GitNexus/gitnexus/bench/tool-accuracy

Deterministic tool accuracy

This small release corpus implements the cheap tool-level checks proposed in discussion #3493, comment 18772530. The fixed answers come from local fixture source and public issue repros; the current engine output never defines the expected answer.

With the repository's dependencies installed, run from gitnexus/:

node --import tsx bench/tool-accuracy/run.ts --check --out bench/tool-accuracy/results
npm test -- test/unit/tool-accuracy-score.test.ts

The runner uses the real runPipelineFromRepo, CSV/database loader, metadata writer and LocalBackend.callTool. It indexes one small temporary TS/Python repository with PDG enabled and one parser worker. Community/process discovery is unnecessary for these checks and is disabled. Native parser workers require the normal build produced by npm ci; run npm run build after worker changes. The registry, database and fixture copy are isolated and removed after the run. Fixture code is parsed, never executed. Extension installation is disabled, accidental JavaScript fetches fail, and neither embeddings nor paid models run.

accuracy.json contains source SHA, working-tree status, fixture SHA-256, issue IDs, expected/actual answers, precision/recall and raw tool outputs. accuracy.md gives the same fixed-answer verdicts and concise failed answers. --out selects a report directory. Without --check, valid observations can be reported while a new fixed-answer failure is investigated; execution/output errors always exit nonzero. --check additionally enforces the regression gate.

Fixture / checks	Public repro	Fixed-answer contract
rename/	#3486	All three Writer.close references; zero edits to comments, strings or unrelated homonyms; typed DeferredWriter.flush parameters refuse without edits or source changes.
src/server/, test/stub.test.ts	#3487	Server beats a same-path test stub; app.all has method *; Map.get is absent; helper-registered literal route is present.
src/web/	#3488	Both the base-URL template passed to a configured wrapper and a direct literal fetch link their consumer files.
src/security.ts, src/server/inline.ts	#3489, #3490, #3491	Named and inline unsafe handlers reach the sink; guarded and constant commands have zero taint; statement and function-hop lines are 1-based.
pkg/	#3497	Direct, chained module-return helper and local module-return helper resolve to the same target.
mod.py, callers.py	#3498	Direct, conditional, boolean and module-qualified annotated factories resolve to C.m.
nested*.py, local_import.py	#3499	Imports resolve within their lexical scopes; unrelated nested declarations and sibling-local imports cannot capture bare calls.
summary.py, 501 deterministic drop records	#3497	Production summarizeUnresolvedReceivers retains 500 names and reports one omitted name; querying a capped-out name requires lower-bound, while a complete empty summary permits exact.

The 501-name case supplies known suppressed outcomes to the production summary writer, persists its actual output as metadata and calls real impact. It tests the cap/consumer contract without requiring 501 parser fixtures. It does not claim parser coverage for those synthetic records. Python checks read actual context.outgoing.calls symbol UIDs, preserving lexical/class ownership.

Rename's positive fixtures construct each Writer receiver directly, including the imported class. A separate DeferredWriter.flush fixture exercises a typed parameter: a TypeScript annotation alone cannot prove which structurally compatible runtime object owns the method. The public rename tool (dry_run: true) must refuse it with planning_status: "unsupported" and code: "ambiguous_reference", zero edits, and unchanged source files. This safety check has no known-gap allowance. Fixture line anchors are updated together; the three exact semantic answers and the same-named method/function, comment and string controls remain required. The runner validates v2 semantic provenance, exact UTF-16 spans and before/after source lines, then scores the whole after line. Semantic rename does not use ripgrep or text-search edits.

Known gaps and repairs

The initial source baseline is dd096e690c06e5d54ab3595c63b83f15e04b6191: 11/31 fixed answers pass, 20 fail. Main at 50aa4be3b2c2c9a1561fc44878e5d8f87b99d16e repaired three #3487 route checks and all four #3499 checks, and their allowances were removed: 18/31 pass, 13 fail. The JSON report records every desired and actual answer independently of gate status. Inline-callback taint now passes and its allowance is removed. Guard taint remains a documented limitation and a visible known gap.

known-gaps.json is an explicit, reviewed list of failing check IDs with issue, reason and observed baseline answers. There is no wildcard, skip, automatic baseline-update command or production-engine fix in this corpus. The gate:

• Rejects new failing checks, unknown/malformed allowances and missing/malformed tool outputs, exceptions, truncation and incomplete fixture extraction.
• Rejects a previously correct fact disappearing or a new incorrect fact appearing inside an existing gap. Partial repairs can still proceed.
• Rejects a passing check whose allowance remains: remove the allowance in the PR that repairs the case, so the fixed answer protects future releases.

Changing a fixture requires reviewing its fixed answers and the line anchors in FIXTURE_ANCHORS (expectations.ts); the runner and a unit test fail when an anchor no longer matches its fixture line. Broadening an allowance requires review of the actual repro; never change an expected answer simply to match current output.