rea/experiments/analysis-efficiency/apktool-string-tables.md

Apktool shared decoding and table-only projection

The measured question is the same in every comparison: inspect an APK's metadata/manifest and its default, German and Brazilian Portuguese string tables. One existing batch removes repeated decoding. An optional table-only projection then removes repeated primary strings without changing the selected facts.

This is a static resource-query measurement, not an agent evaluation or a claim about reconstructing an application's behavior. The machine-readable measured summary is apktool-string-tables.json.

Revisions, inputs and execution

• Baseline: 8f7d12e30d63f3d3b4cfe2a90336362841438f56, latest main at issue start.

• Projection source: 406af90023a545fb280601bf51f445e122c36378.

• Existing shared decode: #4777, integrated by #4804.

• Projection follow-up: #4982.

• Linux x86-64, kernel 6.1.0-53-cloud-amd64, Node 24.18.0, npm 11.16.0, Temurin 21.0.12.1, Apktool 3.0.3. The selected JAR's closeout digest is dbf930b076c6b9be08d57c449cacefc3bdd6b71ebd59b3066fc0e1f5b14f9423; decoder contents were not independently hashed for every request.

• newpipe: 11,507,122 bytes; SHA-256 18447bfb1e06d113edc88df93f471827280de06f6f3d4dc42f56f28b9c1bab79. Valid projected counts: 1126/833/779; unsupported entries: 1/1/1.

• apidemos: 6,521,217 bytes; SHA-256 a9eecf37b26cd084855c530db81c2bb1b91f4c1b095a04f47aa7c20e2791f686. Valid projected counts: 967/29/29; unsupported entries: 8/0/0.

• settings: 3,099,589 bytes; SHA-256 e4024b3ee348929252bcaf5eb422bcab9e4ace96703499a84f3332281c03375f. Valid projected counts: 25/19/19; unsupported entries: 0/0/0.

The counts are default/de/pt-BR respectively. Source paths, XML digests, exact caller paths, runtime hashes and per-run identities/costs are in the JSON summary. The APKs were never installed or executed. Input digests were checked before and after each completed task; provider-owned workspaces were confirmed absent.

Each task uses a fresh production stdio MCP runtime, complete paginated tool discovery, client inspection, resource call(s), close_binary, transport close and owned cleanup. OS file caches were not flushed. Every decode has a fresh JVM, APK snapshot, framework directory and workspace: these are not warm retained engine/database queries, nor cold-disk measurements.

Tasks run serially on one logical CPU, with an 8 GiB family hard limit, 7 GiB early stop and no swap. Admission requires 16 GiB host availability; running tasks stop below 8 GiB host availability. CPU and kernel memory.peak cover the complete owned cgroup, including subprocesses, observer and file cache. Counters are captured at payload exit while the observer is alive; disappeared final whole-scope counters remain unknown. All completed payload receipts show successful owned cleanup and zero observed OOM/limit events at that checkpoint.

Correctness and owning change

Separate native decodes were retained for independent Python Expat validation. It checks direct string entries, source entity spelling, unsupported markup counts, ordering and emitted prefixes against the XML bytes. Manifest text and package identity are checked separately. All six serial/batch pairs agree on selected strings, manifest, metadata and target identity. All nine projection pairs additionally preserve table counts and coverage; batch/table-only counts and coverage match the XML oracle. This qualifies REA's projection of these native outputs, not independent binary ARSC decoder completeness.

The existing default batch repeats its primary table in top-level strings and string_tables[0]. On NewPipe this made the full resource replies larger than three serial replies despite reducing decode calls: 616,059 → 680,188 bytes. ApiDemos' larger common manifest produces the opposite tradeoff. The existing batch remains the compatibility default.

string_tables_only: true (CLI --string-tables-only) keeps the complete requested table facts inline in string_tables and leaves strings empty by request. An unrequested field does not establish that a source table is empty. This also supports one table without locales. include_strings: false still suppresses string projection. Omitting the new option preserves old requests: three real legacy workflows produce identical complete MCP replies, including Evidence identity, at the new source revision.

The resource projection owner computes the output choice and charges aggregate entries for the copies actually emitted. The same 10,000-entry policy can thus return five distinct 2,000-entry tables rather than consuming capacity on a duplicate primary copy. Per-table limits, XML budgets, physical alias reuse, coverage, partial observations and cancellation remain at the existing owner. CLI and MCP share this implementation; there is no new tool or cross-call cache.

Existing shared decode: three serial calls versus one batch

Three alternating fresh-session pairs per APK, all at the same main revision. The table reports marginal medians; paired speedups are reported separately. Elapsed/CPU include discovery, client/Java startup, calls, transcript/hash checks and cleanup. Native oracle creation, tokenization and verification gates are separate costs in the JSON, not hidden inside these task timings.

Target	Resource calls, s	Whole elapsed, s	Whole CPU, s	Family peak, MiB
newpipe	23.19 → 7.40	35.32 → 19.86	34.95 → 19.65	1341.7 → 1309.5
apidemos	17.43 → 6.22	30.32 → 18.85	30.05 → 18.63	1187.1 → 1202.7

Paired resource-call speedup medians: NewPipe 3.23× (2.90–3.23), ApiDemos 2.80× (2.66–3.09). Instrumentation confirms three versus one actual decoder invocations. These call timings include protocol/projection costs; they do not isolate the Java engine phase. NewPipe's whole raw token estimate increases 951,332 → 962,330 with the default batch; ApiDemos decreases 1,054,893 → 948,591. Fewer calls alone do not establish less context.

New selector: default batch versus table-only batch

Three alternating pairs per APK; NewPipe repetitions 1/4/3, others 1/2/3. Repetition 4 replaces a refused start, described below. Both variants still perform one decode and return the same required facts.

Target	Resource call, s	Whole elapsed, s	Whole CPU, s	Family peak, MiB
newpipe	9.18 → 9.35	23.95 → 24.85	23.65 → 23.82	1327.5 → 1323.6
apidemos	6.46 → 5.81	23.28 → 19.79	22.90 → 19.47	1247.8 → 1197.0
settings	4.21 → 4.06	17.66 → 19.39	17.42 → 19.04	1191.6 → 1122.4

There is no consistent latency or whole-family memory improvement established for the selector. NewPipe and Settings have higher whole-task elapsed and CPU medians. Elapsed ranges overlap: NewPipe 23.53–34.46 → 21.59–42.80 s; ApiDemos 18.55–27.07 → 19.25–25.96 s; Settings 16.55–23.39 → 16.14–21.72 s. The cause of this variation was not established; a passing or faster run does not explain a slower one. The selector's supported benefit is reduced duplicate delivery, with its full task costs retained here.

Full resource reply bytes fall 680,188 → 506,481 for NewPipe, 616,841 → 463,588 for ApiDemos and 38,725 → 34,199 for Settings.

Offline tiktoken 0.12.0 / cl100k_base estimates count one complete catalog page-array plus every call name/argument/reply, including inspection and close. Full replies include both textual and structured copies. Text-only replies are also counted separately. These are serialization estimates, not actual model usage; initialization/framing, prompts, client filtering/rendering/replay and continuations are unmeasured.

Target	Text reply tokens	Whole raw transport tokens	Whole estimated reduction
newpipe	73,661 → 55,797	962,330 → 924,352	3.95%
apidemos	65,693 → 49,162	948,591 → 913,693	3.68%
settings	4,387 → 3,942	814,859 → 814,029	0.10%

Discovery still has 207 tools; only decode_android_resources changes. The added schema costs 557 catalog bytes / 113 estimated catalog tokens (805,217 → 805,330). Raw discovery dominates this transport estimate; it does not establish what a client supplies to its model. Smaller replies and these estimates do not prove an actual agent context or investigation advantage.

Verification, attempts and limits

• 94 source tests: provider/resource projection/service, including alias reads, aggregate output exhaustion, malformed/physical failures, partial facts, cancellation and cleanup. The table-only capacity, single-table provider and CLI cases fail on the old implementation (three selected failures, 82 unrelated tests skipped).
• Nine compiled CLI/MCP tests: schema conversion/advertisement, option forwarding, Evidence parameters, legacy locale grammar and existing boolean choices.
• 36 completed real MCP tasks and three real CLI tasks. Nine exact projection oracle pairs; three legacy full-reply comparisons; alias/missing, single-table and omitted-string cases pass. CLI normalized data, subjects and Evidence IDs match MCP exactly on all three APKs.
• check:fast, build:cached and verify:test-discovery passed; the last discovered all 1,579 test files exactly once. npm run check and npm run docs:check also passed.
• An authored test group initially exceeded the 150-line lint limit (153). That gate failed after 189.31 s / 186.19 CPU s, peak 4,487,798,784 bytes; build/compiled checks did not run in that attempt. The capacity test was moved out of that group, and the complete corrected gate passed after 343.97 s / 338.67 CPU s, peak 4,311,293,952 bytes. No rule was disabled.
• projection-newpipe-after-2 was refused before scope/payload creation because host availability was below 16 GiB; a following poll showed about 10.5 GiB. It has no payload timing/counters and is not a completed measurement. The refusal stays recorded; a fresh whole repetition-4 pair ran after recovery. No existing receipts were overwritten and resource thresholds were unchanged.

Raw transcripts, native outputs, intermediate failures and per-request workspace arguments remain in ignored local audit storage; the committed JSON is a curated summary. Small Settings resources and negative time results are retained. Actual models, APK execution, other providers/platforms and general complex software effectiveness were not verified by this checkpoint.