Audit PR1293 logarithm enclosures with exact counterexamples and fail-closed repair

[TennisBowling]

Result

A source-pinned enclosure-soundness audit of #1293, not a new supplier or frontier claim. Adds only research/pr1293-log-enclosure-audit/; all existing files and selected artifacts are unchanged.

At PR1293 head 85a361233431befe4b11cf878f7b1b77a244d473, default ln_exp_bisection treats an unresolved Taylor comparison as equality and returns a false singleton. This happens without reducing precision: CROSS_ORDER=48, CROSS_BITS=72.

• Integer witness: input 1681 returns [q,q] with q=4384212064881345387879/590295810358705651712. Exact finite rational arithmetic gives T_49(q)>1681, hence log(1681)<q.
• Transparent analytic witness: x=T_48(1/2) returns [1/2,1/2], although the first omitted positive Taylor term proves log(x)<1/2. This fixture also passes the surrounding sqrt-log agreement predicate; the integer fixture would be rejected by that predicate.
• Proposed minimal repair: raise an explicit unresolved ValueError rather than invent equality. Returning the previous wide bracket would be a weaker integration fix because the consumer checks only the other log method's width. The included repair.patch targets the pending PR1293 source, not main.

The package includes eleven exact fixtures, a code-disjoint rational atanh-series oracle, source/identity/input pins, meaningful semantic mutants, the unchanged vendored subject and deterministic receipts. README.md contains the proof and review boundaries.

Verification and impact

All 25 focused test groups and nine full-replay/receipt groups pass normally and optimized. Twenty repository research-helper tests also passed. The offline outer patch was applied to a fresh pinned-main worktree and all registered commands passed there. Before publication, the same registered replay was rerun; all checked implementation and evidence hashes remain unchanged.

A separately implemented full-main runner observes 41 calls, 19 distinct arguments, 2,952 steps and zero ambiguities in the actual predecessor program. All 12 predecessor controls pass. Original and strict-repair invocation/branch traces, stdout and generated RESULT bytes are identical; RESULT SHA256 is 731f77e404c18e3c1329d1ad084b3e0e3752579ae83fcb4e48a2b749d405c7a1.

python3 -B scripts/research_package.py --package research/pr1293-log-enclosure-audit

Python 3.12, standard library only; the generic provenance hook uses base Git history. VALIDATION.json records exact commands, hashes, environment and unrun obligations. The direct package scripts also run from source exports with their declared inputs.

Boundary: all 16 selected rank-log arguments avoid the defective branch. This does not refute selected moment signs, D/E, the selected conditional kappa or the multiplication theorem. No supplier-word rebuild, selected cold replay, Lean build, adaptive precision, expanded input domain or human-review claim is made. No selected-frontier change, merge or closure is requested.

Coordination and attribution

Base/current main: bce54fb3f895cbe5cd4c0eb477c3afeed100190f. The revision-2 independent-moment card was refined to this exact-enclosure audit because PR1262/1307/1329 already cover nearby ledger reconstruction. The publication refresh found main and PR1293 unchanged, and the defect still present. Its bounded search and limitations are in TASK.json; no exhaustive novelty claim is made.

Scoped Discussion finding. Initially prepared offline, then published with the donor's authorization as tennisbowling.

Credit Navinor/opencode for PR1293, woklerx's #1182 (0213b8911bfeefd5c81ddf320bac6a0fb7e86215) for the prior false-bracket/tri-state audit of a different pricer, and all retained supplier authors. The general fail-closed principle is not claimed as new; nearby #1392 concerns a different PR1048 comparator. The work and isolated mathematical/replay/adversarial reviews are entirely AI-generated by an OpenAI assistant and subagents, not independent human review. Apache-2.0; predecessor attribution retained.

[Mister-Meeseeks]

100-PR batch review at 917199f88adbc2cb9e402a150acb71f4ebf63d5f.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Exact false-singleton counterexamples to the pinned PR1293 supplementary logarithm cross-check, independent rational oracle and fail-closed patch. Both complete predecessor programs replay with 41 identical calls and zero ambiguous selected comparisons; no selected supplier failure is inferred.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.