Unittest receipt comparison: exact stderr outside the final timing field

[Arham-Tahir64]

DONE — PR #1926, head 387501825842caa3507e34f110b5f9570dc04046. Claim released.

The integrated PR921 benchmark can normalize away substantive diagnostic changes: its unanchored timing regex erases in 1.000s versus in 2.000s in failure text, and UTF-8 replacement decoding collapses distinct ff/fe bytes. Both collisions reproduce with real unittest captures; bounded fault injection makes the frozen benchmark report PASS for changed sequential/concurrent stderr.

The source-bound successor package research/unittest-receipt-comparison retains all captured bytes except the numeric elapsed value in the final standard unittest trailer. Separate raw/summary tags and an explicit prefix length prevent marker collisions. An opt-in adapter preserves the predecessor's commands, test inventory, stdout and exit-code checks, and rejection of remaining mismatches; predecessor files stay untouched.

Verified locally on Windows 11 / CPython 3.12.10: all 20 focused tests pass in both normal and optimized registered replay, with genuine timing-only acceptance, exact bytes, real captures, end-to-end faults and source-drift controls. Metadata/pins, affected-package selection and patch hygiene pass. All 42 onboarding tests pass with PYTHONUTF8=1 (the separate cp1252 issue is PR1834); workspace TEMP/TMP supplied sandbox-compatible isolated fixtures.

Reproduce: python3.12 -B scripts/research_package.py --package research/unittest-receipt-comparison. TASK/SOURCES/RESULT/VALIDATION/VERIFY carry the precise scope, sources, witnesses and limitations. Full repository benchmarking and selected mathematical replay remain unrun; no speedup, new kappa, historical-run fault or full-suite equivalence is asserted. Hosted CI was queued at publication and is not counted as a local pass.

Base main 0b1cccf86e2f5132ed286a26ef0a681a495740ea; portable-replay nonce d478b42a7ca6903045649ed7bd64bc93, priority revision 3. Final fresh overlap covered 573 open PRs and 542 Discussion records, complete pagination and no warnings. Matches were this claim/package, its frozen predecessor and the unrelated data-corner benchmark. This is not proof of novelty. Related PR1852 protects child assertions; it addresses a distinct obligation.

Next question: run the exact adapter on the entire pinned suite in a supported environment and inspect any custom output that conservatively fails comparison. New trailer support must retain diagnostic byte equality.

Credit Aaron Wang/navinor PR921 (reviewed submitted head c1ee2fbfdc5700a1ba0251e7727ee8ae0cb4782f) and Douglas Colkitt's batch932 acceptance hardening. Prepared by Arham Tahir with substantial OpenAI Codex assistance; Apache-2.0. No subagents or independent human review.