Eigenvalue => "field_model_eigenvalue_v1",
+            Self::PerturbationEnergy => "field_model_perturbation_energy_v1",
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct CalibratedOccupancy {
+    pub person_count: usize,
+    pub method: CalibratedOccupancyMethod,
+}
+
 /// Create a FieldModelConfig for single-link mode (one ESP32 node = one link).
 /// This avoids the DimensionMismatch error when feeding single-frame observations.
 pub fn single_link_config() -> FieldModelConfig {
@@ -49,6 +149,18 @@ pub fn single_link_config() -> FieldModelConfig {
    }
 }

+/// Resolve the model status at the observation wall clock. `FieldModel::status`
+/// records the state at collection or restore time and does not advance as a
+/// long-running server crosses the stale and expiry boundaries.
+pub fn calibration_status_at(field: &FieldModel, observed_at_us: u64) -> CalibrationStatus {
+    match field.status() {
+        CalibrationStatus::Uncalibrated | CalibrationStatus::Collecting => field.status(),
+        CalibrationStatus::Fresh | CalibrationStatus::Stale | CalibrationStatus::Expired => {
+            field.check_freshness(observed_at_us)
+        }
+    }
+}
+
 /// Estimate occupancy using the FieldModel when calibrated, falling back
 /// to the score-based heuristic otherwise.
 ///
@@ -58,16 +170,17 @@ pub fn single_link_config() -> FieldModelConfig {
 pub fn occupancy_or_fallback(
    field: &FieldModel,
    frame_history: &VecDeque<Vec<f64>>,
+    observed_at_us: u64,
    smoothed_score: f64,
    prev_count: usize,
 ) -> usize {
-    match field.status() {
+    match calibration_status_at(field, observed_at_us) {
        CalibrationStatus::Fresh | CalibrationStatus::Stale => {
            let frames: Vec<Vec<f64>> = frame_history
                .iter()
                .rev()
                .take(OCCUPANCY_WINDOW)
-                .cloned()
+                .map(|frame| CALIB_NORMALIZER.resample_to_canonical(frame))
                .collect();

            if frames.is_empty() {
@@ -87,82 +200,197 @@ pub fn occupancy_or_fallback(

            // Fallback: perturbation energy thresholds.
            // FieldModel expects [n_links][n_subcarriers] — we use n_links=1.
-            let observation = vec![frames[0].clone()];
-            match field.extract_perturbation(&observation) {
-                Ok(perturbation) => {
-                    if perturbation.total_energy > ENERGY_THRESH_3 {
-                        3
-                    } else if perturbation.total_energy > ENERGY_THRESH_2 {
-                        2
-                    } else if perturbation.total_energy > 1.0 {
-                        1
-                    } else {
-                        0
-                    }
-                }
-                Err(_) => score_to_person_count(smoothed_score, prev_count),
-            }
+            perturbation_occupancy(field, &frames)
+                .unwrap_or_else(|| score_to_person_count(smoothed_score, prev_count))
        }
        _ => score_to_person_count(smoothed_score, prev_count),
    }
 }

-/// Feed the latest frame to the FieldModel during calibration collection.
+/// Return true only when a restored low-authority bootstrap model has a full
+/// runtime window and that model resolves the window as empty. This helper is
+/// deliberately one-way: a bootstrap prior may suppress an obvious background
+/// false positive, but cannot authorize presence or vital signs.
+pub fn bootstrap_empty_prior_applies(
+    field: &FieldModel,
+    frame_history: &VecDeque<Vec<f64>>,
+    observed_at_us: u64,
+) -> bool {
+    bootstrap_background_match(field, frame_history, observed_at_us)
+        .is_some_and(|result| result.matches_empty)
+}
+
+/// Score the current canonical runtime window against the learned empty-room
+/// residual references. Legacy models without scalar references preserve the
+/// binary match but intentionally return no score.
+pub fn bootstrap_background_match(
+    field: &FieldModel,
+    frame_history: &VecDeque<Vec<f64>>,
+    observed_at_us: u64,
+) -> Option<BootstrapBackgroundMatch> {
+    if frame_history.len() < OCCUPANCY_WINDOW
+        || field.check_freshness(observed_at_us) != CalibrationStatus::Fresh
+    {
+        return None;
+    }
+    let frames: Vec<Vec<f64>> = frame_history
+        .iter()
+        .rev()
+        .take(OCCUPANCY_WINDOW)
+        .map(|frame| CALIB_NORMALIZER.resample_to_canonical(frame))
+        .collect();
+    let result = field.empty_room_match(&frames)?;
+    Some(BootstrapBackgroundMatch {
+        matches_empty: result.matches_empty,
+        score: result.score,
+        normalized_residual_z: result.normalized_residual_z,
+        maturity: result.maturity,
+        reliable: result.reliable,
+        residual_energy: result.residual_energy,
+        residual_energy_threshold: result.residual_energy_threshold,
+        window_size: result.window_size,
+        reference_window_count: result.reference_window_count,
+    })
+}
+
+/// Produce a fail-closed occupancy result from a fresh calibrated model.
+pub fn calibrated_occupancy(
+    field: &FieldModel,
+    frame_history: &VecDeque<Vec<f64>>,
+    observed_at_us: u64,
+) -> Option<CalibratedOccupancy> {
+    if field.check_freshness(observed_at_us) != CalibrationStatus::Fresh {
+        return None;
+    }
+    let frames: Vec<Vec<f64>> = frame_history
+        .iter()
+        .rev()
+        .take(OCCUPANCY_WINDOW)
+        .map(|frame| CALIB_NORMALIZER.resample_to_canonical(frame))
+        .collect();
+    if frames.is_empty() {
+        return None;
+    }
+    if let Ok(person_count) = field.estimate_occupancy(&frames) {
+        return Some(CalibratedOccupancy {
+            person_count: person_count.min(MAX_SINGLE_LINK_OCCUPANCY),
+            method: CalibratedOccupancyMethod::Eigenvalue,
+        });
+    }
+    Some(CalibratedOccupancy {
+        person_count: perturbation_occupancy(field, &frames)?,
+        method: CalibratedOccupancyMethod::PerturbationEnergy,
+    })
+}
+
+/// Feed the current CSI frame to the FieldModel during calibration.
 ///
 /// Acts while the model is `Uncalibrated` or `Collecting`. The first fed frame
 /// flips a freshly-started (`Uncalibrated`) model to `Collecting` inside
 /// `feed_calibration`; without accepting the `Uncalibrated` state here the two
 /// gates deadlock and the frame count never leaves 0 (calibration/start yields
-/// an `Uncalibrated` model that nothing would ever advance). Wraps the latest
-/// frame as a single-link observation (n_links=1) and feeds it.
-pub fn maybe_feed_calibration(field: &mut FieldModel, frame_history: &VecDeque<Vec<f64>>) {
+/// an `Uncalibrated` model that nothing would ever advance). Edge-vitals
+/// packets have no CSI observation and must never call this helper. The caller
+/// owns the stateful sequence admission gate. Wraps the current frame as a
+/// single-link observation (n_links=1) and feeds it.
+pub fn maybe_feed_calibration(field: &mut FieldModel, amplitudes: &[f64]) -> bool {
    if !matches!(
        field.status(),
        CalibrationStatus::Uncalibrated | CalibrationStatus::Collecting
-    ) {
-        return;
+    ) || amplitudes.is_empty()
+    {
+        return false;
    }
-    if let Some(latest) = frame_history.back() {
-        // Resample the raw amplitude vector onto the FieldModel's canonical
-        // 56-tone grid before feeding. Real HT40 nodes stream 128-wide frames;
-        // feeding those raw made every `feed_calibration` fail DimensionMismatch
-        // (swallowed at debug level), pinning frame_count at 0 even after the
-        // status-gate deadlock was fixed. Single-link observation: [1][56].
-        let canonical = CALIB_NORMALIZER.resample_to_canonical(latest);
-        let observations = vec![canonical];
-        if let Err(e) = field.feed_calibration(&observations) {
+    // Resample the raw amplitude vector onto the FieldModel's canonical
+    // 56-tone grid before feeding. Real HT40 nodes stream 128-wide frames;
+    // feeding those raw made every `feed_calibration` fail DimensionMismatch
+    // (swallowed at debug level), pinning frame_count at 0 even after the
+    // status-gate deadlock was fixed. Single-link observation: [1][56].
+    let canonical = CALIB_NORMALIZER.resample_to_canonical(amplitudes);
+    let observations = vec![canonical];
+    match field.feed_calibration(&observations) {
+        Ok(()) => true,
+        Err(e) => {
            tracing::debug!("FieldModel calibration feed: {e}");
+            false
        }
    }
 }

-/// Parse node positions from a semicolon-delimited string.
+/// One parsed `--node-positions` entry: a position, and the node id if the
+/// operator named one explicitly.
 ///
-/// Format: `"x,y,z;x,y,z;..."` where each coordinate is an `f32`.
-/// Malformed entries are skipped with a warning log.
-pub fn parse_node_positions(input: &str) -> Vec<[f32; 3]> {
+/// `node_id: None` means the entry was given positionally, so its identity is
+/// its index in the list — which is only the same thing as a node id on a fleet
+/// numbered `0, 1, 2, ...`.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub struct NodePositionEntry {
+    pub node_id: Option<u8>,
+    pub position: [f32; 3],
+}
+
+/// Parse node positions, keeping any explicit node ids.
+///
+/// Two accepted forms, per entry:
+///
+/// * `x,y,z`           — positional; identity is the entry's index.
+/// * `node_id:x,y,z`   — explicit; identity is the stated node id.
+///
+/// The explicit form exists because the positional one is only unambiguous on a
+/// fleet whose ids happen to run `0, 1, 2, ...`. Anywhere else, "the third
+/// entry" and "node 3" are different nodes, and nothing in the string says
+/// which was meant. Naming the id removes the guess rather than relocating it.
+///
+/// Both forms may be mixed; a positional entry keeps its index as its identity.
+/// Malformed entries are skipped with a warning rather than failing the boot —
+/// a typo in one triplet should not take the server down.
+pub fn parse_node_position_entries(input: &str) -> Vec<NodePositionEntry> {
    if input.is_empty() {
        return Vec::new();
    }
    input
        .split(';')
        .enumerate()
-        .filter_map(|(idx, triplet)| {
+        .filter_map(|(idx, raw)| {
+            let entry = raw.trim();
+            if entry.is_empty() {
+                return None;
+            }
+            // An explicit id is everything before the first ':'. Coordinates
+            // never contain one, so this cannot be ambiguous.
+            let (node_id, triplet) = match entry.split_once(':') {
+                Some((id_str, rest)) => match id_str.trim().parse::<u8>() {
+                    Ok(id) => (Some(id), rest),
+                    Err(_) => {
+                        tracing::warn!(
+                            "Skipping node position entry {idx}: '{entry}' has \
+                             an unparseable node id before ':' (expected \
+                             node_id:x,y,z with node_id in 0..=255)"
+                        );
+                        return None;
+                    }
+                },
+                None => (None, entry),
+            };
+
            let parts: Vec<&str> = triplet.split(',').collect();
            if parts.len() != 3 {
                tracing::warn!(
-                    "Skipping malformed node position entry {idx}: '{triplet}' (expected x,y,z)"
+                    "Skipping malformed node position entry {idx}: '{entry}' (expected x,y,z)"
                );
                return None;
            }
            match (
-                parts[0].parse::<f32>(),
-                parts[1].parse::<f32>(),
-                parts[2].parse::<f32>(),
+                parts[0].trim().parse::<f32>(),
+                parts[1].trim().parse::<f32>(),
+                parts[2].trim().parse::<f32>(),
            ) {
-                (Ok(x), Ok(y), Ok(z)) => Some([x, y, z]),
+                (Ok(x), Ok(y), Ok(z)) => Some(NodePositionEntry {
+                    node_id,
+                    position: [x, y, z],
+                }),
                _ => {
-                    tracing::warn!("Skipping unparseable node position entry {idx}: '{triplet}'");
+                    tracing::warn!("Skipping unparseable node position entry {idx}: '{entry}'");
                    None
                }
            }
@@ -170,10 +398,81 @@ pub fn parse_node_positions(input: &str) -> Vec<[f32; 3]> {
        .collect()
 }

+/// Build the `node_id -> position` map the live `NodeInfo` output reads.
+///
+/// Identity is the explicit `node_id:` prefix when the operator gave one, and
+/// the entry's list index otherwise.
+///
+/// This exists as a function rather than a loop inside `main()` because the bug
+/// it fixes was invisible without a test: positions were inserted keyed by list
+/// index and read back by `node_id`, so on any fleet not numbered `0, 1, 2, ...`
+/// every lookup missed and silently fell through to the hardcoded
+/// `[2.0, 0.0, 1.5]` that `--node-positions` exists to replace. The config
+/// looked applied and did nothing.
+pub fn node_positions_by_id(
+    entries: &[NodePositionEntry],
+) -> std::collections::HashMap<u8, [f32; 3]> {
+    let mut map = std::collections::HashMap::new();
+    for (idx, e) in entries.iter().enumerate() {
+        let key = e.node_id.unwrap_or(idx as u8);
+        if let Some(prev) = map.insert(key, e.position) {
+            tracing::warn!(
+                "node position for node {key} given twice ({prev:?} then {:?}); the later entry wins",
+                e.position
+            );
+        }
+    }
+    map
+}
+
+/// Parse node positions from a semicolon-delimited string.
+///
+/// Format: `"x,y,z;x,y,z;..."` where each coordinate is an `f32`.
+/// Malformed entries are skipped with a warning log.
+///
+/// Positional view, for consumers that index the list rather than key it by
+/// node id. Since issue #1866 the fusion path is no longer one of them: it
+/// takes the keyed map from [`node_positions_by_id`] via
+/// `MultistaticFuser::set_node_positions_by_id()`. Use
+/// [`parse_node_position_entries`] where identity is needed.
+pub fn parse_node_positions(input: &str) -> Vec<[f32; 3]> {
+    parse_node_position_entries(input)
+        .into_iter()
+        .map(|e| e.position)
+        .collect()
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;

+    fn fresh_test_model() -> FieldModel {
+        let config = FieldModelConfig {
+            n_links: 1,
+            n_subcarriers: 56,
+            n_modes: 3,
+            min_calibration_frames: 10,
+            min_calibration_duration_s: 0.0,
+            baseline_expiry_s: 2.0,
+        };
+        let mut field = FieldModel::new(config).expect("field model");
+        for index in 0..20 {
+            let phase = index as f64 * 0.1;
+            let observation = vec![(0..56)
+                .map(|subcarrier| {
+                    1.0 + subcarrier as f64 * 0.01 + (phase + subcarrier as f64 * 0.03).sin() * 0.01
+                })
+                .collect()];
+            field
+                .feed_calibration(&observation)
+                .expect("calibration frame");
+        }
+        field
+            .finalize_calibration(1_000_000, 7)
+            .expect("finalize calibration");
+        field
+    }
+
    #[test]
    fn test_parse_node_positions() {
        let positions = parse_node_positions("0,0,1.5;3,0,1.5;1.5,3,1.5");
@@ -203,6 +502,158 @@ mod tests {
        assert_eq!(positions[0], [3.0, 4.0, 5.0]);
    }

+    // ---- Explicit node ids (the indexing-mismatch regression) ----
+
+    /// Positional entries carry no id, so identity falls back to the index.
+    /// This is the ONLY case the old behaviour handled correctly.
+    #[test]
+    fn positional_entries_report_no_explicit_id() {
+        let e = parse_node_position_entries("0,0,1.5;3,0,1.5");
+        assert_eq!(e.len(), 2);
+        assert!(e.iter().all(|x| x.node_id.is_none()));
+        assert_eq!(e[1].position, [3.0, 0.0, 1.5]);
+    }
+
+    /// THE REGRESSION. Positions were stored keyed by list index and read back
+    /// by node_id, so on any fleet not numbered 0,1,2,... every lookup missed
+    /// and fell through to the hardcoded default the option exists to replace.
+    /// Non-sequential ids are what catch that.
+    #[test]
+    fn non_sequential_node_ids_are_preserved() {
+        let e = parse_node_position_entries("11:1,2,3;12:4,5,6;13:7,8,9");
+        assert_eq!(e.len(), 3);
+        assert_eq!(e[0].node_id, Some(11));
+        assert_eq!(e[1].node_id, Some(12));
+        assert_eq!(e[2].node_id, Some(13));
+        assert_eq!(e[2].position, [7.0, 8.0, 9.0]);
+        // The index would have been 0,1,2 -- none of which is a real node here.
+        for (idx, entry) in e.iter().enumerate() {
+            assert_ne!(
+                entry.node_id,
+                Some(idx as u8),
+                "index and node id must not be conflated"
+            );
+        }
+    }
+
+    /// Two-digit ids, and a fleet re-homed to even numbers -- the shape of the
+    /// planned hardware check (nine boards moved from 0..8 to 0,2,4,...,16).
+    #[test]
+    fn two_digit_and_even_numbered_fleet_ids_parse() {
+        let spec = "0:0,0,1;2:1,0,1;4:2,0,1;6:3,0,1;8:4,0,1;10:5,0,1;12:6,0,1;14:7,0,1;16:8,0,1";
+        let e = parse_node_position_entries(spec);
+        assert_eq!(e.len(), 9);
+        let ids: Vec<u8> = e.iter().map(|x| x.node_id.unwrap()).collect();
+        assert_eq!(ids, vec![0, 2, 4, 6, 8, 10, 12, 14, 16]);
+        assert_eq!(e[8].position, [8.0, 0.0, 1.0]);
+    }
+
+    /// A non-zero starting index must work too -- nothing requires node 0.
+    #[test]
+    fn a_fleet_that_does_not_start_at_zero_parses() {
+        let e = parse_node_position_entries("18:1.5,2.5,3.5");
+        assert_eq!(e[0].node_id, Some(18));
+        assert_eq!(e[0].position, [1.5, 2.5, 3.5]);
+    }
+
+    /// Mixing is allowed: an unkeyed entry keeps its index.
+    #[test]
+    fn mixed_forms_keep_their_own_identity() {
+        let e = parse_node_position_entries("0,0,1;7:9,9,9");
+        assert_eq!(e[0].node_id, None);
+        assert_eq!(e[1].node_id, Some(7));
+    }
+
+    /// An unparseable id is skipped rather than silently treated as positional
+    /// -- guessing there would reintroduce the exact ambiguity being removed.
+    #[test]
+    fn an_unparseable_node_id_is_skipped_not_guessed() {
+        let e = parse_node_position_entries("abc:1,2,3;5:4,5,6");
+        assert_eq!(e.len(), 1, "only the well-formed entry survives");
+        assert_eq!(e[0].node_id, Some(5));
+    }
+
+    /// 256 is out of range for a u8 node id and must not wrap to 0.
+    #[test]
+    fn an_out_of_range_node_id_is_rejected() {
+        let e = parse_node_position_entries("256:1,2,3");
+        assert!(e.is_empty(), "256 must not wrap to node 0");
+    }
+
+    /// The positional view stays byte-compatible for the fuser, which indexes
+    /// the list rather than keying it.
+    #[test]
+    fn the_positional_view_is_unchanged_by_explicit_ids() {
+        let p = parse_node_positions("11:1,2,3;12:4,5,6");
+        assert_eq!(p, vec![[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]);
+    }
+
+    /// The regression itself, at the map that `NodeInfo` actually reads.
+    ///
+    /// The old code inserted keyed by list index and looked up by node_id. On a
+    /// fleet numbered 11,12,13 that map holds keys 0,1,2, so every lookup misses
+    /// and the node reports the hardcoded default. This asserts both halves:
+    /// the real ids resolve, and the indices are absent.
+    #[test]
+    fn the_map_is_keyed_by_node_id_not_list_index() {
+        let entries = parse_node_position_entries("11:1,2,3;12:4,5,6;13:7,8,9");
+        let map = node_positions_by_id(&entries);
+
+        assert_eq!(map.get(&11), Some(&[1.0, 2.0, 3.0]));
+        assert_eq!(map.get(&12), Some(&[4.0, 5.0, 6.0]));
+        assert_eq!(map.get(&13), Some(&[7.0, 8.0, 9.0]));
+
+        for idx in 0u8..3 {
+            assert!(
+                map.get(&idx).is_none(),
+                "list index {idx} must not be a key -- that is the bug: a lookup                  by node_id would miss and fall back to the hardcoded position"
+            );
+        }
+    }
+
+    /// Positional input keeps the old behaviour exactly: index becomes the key.
+    /// This is what made the bug invisible on a 0..8 fleet.
+    #[test]
+    fn positional_input_still_keys_by_index() {
+        let entries = parse_node_position_entries("0,0,1;3,0,1;6,0,1");
+        let map = node_positions_by_id(&entries);
+        assert_eq!(map.get(&0), Some(&[0.0, 0.0, 1.0]));
+        assert_eq!(map.get(&1), Some(&[3.0, 0.0, 1.0]));
+        assert_eq!(map.get(&2), Some(&[6.0, 0.0, 1.0]));
+    }
+
+    /// A re-homed fleet at 0,2,4,...,16 resolves on the real ids, and the odd
+    /// indices in between are absent.
+    #[test]
+    fn an_even_numbered_fleet_resolves_by_id() {
+        let spec = "0:0,0,1;2:1,0,1;4:2,0,1;6:3,0,1;8:4,0,1;10:5,0,1;12:6,0,1;14:7,0,1;16:8,0,1";
+        let map = node_positions_by_id(&parse_node_position_entries(spec));
+        assert_eq!(map.len(), 9);
+        assert_eq!(map.get(&16), Some(&[8.0, 0.0, 1.0]));
+        assert_eq!(map.get(&10), Some(&[5.0, 0.0, 1.0]));
+        // Index-keyed would have produced 0..8; 1,3,5,7 are not real nodes here.
+        for odd in [1u8, 3, 5, 7] {
+            assert!(map.get(&odd).is_none(), "node {odd} does not exist in this fleet");
+        }
+    }
+
+    /// A duplicate id keeps the later entry rather than silently holding two.
+    #[test]
+    fn a_duplicate_node_id_keeps_the_later_entry() {
+        let map = node_positions_by_id(&parse_node_position_entries("4:1,1,1;4:2,2,2"));
+        assert_eq!(map.len(), 1);
+        assert_eq!(map.get(&4), Some(&[2.0, 2.0, 2.0]));
+    }
+
+    /// Whitespace around entries and coordinates is tolerated.
+    #[test]
+    fn whitespace_is_tolerated() {
+        let e = parse_node_position_entries(" 3 : 1.0 , 2.0 , 3.0 ");
+        assert_eq!(e.len(), 1);
+        assert_eq!(e[0].node_id, Some(3));
+        assert_eq!(e[0].position, [1.0, 2.0, 3.0]);
+    }
+
    /// Regression: a freshly-started (`Uncalibrated`) field model must begin
    /// collecting once frames arrive. Before the fix, `maybe_feed_calibration`
    /// only fed while already `Collecting`, but only `feed_calibration` sets
@@ -215,10 +666,7 @@ mod tests {

        // n_subcarriers defaults to 56; one single-link frame of that width.
        let frame = vec![0.5_f64; 56];
-        let mut history: VecDeque<Vec<f64>> = VecDeque::new();
-        history.push_back(frame);
-
-        maybe_feed_calibration(&mut field, &history);
+        assert!(maybe_feed_calibration(&mut field, &frame));

        assert_eq!(
            field.status(),
@@ -231,8 +679,8 @@ mod tests {
            "frame count must advance past 0"
        );

-        // Subsequent frames keep accumulating while Collecting.
-        maybe_feed_calibration(&mut field, &history);
+        // A subsequent unique frame keeps accumulating while Collecting.
+        assert!(maybe_feed_calibration(&mut field, &frame));
        assert_eq!(field.calibration_frame_count(), 2);
    }

@@ -247,10 +695,7 @@ mod tests {

        // 128-wide frame (HT40), NOT the model's 56 — would DimensionMismatch raw.
        let wide = vec![0.5_f64; 128];
-        let mut history: VecDeque<Vec<f64>> = VecDeque::new();
-        history.push_back(wide);
-
-        maybe_feed_calibration(&mut field, &history);
+        assert!(maybe_feed_calibration(&mut field, &wide));

        assert_eq!(
            field.status(),
@@ -263,4 +708,129 @@ mod tests {
            "wide frame must accumulate, not be silently dropped"
        );
    }
+
+    #[test]
+    fn maybe_feed_calibration_rejects_an_empty_observation() {
+        let mut field = FieldModel::new(single_link_config()).expect("field model");
+        assert!(!maybe_feed_calibration(&mut field, &[]));
+        assert_eq!(field.calibration_frame_count(), 0);
+    }
+
+    #[test]
+    fn calibrated_occupancy_is_fresh_model_only_and_normalizes_runtime_frames() {
+        let field = fresh_test_model();
+        let mut history = VecDeque::new();
+        history.push_back(vec![1.02; 128]);
+        let result =
+            calibrated_occupancy(&field, &history, 1_500_000).expect("fresh calibrated inference");
+        assert!(result.person_count <= MAX_SINGLE_LINK_OCCUPANCY);
+        assert!(matches!(
+            result.method,
+            CalibratedOccupancyMethod::Eigenvalue | CalibratedOccupancyMethod::PerturbationEnergy
+        ));
+        assert_eq!(calibrated_occupancy(&field, &history, 4_000_000), None);
+        assert_eq!(
+            calibrated_occupancy(&field, &VecDeque::new(), 1_500_000),
+            None
+        );
+    }
+
+    #[test]
+    fn occupancy_falls_back_after_expiry_without_a_process_restart() {
+        let field = fresh_test_model();
+        let baseline = field.modes().unwrap().baseline[0].clone();
+        let history: VecDeque<Vec<f64>> =
+            (0..OCCUPANCY_WINDOW).map(|_| baseline.clone()).collect();
+
+        assert_eq!(field.status(), CalibrationStatus::Fresh);
+        assert_eq!(
+            calibration_status_at(&field, 1_500_000),
+            CalibrationStatus::Fresh
+        );
+        assert_eq!(
+            occupancy_or_fallback(&field, &history, 1_500_000, 0.0, 0),
+            0,
+            "a fresh field model may resolve its learned baseline as empty"
+        );
+
+        assert_eq!(field.status(), CalibrationStatus::Fresh);
+        assert_eq!(
+            calibration_status_at(&field, 4_000_000),
+            CalibrationStatus::Expired
+        );
+        assert_eq!(
+            occupancy_or_fallback(&field, &history, 4_000_000, 0.0, 0),
+            score_to_person_count(0.0, 0),
+            "an expired in-memory model must fall back even though its cached status is fresh"
+        );
+    }
+
+    #[test]
+    fn adaptive_runtime_reference_scores_empty_and_rejects_shift() {
+        let config = FieldModelConfig {
+            n_links: 1,
+            n_subcarriers: 56,
+            n_modes: 3,
+            min_calibration_frames: 500,
+            min_calibration_duration_s: 0.0,
+            baseline_expiry_s: 60.0,
+        };
+        let mut field = FieldModel::new(config).expect("field model");
+        let mut calibration = Vec::new();
+        for frame_index in 0..600 {
+            let time = frame_index as f64 * 0.071;
+            let frame: Vec<f64> = (0..56)
+                .map(|subcarrier| {
+                    let carrier = subcarrier as f64;
+                    18.0 + carrier * 0.08
+                        + (time + carrier * 0.13).sin() * 0.8
+                        + (time * 0.37 + carrier * 0.031).cos() * 0.35
+                })
+                .collect();
+            field.feed_calibration(&[frame.clone()]).unwrap();
+            calibration.push(frame);
+        }
+        field.finalize_calibration(1_000_000, 7).unwrap();
+
+        let empty_history: VecDeque<Vec<f64>> = calibration[550..600].iter().cloned().collect();
+        let empty = calibrated_occupancy(&field, &empty_history, 1_500_000)
+            .expect("adaptive perturbation result");
+        assert_eq!(empty.person_count, 0);
+        let background = bootstrap_background_match(&field, &empty_history, 1_500_000)
+            .expect("background score");
+        assert!(background.matches_empty);
+        assert_eq!(background.reference_window_count, 12);
+        assert!(background.reliable);
+        assert_eq!(background.maturity, 0.6);
+        assert!(background.normalized_residual_z.is_some());
+        assert!(background
+            .score
+            .is_some_and(|score| (0.5..=1.0).contains(&score)));
+
+        let short_history: VecDeque<Vec<f64>> = empty_history.iter().take(49).cloned().collect();
+        assert!(!bootstrap_empty_prior_applies(
+            &field,
+            &short_history,
+            1_500_000
+        ));
+
+        let shifted_history: VecDeque<Vec<f64>> = calibration[550..600]
+            .iter()
+            .map(|frame| {
+                frame
+                    .iter()
+                    .enumerate()
+                    .map(|(index, value)| value + if index % 3 == 0 { 8.0 } else { -5.0 })
+                    .collect()
+            })
+            .collect();
+        let shifted = calibrated_occupancy(&field, &shifted_history, 1_500_000)
+            .expect("shifted perturbation result");
+        assert!(shifted.person_count >= 1);
+        let shifted_background = bootstrap_background_match(&field, &shifted_history, 1_500_000)
+            .expect("shifted background result");
+        assert!(!shifted_background.matches_empty);
+        assert_eq!(shifted_background.score, Some(0.0));
+        assert!(shifted_background.reliable);
+    }
 }
@@ -9,6 +9,7 @@
 //! - Real-time CSI introspection / low-latency tap (`introspection`, ADR-099)

 pub mod bearer_auth;
+pub mod bootstrap_baseline;
 pub mod browser_session;
 pub mod ws_ticket;
 pub mod cli;
@@ -544,6 +544,10 @@ impl AppStateInner {
    pub fn person_count(&self) -> usize {
        use crate::csi::score_to_person_count;
        use crate::field_bridge;
+        let observed_at_us = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .map(|duration| duration.as_micros() as u64)
+            .unwrap_or(0);
        match self.field_model.as_ref() {
            Some(fm) => {
                let history = if !self.frame_history.is_empty() {
@@ -559,6 +563,7 @@ impl AppStateInner {
                field_bridge::occupancy_or_fallback(
                    fm,
                    history,
+                    observed_at_us,
                    self.smoothed_person_score,
                    self.prev_person_count,
                )
@@ -346,6 +346,31 @@ impl VitalSignDetector {
        self.frame_count = 0;
    }

+    /// Retune temporal windows to an observed CSI clock. Changes smaller than
+    /// twenty percent are ignored to avoid resetting on normal arrival jitter.
+    /// A real clock change clears history so samples from different clocks are
+    /// never interpreted in one spectrum.
+    pub fn reconfigure_sample_rate(&mut self, sample_rate: f64) -> bool {
+        if !sample_rate.is_finite() || sample_rate < 1.0 {
+            return false;
+        }
+        let relative_change = (sample_rate - self.sample_rate).abs() / self.sample_rate.max(1.0);
+        if relative_change < 0.20 {
+            return false;
+        }
+        self.sample_rate = sample_rate;
+        self.breathing_capacity =
+            ((sample_rate * self.breathing_window_secs) as usize).max(1);
+        self.heartbeat_capacity =
+            ((sample_rate * self.heartbeat_window_secs) as usize).max(1);
+        self.reset();
+        true
+    }
+
+    pub fn sample_rate_hz(&self) -> f64 {
+        self.sample_rate
+    }
+
    /// Current buffer fill levels for diagnostics.
    /// Returns (breathing_len, breathing_capacity, heartbeat_len, heartbeat_capacity).
    pub fn buffer_status(&self) -> (usize, usize, usize, usize) {
@@ -842,6 +867,22 @@ mod tests {
        assert_eq!(hb_len, 0);
    }

+    #[test]
+    fn measured_clock_reconfigures_once_and_clears_mixed_history() {
+        let mut detector = VitalSignDetector::new(20.0);
+        let amp = vec![10.0; 56];
+        let phase = vec![0.0; 56];
+        for _ in 0..32 {
+            detector.process_frame(&amp, &phase);
+        }
+        assert!(!detector.reconfigure_sample_rate(18.0));
+        assert_eq!(detector.sample_rate_hz(), 20.0);
+        assert!(detector.reconfigure_sample_rate(12.0));
+        assert_eq!(detector.sample_rate_hz(), 12.0);
+        assert_eq!(detector.buffer_status(), (0, 360, 0, 180));
+        assert!(!detector.reconfigure_sample_rate(f64::NAN));
+    }
+
    #[test]
    fn test_vital_signs_default() {
        let vs = VitalSigns::default();
@@ -26,6 +26,7 @@
 //! - `ruvector-attn-mincut` for cross-node spectrogram attention gating
 //! - `ruvector-mincut` for person separation (DynamicMinCut)

+use std::collections::HashMap;
 use std::sync::Arc;

 use super::cir::{CirConfig, CirEstimator};
@@ -72,6 +73,12 @@ pub struct FusedSensingFrame {
    /// Per-node multi-band frames (preserved for geometry computations).
    pub node_frames: Vec<MultiBandCsiFrame>,
    /// Node positions (x, y, z) in meters from deployment configuration.
+    ///
+    /// Parallel to [`Self::node_frames`]: `node_positions[i]` is the
+    /// configured position of `node_frames[i].node_id`, resolved by identity
+    /// rather than by rank (issue #1866). Nodes with no configured position
+    /// appear at the origin. Index `i` is this cycle's cohort rank, which is
+    /// *not* a node id -- read the id from the paired frame.
    pub node_positions: Vec<[f32; 3]>,
    /// Number of active nodes contributing to this frame.
    pub active_nodes: usize,
@@ -147,8 +154,31 @@ impl MultistaticConfig {
    /// maximum legitimate spread between two paired node frames in one cycle is
    /// the full cycle length `tdm_total_slots × slot_duration_us` (last slot vs
    /// first slot). The hard guard is set to that cycle length plus 20% jitter
-    /// headroom; the soft guard to ~⅓ of the cycle (a normal adjacent-slot pair
-    /// fuses cleanly, a near-full-cycle spread is flagged as loose alignment).
+    /// headroom; the soft guard to the schedule's own **designed stagger**,
+    /// `(slots - 1) * slot_duration_us` -- the offset between the first and
+    /// last slot, which is the spread a complete, healthy cycle produces.
+    ///
+    /// # Why not ~⅓ of the cycle (issue #1866-adjacent, corrected)
+    ///
+    /// The soft guard was previously `cycle_us / 3`, reasoned about as "a
+    /// normal *adjacent-slot pair* fuses cleanly". But [`MultistaticFuser::fuse_scored`]
+    /// does not compare pairs -- it takes `max(timestamp) - min(timestamp)`
+    /// across the **whole cohort**, so the spread it measures is always the
+    /// first-to-last offset, `(N - 1)` slots. Against that:
+    ///
+    /// ```text
+    ///   soft < stagger  <=>  N*slot/3 < (N-1)*slot  <=>  N/3 < N-1  <=>  N > 1.5
+    /// ```
+    ///
+    /// which holds for **every** `N >= 2`. The derived soft guard therefore
+    /// never covered the schedule it was derived from, at any fleet size, and
+    /// a complete healthy cycle was flagged as a `TimestampMismatch`
+    /// contradiction -- which demotes the BFLD privacy class one step
+    /// (ADR-137 §2.7 -> ADR-141) and suppresses raw outputs.
+    ///
+    /// The doc example below is the smallest demonstration: the 2-slot mesh
+    /// whose **measured** 18,194 µs spread motivated this constructor derived
+    /// a 12,000 µs soft guard, below its own motivating measurement.
    ///
    /// `tdm_total_slots` is clamped to ≥ 1. All other fields take their
    /// [`Default`] values.
@@ -160,6 +190,9 @@ impl MultistaticConfig {
    /// // reported 18,194 µs 2-slot spread.
    /// let cfg = MultistaticConfig::for_tdm_schedule(2, 18_000);
    /// assert!(cfg.guard_interval_us >= 18_194);
+    /// // ...and the soft guard now covers it too, so a healthy 2-slot cycle
+    /// // is not recorded as a contradiction.
+    /// assert!(cfg.soft_guard_us >= 18_194);
    /// ```
    #[must_use]
    pub fn for_tdm_schedule(tdm_total_slots: usize, slot_duration_us: u64) -> Self {
@@ -167,8 +200,19 @@ impl MultistaticConfig {
        let cycle_us = slots.saturating_mul(slot_duration_us);
        // +20% jitter headroom on the full cycle.
        let guard_interval_us = cycle_us.saturating_add(cycle_us / 5).max(1);
-        // Soft band at ~⅓ cycle, kept strictly below the hard guard.
-        let soft_guard_us = (cycle_us / 3).clamp(1, guard_interval_us.saturating_sub(1).max(1));
+        // Soft band at the schedule's designed stagger -- the first-to-last
+        // slot offset, which is what `fuse_scored` measures. Kept strictly
+        // below the hard guard, so a spread approaching a full cycle is still
+        // flagged as loose alignment.
+        // +20% jitter headroom, mirroring the hard guard: the *nominal* stagger
+        // is not the observed one. The 2-slot mesh below has an 18,000 us
+        // nominal stagger but a MEASURED 18,194 us spread, so a bare nominal
+        // bound would flag the very cycle it was derived from. Always strictly
+        // below the hard guard, since stagger < cycle.
+        let stagger_us = slots.saturating_sub(1).saturating_mul(slot_duration_us);
+        let soft_guard_us = stagger_us
+            .saturating_add(stagger_us / 5)
+            .clamp(1, guard_interval_us.saturating_sub(1).max(1));
        Self {
            guard_interval_us,
            soft_guard_us,
@@ -191,8 +235,16 @@ impl MultistaticConfig {
 /// disable the CIR path and keep the legacy frequency-domain coherence only.
 pub struct MultistaticFuser {
    config: MultistaticConfig,
-    /// Node positions in 3D space (meters).
-    node_positions: Vec<[f32; 3]>,
+    /// Node positions in 3D space (meters), keyed by node id (issue #1866).
+    ///
+    /// Keyed rather than indexed because [`Self::fuse`] receives a *compacted*
+    /// cohort: `node_frames_from_states_with_guard` drops nodes that are
+    /// stale, outside the timestamp guard, or have an empty history. Under
+    /// index addressing, one node missing a single beacon re-ranks every node
+    /// above it and silently hands them their neighbour's coordinates. Each
+    /// `MultiBandCsiFrame` carries its own `node_id`, so the position is
+    /// looked up by identity and a silent node moves nobody.
+    node_positions: HashMap<u8, [f32; 3]>,
    /// Optional shared CIR estimator (ADR-134).  `None` = legacy path only.
    cir_estimator: Option<Arc<CirEstimator>>,
 }
@@ -212,7 +264,7 @@ impl MultistaticFuser {
    pub fn new() -> Self {
        Self {
            config: MultistaticConfig::default(),
-            node_positions: Vec::new(),
+            node_positions: HashMap::new(),
            cir_estimator: None,
        }
    }
@@ -221,7 +273,7 @@ impl MultistaticFuser {
    pub fn with_config(config: MultistaticConfig) -> Self {
        Self {
            config,
-            node_positions: Vec::new(),
+            node_positions: HashMap::new(),
            cir_estimator: None,
        }
    }
@@ -267,11 +319,39 @@ impl MultistaticFuser {
        fuser
    }

-    /// Set node positions for geometric diversity computations.
+    /// Set node positions from a positional list (**legacy compatibility**).
+    ///
+    /// The list index is the node id: `positions[i]` belongs to node `i`. That
+    /// is the meaning the `--node-positions` flag has always carried when its
+    /// entries are given without an explicit `node_id:` prefix, so this
+    /// mapping is preserved exactly rather than reinterpreted.
+    ///
+    /// This is only correct for fleets whose ids are dense and zero-based.
+    /// Prefer [`Self::set_node_positions_by_id`], which states the identity
+    /// instead of inferring it from ordering.
    pub fn set_node_positions(&mut self, positions: Vec<[f32; 3]>) {
+        self.node_positions = positions
+            .into_iter()
+            .enumerate()
+            .map(|(idx, position)| (idx as u8, position))
+            .collect();
+    }
+
+    /// Set node positions keyed by node id (issue #1866, preferred).
+    ///
+    /// The position travels with the board it was measured for. Nodes absent
+    /// from the map fuse at the origin, exactly as before; nodes present but
+    /// not currently reporting cost nothing and shift nobody.
+    pub fn set_node_positions_by_id(&mut self, positions: HashMap<u8, [f32; 3]>) {
        self.node_positions = positions;
    }

+    /// Return the configured position for one node id, if any.
+    #[must_use]
+    pub fn node_position(&self, node_id: u8) -> Option<[f32; 3]> {
+        self.node_positions.get(&node_id).copied()
+    }
+
    /// Return the configured hard timestamp guard in microseconds.
    ///
    /// Callers that assemble frames before invoking [`Self::fuse`] use this
@@ -283,11 +363,33 @@ impl MultistaticFuser {
        self.config.guard_interval_us
    }

-    /// Return the current node positions.
-    pub fn node_positions(&self) -> &[[f32; 3]] {
+    /// Return the configured positions keyed by node id.
+    #[must_use]
+    pub fn node_positions_by_id(&self) -> &HashMap<u8, [f32; 3]> {
        &self.node_positions
    }

+    /// Return the configured positions as a dense id-indexed list
+    /// (**legacy compatibility view**).
+    ///
+    /// Entry `i` is node `i`'s position, with unconfigured ids filled at the
+    /// origin, so this round-trips [`Self::set_node_positions`]. It is a view
+    /// of the configuration, *not* of what any particular fusion cycle used —
+    /// a cycle's cohort is compacted, so read
+    /// [`FusedSensingFrame::node_positions`] alongside its `node_frames` for
+    /// that.
+    #[must_use]
+    pub fn node_positions(&self) -> Vec<[f32; 3]> {
+        let Some(&max_id) = self.node_positions.keys().max() else {
+            return Vec::new();
+        };
+        let mut dense = vec![[0.0_f32, 0.0, 0.0]; max_id as usize + 1];
+        for (&id, &position) in &self.node_positions {
+            dense[id as usize] = position;
+        }
+        dense
+    }
+
    /// Fuse multiple node frames into a single `FusedSensingFrame`.
    ///
    /// When only one node is provided, falls back to single-node mode
@@ -360,11 +462,19 @@ impl MultistaticFuser {
        timestamps.sort_unstable();
        let timestamp_us = timestamps[timestamps.len() / 2];

-        // Build node positions list, filling with origin for unknown nodes
-        let positions: Vec<[f32; 3]> = (0..n_nodes)
-            .map(|i| {
+        // Build the node positions list parallel to `node_frames`, resolving
+        // each entry by that frame's own `node_id` (issue #1866). The cohort
+        // handed to `fuse` is compacted -- stale, out-of-guard and
+        // empty-history nodes are already dropped -- so a frame's index here
+        // is its rank in this cycle, not its identity. Addressing by index
+        // gave every node above a gap its neighbour's coordinates whenever a
+        // single board missed a beacon. Unknown nodes still fill with the
+        // origin, unchanged.
+        let positions: Vec<[f32; 3]> = node_frames
+            .iter()
+            .map(|frame| {
                self.node_positions
-                    .get(i)
+                    .get(&frame.node_id)
                    .copied()
                    .unwrap_or([0.0, 0.0, 0.0])
            })
@@ -964,6 +1074,88 @@ mod tests {
        );
    }

+    /// A derived schedule must not flag its own healthy cycle as a
+    /// contradiction.
+    ///
+    /// Every prior assertion on `soft_guard_us` was *relational* (`soft <
+    /// hard`, `soft >= 1`), so nothing compared it to a spread the schedule
+    /// actually produces -- which is how `cycle/3` survived. `fuse_scored`
+    /// measures `max - min` across the whole cohort, i.e. `(N-1)` slots, so
+    /// that is what the soft guard must cover.
+    #[test]
+    fn derived_soft_guard_covers_the_schedules_own_stagger() {
+        for (slots, slot_us) in [
+            (2_usize, 18_000_u64), // the measured 2-slot mesh in the doc comment
+            (3, 27_000),
+            (4, 12_500),
+            (6, 33_300),
+            (9, 35_600),  // a nine-node fleet: 320 ms cycle
+            (13, 36_900),
+        ] {
+            let cfg = MultistaticConfig::for_tdm_schedule(slots, slot_us);
+            // Nominal stagger; the derived guard must clear it WITH headroom,
+            // because observed spread runs slightly over nominal (18,194 us
+            // measured against an 18,000 us nominal 2-slot stagger).
+            let stagger = (slots as u64 - 1) * slot_us;
+            assert!(
+                cfg.soft_guard_us > stagger,
+                "{slots} slots x {slot_us} us: soft guard {} us is below the \
+                 schedule's own first-to-last stagger of {stagger} us, so a \
+                 complete healthy cycle is recorded as a TimestampMismatch \
+                 contradiction and demotes the privacy class",
+                cfg.soft_guard_us
+            );
+            assert!(cfg.soft_guard_us < cfg.guard_interval_us);
+        }
+    }
+
+    /// End to end: a full, evenly-staggered cycle fuses with **no**
+    /// contradiction under its own derived config.
+    #[test]
+    fn a_healthy_full_cycle_raises_no_contradiction() {
+        let (slots, slot_us) = (9_usize, 35_600_u64);
+        let cfg = MultistaticConfig::for_tdm_schedule(slots, slot_us);
+        let fuser = MultistaticFuser::with_config(cfg.clone());
+
+        // Node k transmits in slot k -- the schedule working exactly as designed.
+        let frames: Vec<MultiBandCsiFrame> = (0..slots)
+            .map(|k| make_node_frame(k as u8, k as u64 * slot_us, 56, 1.0))
+            .collect();
+
+        let (_fused, score) = fuser
+            .fuse_scored(&frames, 0.0)
+            .expect("a healthy cycle must fuse");
+        assert!(
+            !score.forces_privacy_demotion(),
+            "a complete {slots}-slot cycle is the schedule operating normally \
+             and must not demote the privacy class; flags: {:?}",
+            score.contradiction_flags
+        );
+    }
+
+    /// The soft band still does its job: a spread beyond the designed stagger
+    /// is flagged.
+    #[test]
+    fn a_spread_beyond_the_stagger_is_still_flagged() {
+        let (slots, slot_us) = (9_usize, 35_600_u64);
+        let cfg = MultistaticConfig::for_tdm_schedule(slots, slot_us);
+        let hard = cfg.guard_interval_us;
+        let fuser = MultistaticFuser::with_config(cfg);
+
+        // Two nodes, inside the hard guard but wider than a full cycle's stagger.
+        let late = hard - 1;
+        let frames = vec![
+            make_node_frame(0, 0, 56, 1.0),
+            make_node_frame(1, late, 56, 1.0),
+        ];
+        let (_f, score) = fuser.fuse_scored(&frames, 0.0).expect("within hard guard");
+        assert!(
+            score.forces_privacy_demotion(),
+            "a spread of {late} us exceeds the designed stagger and must still \
+             be recorded as loose alignment"
+        );
+    }
+
    /// The derived soft guard stays strictly below the hard guard, and a
    /// degenerate (0-slot) schedule clamps to a usable config.
    #[test]
@@ -1064,10 +1256,16 @@ mod tests {

    #[test]
    fn node_positions_set_and_retrieved() {
+        // The legacy positional setter round-trips through the dense view:
+        // index i in, index i out. Storage is keyed by id underneath, and
+        // `positions[i]` is documented to mean node i.
        let mut fuser = MultistaticFuser::new();
        let positions = vec![[0.0, 0.0, 1.0], [3.0, 0.0, 1.0]];
        fuser.set_node_positions(positions.clone());
-        assert_eq!(fuser.node_positions(), &positions[..]);
+        assert_eq!(fuser.node_positions(), positions);
+        assert_eq!(fuser.node_position(0), Some([0.0, 0.0, 1.0]));
+        assert_eq!(fuser.node_position(1), Some([3.0, 0.0, 1.0]));
+        assert_eq!(fuser.node_position(2), None);
    }

    #[test]
@@ -1083,6 +1281,255 @@ mod tests {
        assert_eq!(fused.node_positions[1], [0.0, 0.0, 0.0]); // default
    }

+    /// Issue #1866 regressions: a fused contribution must carry the position
+    /// belonging to its own `node_id`, under every cohort shape the live
+    /// collector can produce.
+    ///
+    /// `node_frames_from_states_with_guard` hands `fuse` a *compacted* cohort:
+    /// stale, out-of-guard and empty-history nodes are already gone. So the
+    /// index of a frame is its rank in this cycle, never its identity.
+    mod positions_keyed_by_node_id {
+        use super::*;
+
+        fn positions(entries: &[(u8, [f32; 3])]) -> HashMap<u8, [f32; 3]> {
+            entries.iter().copied().collect()
+        }
+
+        /// Every frame gets its own node's coordinates -- the core assertion.
+        fn assert_each_frame_owns_its_position(
+            configured: &HashMap<u8, [f32; 3]>,
+            frames: &[MultiBandCsiFrame],
+        ) {
+            let mut fuser = MultistaticFuser::new();
+            fuser.set_node_positions_by_id(configured.clone());
+            let fused = fuser.fuse(frames).expect("cohort fuses");
+
+            assert_eq!(fused.node_positions.len(), frames.len());
+            for (frame, position) in fused.node_frames.iter().zip(&fused.node_positions) {
+                let expected = configured
+                    .get(&frame.node_id)
+                    .copied()
+                    .unwrap_or([0.0, 0.0, 0.0]);
+                assert_eq!(
+                    *position, expected,
+                    "node {} received the wrong position",
+                    frame.node_id
+                );
+            }
+        }
+
+        #[test]
+        fn sparse_node_ids_each_get_their_own_position() {
+            // Real fleets use logical ids like 11/12/13. Under rank addressing
+            // every one of these resolved to the origin, because indices
+            // 11..13 do not exist in a 3-entry list.
+            let configured = positions(&[
+                (11, [1.0, 0.0, 0.0]),
+                (12, [2.0, 0.0, 0.0]),
+                (13, [3.0, 0.0, 0.0]),
+            ]);
+            let frames = vec![
+                make_node_frame(11, 100, 56, 1.0),
+                make_node_frame(12, 101, 56, 1.0),
+                make_node_frame(13, 102, 56, 1.0),
+            ];
+            assert_each_frame_owns_its_position(&configured, &frames);
+
+            let mut fuser = MultistaticFuser::new();
+            fuser.set_node_positions_by_id(configured);
+            let fused = fuser.fuse(&frames).expect("cohort fuses");
+            assert_eq!(fused.node_positions[0], [1.0, 0.0, 0.0]);
+            assert_eq!(fused.node_positions[2], [3.0, 0.0, 0.0]);
+        }
+
+        #[test]
+        fn configuration_order_does_not_change_any_position() {
+            // Insertion order must be irrelevant: identity is stated, not
+            // inferred from where an entry sat in the config.
+            let ascending = positions(&[
+                (3, [3.0, 0.0, 0.0]),
+                (5, [5.0, 0.0, 0.0]),
+                (9, [9.0, 0.0, 0.0]),
+            ]);
+            let reversed = positions(&[
+                (9, [9.0, 0.0, 0.0]),
+                (5, [5.0, 0.0, 0.0]),
+                (3, [3.0, 0.0, 0.0]),
+            ]);
+            let frames = vec![
+                make_node_frame(3, 100, 56, 1.0),
+                make_node_frame(5, 101, 56, 1.0),
+                make_node_frame(9, 102, 56, 1.0),
+            ];
+
+            let mut a = MultistaticFuser::new();
+            a.set_node_positions_by_id(ascending);
+            let mut b = MultistaticFuser::new();
+            b.set_node_positions_by_id(reversed);
+
+            assert_eq!(
+                a.fuse(&frames).expect("fuses").node_positions,
+                b.fuse(&frames).expect("fuses").node_positions
+            );
+        }
+
+        #[test]
+        fn a_silent_node_does_not_move_the_others() {
+            // THE REASON THIS IS KEYED BY ID. Node 2 misses a beacon and drops
+            // out of the cohort. Under rank addressing nodes 3 and 4 shift
+            // down one slot and silently inherit their neighbours' coordinates
+            // -- a fleet-wide position error caused by one dropped frame.
+            let configured = positions(&[
+                (1, [1.0, 0.0, 0.0]),
+                (2, [2.0, 0.0, 0.0]),
+                (3, [3.0, 0.0, 0.0]),
+                (4, [4.0, 0.0, 0.0]),
+            ]);
+            let mut fuser = MultistaticFuser::new();
+            fuser.set_node_positions_by_id(configured);
+
+            let all_up = vec![
+                make_node_frame(1, 100, 56, 1.0),
+                make_node_frame(2, 101, 56, 1.0),
+                make_node_frame(3, 102, 56, 1.0),
+                make_node_frame(4, 103, 56, 1.0),
+            ];
+            let two_silent = vec![
+                make_node_frame(1, 100, 56, 1.0),
+                make_node_frame(3, 102, 56, 1.0),
+                make_node_frame(4, 103, 56, 1.0),
+            ];
+
+            let full = fuser.fuse(&all_up).expect("fuses");
+            let gapped = fuser.fuse(&two_silent).expect("fuses");
+
+            assert_eq!(full.node_positions[2], [3.0, 0.0, 0.0]);
+            assert_eq!(full.node_positions[3], [4.0, 0.0, 0.0]);
+            // Node 3 is now at rank 1 and node 4 at rank 2, but both keep
+            // their own coordinates.
+            assert_eq!(gapped.node_positions[0], [1.0, 0.0, 0.0]);
+            assert_eq!(gapped.node_positions[1], [3.0, 0.0, 0.0]);
+            assert_eq!(gapped.node_positions[2], [4.0, 0.0, 0.0]);
+        }
+
+        #[test]
+        fn join_and_leave_churn_never_reassigns_a_position() {
+            // Walk a node through leave -> rejoin. Every cohort shape must
+            // still pair each frame with its own configured coordinates.
+            let configured = positions(&[
+                (0, [0.0, 0.0, 0.0]),
+                (1, [1.0, 1.0, 0.0]),
+                (2, [2.0, 2.0, 0.0]),
+                (3, [3.0, 3.0, 0.0]),
+            ]);
+            let cohorts: Vec<Vec<u8>> = vec![
+                vec![0, 1, 2, 3],
+                vec![0, 2, 3],
+                vec![2, 3],
+                vec![3],
+                vec![0, 3],
+                vec![0, 1, 2, 3],
+            ];
+            for ids in cohorts {
+                let frames: Vec<MultiBandCsiFrame> = ids
+                    .iter()
+                    .enumerate()
+                    .map(|(i, &id)| make_node_frame(id, 100 + i as u64, 56, 1.0))
+                    .collect();
+                assert_each_frame_owns_its_position(&configured, &frames);
+            }
+        }
+
+        #[test]
+        fn a_configured_but_absent_node_costs_nothing() {
+            // Positions may be configured for boards that are not reporting.
+            // They must not consume a slot or displace anyone.
+            let configured = positions(&[
+                (1, [1.0, 0.0, 0.0]),
+                (7, [7.0, 0.0, 0.0]),
+                (42, [42.0, 0.0, 0.0]),
+            ]);
+            let frames = vec![
+                make_node_frame(1, 100, 56, 1.0),
+                make_node_frame(7, 101, 56, 1.0),
+            ];
+            assert_each_frame_owns_its_position(&configured, &frames);
+
+            let mut fuser = MultistaticFuser::new();
+            fuser.set_node_positions_by_id(configured);
+            let fused = fuser.fuse(&frames).expect("fuses");
+            assert_eq!(fused.node_positions.len(), 2);
+            assert_eq!(fused.node_positions[1], [7.0, 0.0, 0.0]);
+        }
+
+        #[test]
+        fn an_unconfigured_node_still_fuses_at_the_origin() {
+            // Unchanged fallback behaviour: unknown node -> origin, and it
+            // must not disturb its neighbours.
+            let configured = positions(&[(1, [1.0, 0.0, 0.0]), (3, [3.0, 0.0, 0.0])]);
+            let frames = vec![
+                make_node_frame(1, 100, 56, 1.0),
+                make_node_frame(2, 101, 56, 1.0),
+                make_node_frame(3, 102, 56, 1.0),
+            ];
+            assert_each_frame_owns_its_position(&configured, &frames);
+
+            let mut fuser = MultistaticFuser::new();
+            fuser.set_node_positions_by_id(configured);
+            let fused = fuser.fuse(&frames).expect("fuses");
+            assert_eq!(fused.node_positions[1], [0.0, 0.0, 0.0]);
+        }
+
+        #[test]
+        fn legacy_positional_setter_still_means_index_equals_node_id() {
+            // Documented compatibility mapping: a bare `--node-positions`
+            // list has always meant "the i-th entry is node i". Preserved
+            // exactly, so existing deployments are unaffected.
+            let mut fuser = MultistaticFuser::new();
+            fuser.set_node_positions(vec![
+                [0.0, 0.0, 0.0],
+                [1.0, 0.0, 0.0],
+                [2.0, 0.0, 0.0],
+            ]);
+            let frames = vec![
+                make_node_frame(0, 100, 56, 1.0),
+                make_node_frame(1, 101, 56, 1.0),
+                make_node_frame(2, 102, 56, 1.0),
+            ];
+            let fused = fuser.fuse(&frames).expect("fuses");
+            assert_eq!(fused.node_positions[0], [0.0, 0.0, 0.0]);
+            assert_eq!(fused.node_positions[1], [1.0, 0.0, 0.0]);
+            assert_eq!(fused.node_positions[2], [2.0, 0.0, 0.0]);
+        }
+
+        /// Negative control: the pre-fix behaviour must actually be gone.
+        ///
+        /// Rank addressing over a dense zero-based config only diverges from
+        /// id addressing once the cohort is compacted, so this pins the exact
+        /// case the old code got wrong rather than restating the happy path.
+        #[test]
+        fn rank_addressing_would_have_failed_this_cohort() {
+            let configured = positions(&[
+                (0, [0.0, 0.0, 0.0]),
+                (1, [1.0, 0.0, 0.0]),
+                (2, [2.0, 0.0, 0.0]),
+            ]);
+            let mut fuser = MultistaticFuser::new();
+            fuser.set_node_positions_by_id(configured);
+
+            // Node 1 is silent; the cohort is [0, 2].
+            let frames = vec![
+                make_node_frame(0, 100, 56, 1.0),
+                make_node_frame(2, 101, 56, 1.0),
+            ];
+            let fused = fuser.fuse(&frames).expect("fuses");
+
+            // Rank addressing would hand rank 1 the position of node 1.
+            assert_ne!(fused.node_positions[1], [1.0, 0.0, 0.0]);
+            assert_eq!(fused.node_positions[1], [2.0, 0.0, 0.0]);
+        }
+    }
+
    #[test]
    fn geometric_diversity_single_node() {
        assert_eq!(geometric_diversity(&[[0.0, 0.0, 0.0]]), 0.0);
@@ -1 +1 @@
 .8.8
 .8.10