DONE: numeric task references in overlap search

[petalas]

DONE / RELEASED / 2026-10-16 / priority revision 6

PR3002, exact commit 0bc34c304441740273e30e06cb28710bdd3d788d, fixes numeric task references in the overlap helper. Evidence and reproduction.

Canonical positive integer card terms match identifiers rather than decimal/hash fragments. Explicit positional queries and leading-zero witness words remain literal. A same-corpus 2977 comparison retains six genuine leads and removes seventeen fragment hits. All 25 focused optimized tests pass; the broader 164-test workflow suite passes with eighteen existing skips. Separate read-only review found no blocking issue. Evidence is author-checked, with no independent human or scientific-theorem review claimed.

No active claim. Identifier matching remains syntactic, broad terms can still fill the result limit, and scientific/selected-frontier obligations are unchanged. Final refreshed coverage was 244 open PRs, 2,440 closed PRs and 1,707 discussion records, complete and without warnings; canonical main was unchanged.

---

Original claim follows, retained for scope history.

CLAIM / 2026-10-16 / priority revision 6

Question: can numeric references supplied by task cards avoid matching unrelated kappa digits and commit hashes while retaining real PR/Discussion references? The current literal matching of the supplier-boundary-audit term 2977 returns unrelated long numerical claims. This is a coordination-tool repair, not a mathematical claim.

Base main 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11. Original draw 4c5a1c825963fc47f3423c2a00a41d2b, adversarial audit. Scope changed from the supplier-boundary audit after fresh overlap identified Arham Tahir's PR2992 successor and a separately claimed census audit. The stock leads retain their existing implementation obligations; this claim does not take those over.

Method/domain: add identifier-aware matching only for numeric task-card overlap terms; preserve explicit literal queries, text phrases, package names, result scoring, freshness and path search. Public-CLI regressions must show a real successor below twenty spurious numeric title hits before the repair and visible after it, plus genuine-reference and literal-override controls.

Expected artifact: focused script/test/documentation PR. Exit predicate: reproducible failure, passing normal/optimized focused tests and relevant research-workflow tests, refreshed overlap and reviewed exact diff. No frozen scientific inputs or frontier changes.

Evidence LOCAL-ONLY until publication. Nick Petalas; substantial automated assistance in implementation and checking. Claim expires 2026-10-16 19:10 UTC, advisory.

[petalas]

FINDING / DONE / RELEASED

PR3002, head 0bc34c304441740273e30e06cb28710bdd3d788d, adds identifier-aware matching for canonical positive integer task-card terms. It preserves explicit literal overrides and leading-zero witness words such as 01021323, and uses the actual match for title scoring and body excerpts. Numeric matches no longer borrow their score or excerpt from a hash/decimal occurrence.

The public CLI regression originally returned 22 hits, with 21 irrelevant titles hiding the genuine successor below the twenty-result limit. The repaired fixture returns that successor alone. In the dated 244-open-PR/1706-discussion/3126-local corpus, numeric 2977 hits fell from 23 to six; seventeen decimal/hash fragments were removed. The evidence note distinguishes this dated cache observation from the network-free deterministic regression.

Verification: 25 focused optimized tests; 164 workflow tests with 18 existing environment/platform skips; Ruff correctness, whitespace, publication-byte and selected-source-binding checks. A separate read-only reviewer found no blocking issue and identified the leading-zero exception, now covered by a regression. No hosted-CI or scientific replay result is claimed.

Scope changed from the drawn supplier audit to this observed coordination defect after overlap identified Arham Tahir's PR2992 repair and an active independent census audit. Douglas Colkitt retains the helper/workflow credit. Substantial AI assistance was used in implementation, verification and writing; no independent human review is claimed. The selected frontier is unchanged.

Remaining limitation: syntactic identifier matching does not establish novelty, and broad terms may still fill the output limit. No further ranking change or research-family expansion is proposed without a concrete missed lead.