rea/experiments/analysis-efficiency/native-inventory-views.md

Native inventory summaries and retained rows

Continues #4249. The analyst question is to report observed row counts and the first four address/value pairs from procedures, names, strings, and literal main procedure/string searches. Complete delivery reads every row even when only those facts are needed. A real jq name inventory exceeds the existing 10 MiB MCP result budget.

Change and preserved semantics

The five existing tools accept opt-in detail: "summary". Their original complete Evidence is retained before a summary is returned. Delivery detail does not enter provider parameters or the original Evidence identity. Default complete requests keep their existing response. Oversized complete inventory replies gain an actionable retained-page continuation after recording succeeds.

inspect_analysis_view projects native-inventory rows with offsets, limits and selected fields. Parent Evidence ID and original source_index identify duplicate-address alternatives; original order is preserved. Address is always retained, and selecting a string's value also retains its string/decoding qualifications. Counts describe retained provider observations. Exhausting a page does not establish complete code/string discovery. Unknown source artifact identity stays unknown rather than acquiring an invented path.

CLI and MCP use the shared application/domain projector. Its source-format validation is reused only after successful representation-preserving validation of an authenticated immutable JSON snapshot. Two weak sets distinguish names and the stricter strings format. They retain no second parsed inventory or live provider state; entries disappear with their source arrays. Mutable, foreign, malformed or transformed values do not acquire this proof. A parent change rebinds provenance rather than reusing an earlier view envelope.

This reduces delivery after the provider has produced its full inventory. It does not reduce Ghidra discovery, eliminate the retained full inventory, or establish a universal CPU/memory improvement. The catalog becomes larger and summary-plus-page requires more RPCs than a small complete reply.

Real inputs, controls and oracle

Read-only Linux x86-64 targets, without target execution:

Target	Bytes	SHA-256	Provenance
true	35,664	c79bf44242829108e323378531f4ac839513ca1fba45efd6583643526e1e9fd2	Installed Debian coreutils 9.1-1
jq	2,267,912	b1c22172dd303f3be49e935aa56aa48a8b7a46e0bc838b4997d3bb451495870f	Official jq 1.8.2 asset, release API size/digest matched

Baseline runtime is main 58d6043303e722115f1da18b9505c24abeeab423. The candidate runtime is frozen from that base plus this implementation before measurement. Runtime file manifests, the exact patch, driver revisions, original Evidence, transcripts, errors and resource observations are retained outside Git. The curated receipt pins the recordings. Later rebases do not relabel these timings as measurements of a different base.

Node 24.18.0/npm 11.16.0, MCP SDK 2.3.1, Ghidra 12.1.4 and Eclipse Adoptium 21.0.12.1+1-LTS. Each pair uses a fresh MCP server and Ghidra analysis, followed by a warm repetition in that session. OS caches are unflushed. Java heap is 768 MiB and one active processor; Node heaps are 2 GiB. The family guard uses one logical CPU, 4 GiB hard/3 GiB early-stop memory, no swap, and 192 tasks. Heavy jobs are serial. Kernel peaks include file cache, controller verification and receipt work, rather than isolated provider heap. Payload-exit observations verify owned process cleanup; final whole-scope counters remain unavailable.

The live named SDK contracts validate replies. Both modes discover the full catalog, open the same target, complete the question twice, export original Evidence and close. Summary mode requests four rows only for nonempty results. Complete mode includes acknowledged export recovery and local JSON projection when list_names exceeds transport capacity. Its errors, export calls, bytes read and local projection output remain in costs. Model generation of that local program and filesystem-tool schemas are unmeasured.

The small pair preserves every field of every original source Evidence, including IDs, and both phases' complete selected fact vectors. The large completed pair compares the requested facts exactly; it must not claim identical independent Ghidra analyses. An earlier candidate attempt found alias-primary and two generated switch-label differences outside the requested selection. Their underlying cause is unestablished. To qualify projection independently of that variation, the portable verifier imports the actual baseline Evidence, reads every canonical row through MCP, and compares it with the recorded source. It also compares actual CLI file-source views with retained MCP views, without Ghidra/JDK configuration, and reexports original source records unchanged.

Measurements and limits

Each recording contains discovery, both first/warm questions, original Evidence export and cleanup. The token proxy below adds input-facing catalog JSON, name/argument JSON, text replies and local recovery output; it does not count both text and structured copies as if every host supplied both to a model.

Recording / metric	Before	After
true: RPC calls	13	23
true: RPC result bytes	921,748	171,130
true: combined token proxy	306,496	208,002
true: first question	35.534 s	30.791 s
true: warm question	0.236 s	0.224 s
true: whole scope wall	53.662 s	36.777 s
true: family CPU	44.202 s	35.483 s
true: kernel family peak	1103.9 MiB	1186.5 MiB
jq: RPC calls, including export recovery	15	21
jq: RPC result bytes	3,032,150	172,180
jq: combined token proxy	590,032	208,302
jq: first question	334.659 s	318.798 s
jq: warm question	2.521 s	6.916 s
jq: whole scope wall	355.586 s	333.104 s
jq: family CPU	318.976 s	311.319 s
jq: kernel family peak	1498.2 MiB	1692.3 MiB

The catalog remains complete at 209 tools: 3,263,492 → 3,293,844 bytes. Its input-facing projection grows by 642 proxy tokens, from 185,367 to 186,009. Including that increase and every call, the combined serialization proxy falls 32.1% for true and 64.7% for jq. Actual model usage is unmeasured.

Preserve the negative results: the large warm question is slower and both family memory peaks increase. The first native view also pays source-format validation and SDK schema setup; these phases are not isolated by this driver. Fresh analysis dominates total wall time and varies between sessions. These observations do not establish a latency, CPU or memory improvement. The owning benefit is selected delivery with complete source access and actionable recovery, rather than eliminating full provider inventory production.

Individual completed samples, tokenizer estimates and portable checks are in the receipt. A cl100k_base estimate over recorded serialization is not actual model usage. Counts include discovery, call arguments, text replies, errors, continuations and local recovery output; SDK result JSON is also reported separately. Actual client schema selection/rendering, wire framing, model reasoning and isolated correctness-oracle overhead are unmeasured. Single pairs are observations rather than latency distributions or evidence of improved model investigation outcomes.

Incomplete attempts remain separate. The first baseline jq driver stopped at the real oversized reply because it lacked export recovery. The first candidate jq attempt produced its selected facts and source export, then hit the 3 GiB early-stop threshold while comparing differing full inventories; no final report or payload-exit cleanup proof was produced. Expanding the enormous assertion diagnostic is a source-derived explanation, not an observed final stack trace. The harness now emits a bounded mismatch with source digests while retaining strict equality and both full files. The subsequent recording explicitly does not request cross-session full-source equality.

Two static-check attempts were stopped by the host-memory guard; later checks passed after host memory recovered. No guard-stopped attempt is a successful performance sample. Engine-version/platform equivalence, other providers, complete application reconstruction and post-GC retained heap are unqualified.

Executed verification: 24 focused files/305 tests passed, including advertised SDK-converted JSON Schema validation, all five tools, nonretaining refusal, oversized recovery, same-address alternatives, decoding qualifiers, immutable validation ownership and unknown-artifact sources. Full npm run check passed, including module/architecture guards; npm run docs:check also passed. The portable runs passed with 27/444 MCP calls, ten actual CLI comparisons each, every row reconstructed, original records preserved and payload cleanup verified. Their scope times are 25.194/77.537 s and kernel family peaks are 978.4/1940.2 MiB; these additional correctness runs are not part of the selected-question timing pairs.

After rebasing onto main 0134732fc (including Protobuf/gRPC views), 28 focused files/325 tests passed. Final field-selection normalization removes repeated per-row scans of duplicate requested fields while retaining the exact request; the affected two files/nine tests passed again. Final npm run check and npm run docs:check passed. Both real old-source portable checks passed again against the final build, including every canonical row and all 20 CLI/MCP comparisons. The observed catalog is then 209 tools/3,317,414 bytes. Those final source and dependency pins are separate from the earlier frozen timing/token pairs: final-main performance pairs and actual model usage were not rerun.

Reproduce

Build separate baseline/candidate runtimes, then run each mode with the same absolute artifact, Ghidra and JDK paths:

python3 experiments/analysis-efficiency/resource-run.py --observe-payload-exit -- \
  node experiments/analysis-efficiency/native-inventory-views.mjs \
  --runtime /absolute/baseline --target /absolute/application \
  --ghidra /absolute/ghidra --java /absolute/jdk \
  --mode complete --output /absolute/before

python3 experiments/analysis-efficiency/resource-run.py --observe-payload-exit -- \
  node experiments/analysis-efficiency/native-inventory-views.mjs \
  --runtime /absolute/candidate --target /absolute/application \
  --ghidra /absolute/ghidra --java /absolute/jdk \
  --mode summary --output /absolute/after

python3 experiments/analysis-efficiency/resource-run.py --observe-payload-exit -- \
  node experiments/analysis-efficiency/native-inventory-portable.mjs \
  --runtime /absolute/candidate --source /absolute/before \
  --output /absolute/portable

--control /absolute/before additionally requires every source Evidence field to match across the two live analyses. Preserve failure if the engine produces different observations. --verify-pages performs extra live full-page checks; keep those correctness runs distinct from selected-question timing pairs. With private tiktoken==0.12.0, native-inventory-costs.py --output FILE CASE... counts completed recordings and their sibling .log resource observations.

The subsequent main 4eaf3a117 oversized-error protocol-framing fix was also merged into the branch. Its producing-server, native inventory and large-result checks passed (three files/11 tests), preserving both protocol revisions. The real old true source verifier and npm run docs:check passed again. The native projector and earlier performance pins are unchanged; latest-head pre-push/CI checks provide the final branch status.