[corpus] Resolve Python cross-function `batch` FN: relabel sibling helpers or accept as unrecoverable

[AndresL230]

Migrated from recost-dev/extension#134 — the original explicitly flagged that this corpus-labeling decision belongs in the repo where the fixtures live.

Context

Follow-up to recost-dev/extension#117 (FN #1) and recost-dev/extension#132. #117 bundled two C1 false negatives; #132 recovered the unbatched_parallel FN but deferred the Python cross-function batch FN after attempting and reverting a fix.

The reverted attempt (a module-scope (provider, methodChain) batching pass in recost-dev/extension's src/scanner/python-waste-detector.ts) proved this is not a detector gap — it's a corpus-labeling inconsistency. This issue tracks the corpus-side decision that #117's original "design sketch" did not anticipate.

The problem

The corpus labels exactly one batch FN:

• flask-mixed-providers/src/providers/anthropic_helper.py:11 — two calls to _client.messages.create(...) across two module-level functions (summarize, summarize_with_style).

But sibling helpers in the same fixture — openai_helper.py, cohere_helper.py — have structurally identical two-calls-across-two-functions shapes and are not labeled as batch findings. Any cross-function pass precise enough to fire on the Anthropic helper also fires on the siblings → 6 FPs, batch precision drops to 14.3%. No AST signal distinguishes the labeled case from the unlabeled-but-identical ones (would require cross-file call-graph signal the scanner doesn't currently expose).

Decision needed — pick one

Option A — Relabel (recover the FN). Add batch-finding labels to the structurally-identical sibling helpers (openai_helper.py, cohere_helper.py, and any others) so the cross-function pass becomes precise. Then re-implement the reverted (provider, methodChain) module-scope pass in python-waste-detector.ts (in recost-dev/extension).

Option B — Accept (drop the FN from the recall target). Formally mark the anthropic_helper.py:11 batch case as unrecoverable without cross-file call-graph signal in the corpus, removing it from the batch recall denominator. python-waste-detector.ts stays byte-identical to main.

Acceptance criteria

• Decision recorded (A or B) with rationale, here in extension-benchmark.
• If A: sibling helpers labeled consistently; the reverted module-scope pass re-implemented (in recost-dev/extension); npm run benchmark shows batch at TP 1+ / FP 0 / FN 0; no regression to the 7 existing C1 tests; per-type precision stays 100%.
• If B: corpus annotation marks the case unrecoverable; batch recall target adjusted; baseline.json regenerated so the row is no longer counted as an outstanding FN.

Do NOT

Re-add the (provider, methodChain) pass without first resolving the labeling — recost-dev/extension#132 already proved it regresses batch precision to 14.3% on the current corpus.

Links

• Original: [benchmark corpus] Resolve Python cross-function batch FN: relabel sibling helpers or accept as unrecoverable extension#134
• Parent: [Findings] Recover the two C1 false negatives — cross-function batch + inline-parallel unbatched extension#117
• Reverted attempt + analysis: feat(wave4): recover unbatched_parallel FN; defer Python batch FN (#117) extension#132

[AndresL230]

Blocker found: strict Option A cannot satisfy its own acceptance criteria

While scoping the implementation, I hit a hard contradiction in Option A and want it on the record before anyone picks it up.

Option A's mechanism collides with the existing C1 PR-3 test. To recover the anthropic_helper.py:11 FN you need a cross-function (module-scope) batch pass. But extension/src/test/fixtures/c1-pr3/py_diff_functions.py is structurally identical to the flask helper, and c1-pr3-batch-tightening.test.ts:50 asserts it must produce zero batch findings:

# flask-mixed-providers/src/providers/anthropic_helper.py   ← A wants this to FIRE
_client = anthropic.Anthropic()
def summarize(text):            _client.messages.create(...)   # L11
def summarize_with_style(text): _client.messages.create(...)   # L20

# extension/src/test/fixtures/c1-pr3/py_diff_functions.py    ← C1 asserts this must NOT fire
_client = anthropic.Anthropic()
def summarize(text):  _client.messages.create(...)
def translate(text):  _client.messages.create(...)
def explain(text):    _client.messages.create(...)

Both are "N module-level functions, one anthropic.messages.create each, shared module client." The only difference is 2 vs 3 functions — and 3 is the stronger batch case. Any pass precise enough to fire on the flask helper necessarily fires on the C1 fixture → that test fails → which Option A's own criterion ("no regression to the 7 existing C1 tests") forbids.

This is exactly what the revert commit aefb3df ("revert(wave4): back out cross-function Python batch pass") stated: "no AST signal separates TP from FP." I confirmed there is none — the cases are equivalent in every signal the detector sees (provider, methodChain, module-level client, by-function bucketing). The detector's detectSequentialBatching even buckets by (provider, enclosingFunction) on purpose, with the comment "calls in different functions can't be batched together — they execute on independent code paths."

The three real options

1. Option B (recommended). Accept anthropic_helper.py:11 as unrecoverable without cross-file call-graph signal; drop it from the batch recall denominator; regenerate baseline.json. Corpus + baseline only, no detector change, no C1 regression. Note this raises global findingRecall 0.667 → 1.0 (the FN currently drags it down). Also consistent with reality: summarize and summarize_with_style are independent route entry points that are never co-invoked, so they aren't batchable even with call-graph signal.
2. Option A-modified (accept the cost). Relabel the flask siblings as batch findings and flip the C1 PR-3 Python cross-function assertion to expect firing. Recovers the FN on the benchmark, but reintroduces cross-function batch detection that fires on any multi-function helper module in real code — the precision problem #132 reverted for. Knowingly overrides the "no C1 regression" criterion.
3. Add cross-file call-graph signal. Out of scope per this issue, and per point 1 it wouldn't even recover this particular FN.

Status: parked pending a decision. The A3/A5 fixture work (#7) shipped independently in #9 and is unaffected by this.