Tooling: 1.27x faster t16 invoice-boundary replay via compresslevel=1 (identical receipt)

[kumagi]

Summary

registered t16 replay writes 31 mutated .gz fixtures at zlib level 9
  -> zlib.compress was 24.7 s of a 96.1 s profile
  -> compresslevel=1, decompressed payloads unchanged, receipt byte-identical

One-line tooling change to research/t16-invoice-boundary-audit/verify.py: gzip.compress(data, mtime=0) becomes gzip.compress(data, mtime=0, compresslevel=1). Same-fixture measurements, direct registered command, stdout byte-identical on every interpreter:

interpreter	before	after	speedup	stdout
CPython 3.11.15	174.19 s	149.51 s	1.17x	identical
CPython 3.12.13	107.41 s	84.40 s	1.27x	identical
CPython 3.14.4	67.72 s	43.90 s	1.54x	identical

The verifier's own receipt comparison passes unchanged (PASS_T16_INVOICE_BOUNDARY_AUDIT {"assembly": {"ACCEPTED": 10, "REJECTED": 16}, "cases": 54, "invoice": {"ACCEPTED": 26, "REJECTED": 28}}), and python3.12 scripts/research_package.py --package research/t16-invoice-boundary-audit ends PASS_REGISTERED_RESEARCH_PACKAGE ... outcome=completed. No input, fixture, case, control or receipt byte changes; only the compression level of freshly written work-directory fixtures.

Reviewer checklist

• scripts/check_research_commit.py --package research/t16-invoice-boundary-audit -> PASS_RESEARCH_COMMIT_BYTES (16 files at 95568ae8).
• scripts/research_package.py --package ... --check-only -> PASS_PACKAGE_METADATA_AND_PINS.
• git diff HEAD~1 --check clean (no whitespace errors).
• Direct replay passes on 3.11.15, 3.12.13 and 3.14.4 with identical stdout.
• Registered replay passes on 3.12.13 through the generic runner.
• Hosted CI: the Registered research packages run (38939671806) is still queued; I will fix any failure it reports.

Evidence

• Profile (cProfile, CPython 3.14.4): zlib.compress 24.7 s / 31 calls, json.raw_decode 10.8 s / 1,968 calls, vendored invoice_allcold.recount_provider 40.7 s cumulative / 52 calls, fractions ~20 s cumulative.
• This patch removes only the first item. The vendored engine is untouched; the remaining raw_decode cost needs the author's decision (a parsed-source cache across cases would have to respect the per-case mutation writes).
• Byte-identity policy: the repository already compares decompressed payloads rather than container bytes across Python/zlib versions (community reproducibility audit), so the level change does not alter any proof-relevant content.
• The package does not self-pin its own verify.py bytes (no MANIFEST/PINS self-check).

Merge danger

Door: two-way. One line, removable without changing any supplier, receipt or claim. Blast radius: isolated to this package's registered replay runtime.

Convention check: many tooling PRs add a separate research/*-fast-* package, but direct patches to an existing package verifier with an identical receipt are also merged precedent - e.g. #2357 modifies research/complex-center-quarter/verify.py (with re-pinning), #2409 modifies vendored audit code and manifests, and #2100 modifies scripts/experiments/ternary_side.py. This package does not self-pin its own verify.py bytes, so no manifest update is needed. If the maintainers prefer the package form here, I can repackage the same measurement instead.

Credit: Tairitsu-Aya for the package and its 54-case mutation design. Prepared by @kumagi with substantial AI assistance (DeepSeek V4.1 Flash in the OpenCode harness). Apache-2.0.

[kumagi]

Review-readiness update (no code change since head 95568ae8):

• Completed the interpreter matrix with the before/after pairs: CPython 3.11.15 174.19 -> 149.51 s, 3.12.13 107.41 -> 84.40 s, 3.14.4 67.72 -> 43.90 s; stdout byte-identical on all three, and the registered replay passes on 3.12.
• Local reviewer checks pass: check_research_commit.py (PASS_RESEARCH_COMMIT_BYTES, 16 files at this head), research_package.py --check-only, and a clean git diff HEAD~1 --check.
• Convention check for the likely question: direct patches to an existing package verifier with an identical receipt are merged precedent (Tooling: 2.26× faster production emitted-local audit via integer rows with identical receipts #2357 to research/complex-center-quarter/verify.py with re-pinning, Tooling: 2.49× faster production HPE audit via integer rows with identical receipts #2409 to vendored audit code, Tooling: 2.87× faster ternary verify_invocation_frames with identical receipts #2100 to scripts/experiments/ternary_side.py). This package does not self-pin its own verify.py, so no manifest update is needed; I can repackage the measurement as a research/*-fast-* package instead if preferred.
• Hosted CI: the Registered research packages run (38939671806) is still queued; I will address anything it reports. The remaining profile items (json.raw_decode 10.8 s / 1,968 calls; vendored recount_provider 40.7 s) are left as a handoff for the author.