rea/experiments/analysis-efficiency/immutable-parent-validation.md

Reuse successful immutable parent validation

This checkpoint addresses #4810, found in the retained-investigation audit. It changes the shared format interpretation used by CLI and MCP views, without adding a tool, changing a schema, projecting away source data, or rerunning a provider.

Question, ownership, and compatibility

The task resumes authentic exported Evidence in a fresh production stdio MCP server, discovers its source, reads six original symbols or modules in two-row pages, repeats that traversal three times, queries the existing Unknowns, exports the investigation, and closes it. The original source records must survive the round trip exactly. Pages are compared to original retained facts. Independent ELF byte decoding checks symbol names, values, and ranges; JavaScript checks cover selected source bytes and the lib/cli.js module identity, rather than independently proving the entire application inventory or semantic graph.

Imported JavaScript graphs lack the producer's realm-private format proofs. Previously each view repeated full schema, graph/hash, and cross-graph binding validation. Native layout views likewise parsed the entire layout for every page. An immutable Evidence root alone was insufficient: its selected normalized result did not carry the JSON owner's private sealing proof.

The JSON owner now transfers that proof to a selected own member without another walk or per-node registration. Native layouts and imported JavaScript results record a successful full validation of the exact immutable source in their owning WeakSets. Already owned JavaScript graphs keep the existing metadata normalization and binding checks, returning parsed metadata with their original graph references. That path avoids rewalking a large owned graph just to borrow its wrapper. There is no second retained parsed graph, strong cache, cross-process cache, or live-state cache. The schema implementation is fixed for the process; a different source object must validate independently. Weak entries do not keep Evidence alive, and session lookup continues to gate access after close.

Cache admission follows full validation and requires that parsing preserved the source representation. A shared iterative JSON comparison checks own data keys, scalar values, prototypes, and enumerability without calling accessors. Object member order is immaterial; array order remains exact. Two-way identity maps visit preserved descendants once and reject source-node splits or merges. Both comparison maps are temporary and discarded on return. Parser defaults, hidden-field removal, or other normalization retain the parser's result and do not cache the source. Writable/configurable descriptor differences are immaterial: the source is sealed while the temporary parse is mutable. The JSON sealing boundary rejects proxies before reflection, so arbitrary mutable proxy behavior cannot acquire an immutable proof. Mutable, merely Object.freeze'd, failed, and abandoned inputs do not qualify for successful format reuse.

No Evidence identity, source linkage, coverage, unknown, omission, error, or caller-selected page changes are intended. Complete source Evidence remains accessible through the existing contracts. The first imported view still pays for full validation and the representation comparison; a one-off CLI invocation cannot benefit from a cache in a previous process.

Inputs and measurement scope

Baseline: 2c4338db1fdb7773e05376301c558cff2a0b4b80 (REA 6.6.0). Node 24.18.0, npm 11.16.0, Linux x86-64 (AMD EPYC 7B12, kernel 6.1.0-53-cloud). The ELF producer reports pwntools 4.15.0, pyelftools 0.33, and Unicorn 2.1.2. The baseline's built runtime was preserved before production changes. Candidate and baseline use the same locked packages.

• SQLite /usr/lib/x86_64-linux-gnu/libsqlite3.so.0.8.6: SHA-256 2e6eef9a727f081f0d453b4e5e6cbd8b9ef8b6f86cbf7681cbad444d3b0b55c8; 26 sections, 1,479 symbols, 2,963 relocations. Original history: 1,745,224 bytes.
• Shipped npm 11.16.0: inventory digest f8e576a7d162542c616168b25a95a5f62b9accf29467cc37e149905ad4b8e603; five explicitly selected paths (package.json, bin/npm-cli.js, lib/cli.js, lib/cli/entry.js, lib/commands/install.js), 2,822 application nodes, 1,796 semantic nodes, 13,682 semantic unknowns. The 1,416 unselected files stay unknown. Original history: 21,058,430 bytes.
• Small case /usr/bin/true: SHA-256 c79bf44242829108e323378531f4ac839513ca1fba45efd6583643526e1e9fd2; 31 sections, 53 symbols, 66 relocations. Produced through the real production inspect_binary_layout MCP operation. An independent raw ELF reader checked every symbol name, value, and source range plus section/relocation counts.

The two larger histories were produced and independently checked in the linked previous audit. They are historical Evidence, not a claim of fresh provider analysis during a resumed view. All tasks here are static; no target executable is run.

Each completed comparison includes startup, the complete 207-tool catalog, import, queries, oracle checks, export, close, receipt writing, and SDK cleanup. RPC timings exclude subsequent client-side oracle/receipt work. Fresh servers separate first-view validation from later pages; OS file caches are uncontrolled. Runs alternate baseline/candidate order. One CPU, 8 GiB hard family memory, 7 GiB early stop, no swap, and a 192-task ceiling bound each serial task. Resource counters are observed at payload exit while the observer remains alive; final whole-scope counters are unavailable. Kernel cgroup peaks include charged file cache and descendants, rather than just the server's JS heap.

Results and executed validation

Candidate production code: 3930a640fb8387cbd3079eee13fb90f82d1881dc. The curated receipt retains every final run, offline token estimates, compatibility checks, and costs of previous attempts. Values below are medians with min–max ranges across three fresh alternating pairs.

Source	First page, before → after (ms)	Eight later pages, before → after (ms)
SQLite ELF	183.56 [170.13–244.27] → 209.05 [203.11–216.88]	469.89 [393.47–517.92] → 85.04 [65.10–89.01]
npm	2441.08 [1681.26–2479.69] → 1978.73 [1864.89–2720.54]	15534.41 [10378.13–16864.91] → 82.69 [75.19–84.13]
Small ELF	133.33 [100.81–242.58] → 226.60 [111.79–556.22]	122.93 [76.38–299.47] → 62.58 [57.69–96.24]

Source / runtime	Whole elapsed (s)	Family CPU (s)	Kernel family peak (MiB)
SQLite ELF / before	19.08 [15.26–20.00]	18.75 [15.01–19.55]	1013.8 [1009.5–1058.6]
SQLite ELF / after	14.38 [14.06–17.25]	14.09 [13.77–17.00]	984.0 [957.1–1047.1]
npm / before	39.63 [31.21–40.86]	39.27 [30.87–40.49]	1338.0 [1301.0–1543.7]
npm / after	20.71 [20.01–23.02]	20.50 [19.73–22.23]	1286.8 [1250.6–1344.9]
Small ELF / before	14.30 [12.00–17.59]	13.94 [11.77–17.31]	1013.2 [937.8–1017.4]
Small ELF / after	12.31 [11.90–16.25]	12.04 [11.71–16.07]	956.3 [931.4–1007.5]

Source	Paired warm-read ratio, before / after	Paired first-view change, after − before (ms)
SQLite ELF	5.53× [4.42–7.96]	+25.49 [-41.16–+46.75]
npm	184.65× [125.50–224.30]	-462.34 [-614.80–+1039.29]
Small ELF	1.28× [1.22–5.19]	+10.98 [-15.97–+422.88]

Each pair matches the before/after runs with the same repetition number. The paired effect can differ from the ratio of marginal medians when the host speed changes. Positive first-view changes mean higher latency.

The repeated-read benefit is measured per source and pair above. First-view validation also performs the representation comparison and can be more expensive. Whole task, CPU and memory figures above are measured SDK workflow costs, not a general provider-analysis or agent-investigation speedup. The previous e30 checkpoint had a slow npm repetition on both sides, including a 65-second candidate task; its cause was not profiled. The original checkpoint and every fresh run remain retained. The previous 646312dad checkpoint also had slower small-case warm reads in two of three pairs (135.54 → 373.14 ms marginal medians). Those negative results remain available in its linked historical receipt; the current cohort improves all three small pairs, without establishing a general small-input or whole-task benefit. No model-token saving, general memory improvement, or complex-software investigation advantage is established.

All 18 fresh workflow runs preserve original records and canonical answer digests. Every corresponding query's structured result and complete catalog is equal. The three public CLI source pairs are also byte-for-byte identical. The two additional mixed-history tasks preserve all 97 original records, check 64 follow-up heads over 32 pages and revision two, reject stale updates and tampered imports atomically, clear Unknowns on close, and refuse old references.

Offline tiktoken 0.12.0 / cl100k_base estimates for one pass over the complete catalog, named-call arguments and replies are unchanged within each pair: 849,736 native, 856,384 JavaScript, and 840,303 small-case tokens. The catalog contributes 805,215 tokens alone. These are raw transport serialization estimates, not actual model context or usage; initialization/framing, prompts, client discovery, rendering/replay, and continuations are unmeasured. Metadata paths differ from the historical checkpoint; compare equivalent runs within each checkpoint.

Executed checks at the final production code:

• npm run check and npm run docs:check: pass, including five README consistency tests.
• Three latest-main integration files through npm run test:focused, using default deadlines: 27 tests pass at the current production source.
• 130 source tests across immutable JSON, native/JavaScript format validation, and the shared application view service; npm run check:fast and npm run build:cached pass.
• Four compiled CLI/MCP files (evidenceBundleInline, serverUnknownsMcp, analysisViewMcp, and analysisViewCli): 36 tests pass.
• Real public stdio MCP: 18 fresh paired tasks and two additional mixed tasks; public CLI: six invocations, with three exact output pairs. These corpus workflows and four-file boundary checks have exit zero, verified payload cleanup and zero observed payload-exit OOM events.

Internal and automated review reproduced candidate-only defects before integration. Generic deep equality ignored hidden fields removed by parsing; the JSON owner comparison now preserves that normalization. Review also reproduced aliases crossing preserved subtrees and cloned metadata; four regressions failed against the prior source. The comparer now visits preserved descendants with two-way identity mapping. Already-owned JavaScript graphs keep their parsed-metadata fast path, so the generic comparison and full-result cache only apply to the expensive imported fallback. Its initial iterative traversal also repeatedly expanded shared nodes: a valid 27-object graph exhausted an instrumented 10,000-inspection budget. Tracking source-to-parsed node pairs reduces that same compiled probe to 54 inspections and successful borrowing. A parser that splits or merges source nodes keeps its parsed result rather than expanding repeated paths. Deterministic regressions enforce the inspection budget instead of relying on timing or letting a regression consume exponential CPU. These are internal JSON-owner probes, not claimed baseline public-provider reproductions.

The initial exploratory cohort and historical e30 paired checkpoint are not the final comparison. One earlier source-check attempt was mistakenly started during an initial baseline run; the newly started owned scope was stopped after 17.27 s and that check remains incomplete (exit 241). Observed counters, the contended baseline, subsequent successful retries, and prior checkpoints remain retained; no OOM or terminal-counter claim is inferred from that stopped task. Two later tasks stopped because the host-available-memory guard reported its early-stop condition: candidate JavaScript repetition three and baseline small repetition three. Both incomplete attempts and the unmatched completed JavaScript baseline remain separate. After memory recovered, both sides were rerun as fresh pair four; final JavaScript and small samples use pairs 1, 2, 4. Neither interrupted task is a completed comparison. The guard later gained read-only host-memory/cgroup receipt fields; all resource thresholds and its observer/stop logic remained unchanged.

Real Ghidra, Hopper, JEB, fresh JavaScript producer performance, and actual agent trials are outside this checkpoint. Private WeakSet metadata/heap retention was not measured independently; the absence of a strong retained parsed graph is source-derived. The comparison maps exist only during the first interpretation; measured family peaks include their temporary cost.

Latest-main integration and an unexplained timeout

Merge 71e0b2739 brings in main bb212ffec's three test/config/guide fixes. The current source passes all 27 tests in those three integration files through npm run test:focused with the default deadlines. The earlier 646312dad checkpoint initially passed 26 and timed out the live identity/catalog test at 30 seconds. Matching official-command controls subsequently passed all 27 on both unmodified main and that candidate. Direct/extended-deadline diagnostics are recorded separately. Their passing results do not establish the cause of the original timeout, which remains preserved and unexplained. No test-policy or product file was changed to obtain a pass.

Reproduction and retained evidence

Use retained-investigation.mjs from this directory with an authentic exported bundle and its original source Evidence ID. Build both selected runtime revisions with the pinned toolchain, preserving a separate baseline. For each runtime run:

python3 experiments/analysis-efficiency/resource-run.py --observe-payload-exit -- \
  node experiments/analysis-efficiency/retained-investigation.mjs \
  --runtime-root /absolute/runtime --revision REVISION \
  --bundle /absolute/history.json --evidence-id ORIGINAL_EVIDENCE_ID \
  --scenario native --delivery focused --output /absolute/new-receipt-directory

Use --scenario javascript for the npm history, and --verify-mixed with an actual mixed investigation. Reuse the same bytes and source identity on both sides. Do not regenerate parent identities or synthesize bundle records. The checked-in wrapper defaults to a 4 GiB hard / 3 GiB early stop. This checkpoint used a private resource-only copy with 8 GiB / 7 GiB thresholds, a 16 GiB available-memory start floor and 8 GiB host early stop. Its source and unchanged observer hashes are recorded in the curated receipt. This local resource policy does not change the public workflow or global wrapper defaults. Raw histories, transcripts, oracle outputs, and resource logs stay in the private local audit cache; only curated receipts are committed.