test(research): audit centre-circle demands and late responses

[petalas]

Summary

Add research/centre-circle-endpoint-audit/, a bounded independent audit of PR3147 at 5cbc48b18cb05327a8dacab3f626412d3bc5e924.

Original inputs        F0 demand = R x, F2 demand = -R y
                       joint rank = 16, minor determinant = 1024
Late centre response   u -> (-u, +u, -u, +u) through three bridges
                       rank = 8 per replica, 128 across sixteen replicas

This supplies a target-only input witness and corrects the hosted schedule's reported absence of later bridges. Source-only cancellation and no later recursive calls do not justify direct centre elision under the unchanged returned-data contract. A different program or consumer remains open.

PR3147 already reports joint rank 16 and local response rank 8. Those observations retain its credit. The added result is the separately certified rational minors, explicit witness and literal 48-bridge response through the last bridges, before terminal consumption. Other PR3147 claims are not reviewed here.

No new kappa, native/compiler proof, supplier admission or selected-frontier change. All non-scalar recursive events are identity in this explicitly scoped scalar specialization. Canonical investigation: #3149.

Evidence

• Before: PR3147 prices reduced centre profiles using restricted cancellation/zero-feedback observations; its hosted stage-8 text reports no later bridges.
• After: the frozen hosted archive has three later bridges per replica. Exact replay preserves 128 independent response coordinates and gives T0[0]=1 as a zero-F0/nonzero-F2 witness.
• Registered CPython 3.12.15 replay passes:

  python3.12 -B scripts/research_package.py --package research/centre-circle-endpoint-audit

• A separate dense scalar interpreter and integer Bareiss determinants agree with the sparse rational replay. Eight behavioral groups pass normally and under optimized CPython 3.13.5, including semantic mutants, read-only Git-free export and source/receipt corruption controls.
• All 24 assignment/coordination tests, warnings-as-errors, strict applicable Ruff checks, formatting and exact publication-byte checks pass. Exact sources, receipt, commands and limits are committed. No independent human or formal-kernel review is claimed; hosted CI is separate.

Credit dcharb78/PR3147, Aaron Wang/PR3120/3075, the observation/return-boundary contributors and the retained Rohan Arun/Jeremy Kothe/Hao Yan selected-construction lineage. Full attribution is in the README. No predecessor bytes change.

Merge danger

Door: two-way. Blast radius: one isolated research package and its opt-in replay. Shared verification, historical evidence and the selected record are unchanged.