Certify the P104 matching-pool limit and a 45-missing-pair witness

[petalas]

Summary

Adds only research/p104-rematched-dependency-screen/.

PR2981's fixed covers: 2,309 compatible pairs
  every pair has a checked perfect-matching witness
  published pairing: 3,095 missing required dependencies
  explicit alternative: 45 missing, all at Y40
  pooled union: zero missing

Pooling matching-supported pairs loses the obstruction. Deleting edges absent from every perfect matching deletes none, so that pooled support test cannot exclude all matchings of these covers. A stronger test must preserve matching coherence or other execution constraints.

The named alternative is a calibration, not a feasible correction or an optimal matching. It has all source supports at 103 targets; Y40 retains only X8/X41/X45. The search found no passing pairing and makes no exhaustive exclusion. No new supplier, lifetime floor or kappa is claimed.

Evidence

• Before: PR2996 at 4d772802fe86e5185d5468db86b36472e6c81b96 excludes completions of one fixed endpoint pairing and leaves rematching open.
  After: explicit alternating cycles certify all 2,309 edges as individually matching-supported. Interval contacts and literal quotient-frame enumeration regenerate the entire relation for the original, alternative and pooled cases, including the 3,095-pair control. The engines share parsing, matching verification, correction arithmetic and propagation; this is not two independent complete replays.
• Registered CPython 3.12.15 replay passes on Linux/WSL2, including eight behavioral groups, small exhaustive controls, Git-free export and rejecting source/witness/receipt/oracle mutations. Optimized 3.13.5 checks, Ruff, warnings-as-errors source compilation and publication-byte preflight pass. All twelve package/input files match commit 504dc9a15f7a874a47564cdf3ae6857bbbc8d000.
• Reproduce: python3.12 -B scripts/research_package.py --package research/p104-rematched-dependency-screen.

Chronology, jointly consistent full flags, scalar coefficients, dirty restoration, native lowering, physical hosting, precision, the complete consumer bill and all-size admission remain unproved. No selected cold replay, hosted-CI conclusion, independent human review or formal-kernel proof is claimed.

Started from canonical main 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11, priority revision 6. Final main and predecessor heads are unchanged. Refreshed overlap covers 253 open PRs, 2,440 closed/merged PRs and 1,724 Discussion records with complete pagination and no warnings; it does not cover arbitrary diffs or unpublished work. Scope change, bounded stopping criteria and the next matching-coherent question are in TASK.json.

Credit Nick Petalas/PR2981/2920/2996 and Arham Tahir's complementary PR2997 exact-flag audit, with the earlier information-flow, geometry, channel and native/compiler lineage retained in the README. Canonical investigation and handoff.

Merge danger

Door: Two-way. One additive opt-in package.

Blast radius: Package. Frozen predecessors, shared checkers, priorities and the selected frontier are unchanged. The pooled positive result does not justify construction expansion.