[gmh5225]

Distinct STL and compiler-generated routines can share their complete byte patterns. Expand the independent source counterexamples and constrain existing ambiguous cohorts using complete reference requirements or incoming context queries.

The capture receipt records 43 configurations, 107 reproduced objects and their assembly/header evidence; the selector audits 228 unpruned source collections across x86, x64, ARM and ARM64. Historical provenance remains intact. In the changed packs, global scan records decrease from 1,425,169 to 1,423,736; additional source declarations use the name index.

Validation with NeverD #1861:

• All 60 linked probes: wrong preferred candidates 333 → 0, with no newly wrong candidate. Correct library candidates 171,702 → 171,309. Across library and user routines, 750 correct suggestions are withdrawn and 316 gained. Strict abstentions are reported separately from coverage.
• The original gap set retains 158/161 correct candidates. Three identities remain ambiguous under identical instruction/reference evidence; no aliases or independent identities are invented.
• All 98 changed pattern files passed native parsing; 318 native signature tests and 150 consumer/selector Python tests passed. Data tools: 240 passed, one optional external-format roundtrip unavailable.
• Twelve performance probes, 1/4 workers: paired warm matching CPU changes range from −1.07 to +3.21 ms. Ordinary release samples add 0.23–0.72 ms. Fresh-process paired wall changes range from −0.04 to +0.20 s, with peak RSS increases of 2.89–13.14 MiB. OS/signature caches remain enabled; host contention and same-engine controls are recorded.

Complete source hashes, accuracy counts, losses/gains, native output hashes and performance controls are in validation/library-features/expanded-collision-{capture-receipt,refinement}.json. These bounded checks do not establish exhaustive recognition or competitor superiority.