fix(#80): A1 PUBLICATION_GRADE requires a positive symmetrised-gap certificate, not F1-F5 threshold exhaustiveness

[marcohost33-maker]

Closes #80.

Summary

Dedicated classification-semantics PR (AGENTS.md §3); the twin of #89/#88 and the implementation of #80 to-do 2 (reverse the burden of proof). The A1 early branch awarded CONFIRMED / PUBLICATION_GRADE (0.95) whenever gap_rate_consistency < 0.05 + single-exp held and none of the F1–F5 thresholds fired — i.e. the publication-grade claim rested on the unprovable exhaustiveness of the threshold set (kreiss>5, petermann>5, henrici>1, trans>5, reach>2·ratio, gns>1.2). Both independent #79 reviewers (Equalita facts-PASS; cross-family gpt-5.5 ADEQUATE) flagged exactly this residual: a hypothetical weakly-non-normal gap failure below all thresholds with a single-exp-at-gap trajectory would self-certify at publication grade.

Physics rationale

A measured symmetrised gap equal to the spectral gap is operator-intrinsic positive evidence of gap control: it certifies exponential contraction at rate ~Δ in the GNS (Mori-Shirai) or KMS (Fagnola s=1/2) geometry. Any genuine symmetrised-gap-failure mechanism — including a weakly non-normal one below all F1–F5 thresholds — must show up as a reduction (or an uncertifiable floor) in these measured quantities. So:

• New evidence key sym_gap_corroborated = 1.0 iff gap > 0 and (certified GNS with gap_to_gns_ratio ≤ 1.2 or gap_to_kms_ratio ≤ 1.2). The 1.2 boundary is deliberately the F3 firing threshold — "corroborated" and "F3-grade reduction" are complementary, no gap in between. Floored/uncertified GNS (fix(#88): classifier gates A2/F3 on a certified GNS gap — floor sentinel no longer self-certifies PUBLICATION_GRADE #89) and floored KMS both fail, fail-closed.
• _confidence: A1 scores 0.95 only with the certificate; uncorroborated A1 caps at 0.70 → CANDIDATE / CONFIRMATION — the honest grade: the single-exp-at-gap observable is measured, operator-intrinsic gap control is not.
• gap_to_kms_ratio graduates from advisory (fix(#88): classifier gates A2/F3 on a certified GNS gap — floor sentinel no longer self-certifies PUBLICATION_GRADE #89) to class-influencing — exactly the "dedicated PR with FP coverage" the advisory contract demanded. Using it as an F3 veto (issue Classifier F3 (A2/CONFIRMED/PUBLICATION_GRADE) fires off the gns_gap conservative-floor sentinel, not a measured symmetrised-gap reduction #88 option 2) remains deferred.

Behaviour change

Input	Before	After
A1 + certified/equal symmetrised gaps (thermal reference: Δ_GNS = Δ_KMS = Δ)	0.95 CONFIRMED / PUBLICATION_GRADE	unchanged
A1 + GNS floored, KMS == Δ (#88 repro regime, if it lands A1)	0.95 CONFIRMED	unchanged (KMS certificate suffices)
A1 + GNS floored/uncertified and KMS reduced or floored (the #80 adversary)	0.95 CONFIRMED / PUBLICATION_GRADE	0.70 CANDIDATE / CONFIRMATION

Anchor-preserving: the gap-controlled reference keeps A1/none CONFIRMED (its symmetrised gaps equal Δ exactly — verified numerically); V1–V5 golden labels are A5/A11 and never take the A1 confidence path; sacred anchor suite 21/21 green. No known real system flips — the capped region is precisely the adversarial class both reviewers hypothesised.

Remaining from #80 (not in this PR)

• To-do 1/3 (cross-family review of threshold exhaustiveness; numerical adversarial search for a real sub-threshold system) — the synthetic adversary is pinned in tests; a systematic operator search is research-grade follow-up. With this PR, such a system can no longer self-certify at publication grade, which defuses the severity.

Test plan

• pytest -q — 489 passed (exit 0)
• pytest tests/test_anchors.py -v — 21/21 incl. QuTiP cross-checks (exit 0)
• ruff check src tests benchmarks — clean (exit 0)
• mypy src/liouscope — clean, 50 files (exit 0)
• New tests: synthetic A1/PUBLICATION_GRADE depends on F1-F5 threshold exhaustiveness (residual from #69/#79 dual-review) #80 adversary → 0.70 CANDIDATE; KMS-certificate-alone → 0.95; both-gaps-floored fail-closed → 0.70
• Existing A1 branch/e2e tests unchanged-green (corroborated defaults / thermal reference)
• CHANGELOG.md updated (Unreleased → Fixed)
• Required status checks (Python 3.10–3.14 matrix + QuTiP 3.11/3.12)

Merge authorized by User1/Marco (Session 2026-07-07, „abarbeite weitere issues und merge") on green required checks; post-merge cross-family review of the semantics stays open in the #80 thread (claim_status: pending, KANON anchor RESEARCH-LIOUSCOPE). No manifest-schema, taxonomy-version or CITATION.cff change (same footprint as #77/#85/#89). No history-touching git operations.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BuG9j9cRJ6x6QMuYk7vXdx

---

Generated by Claude Code