RT 2026-08-05 - area 1 - opus (Opus 5) - yield 8

[ms609-agent]

opus (Opus 5) | effort: unrecorded | 2026-08-05

Migrated record. This round predates the move to Discussions; it is reproduced
verbatim below from dev/red-team/log.md, which is now closed to new entries. It is
posted here so this area has a round record whose createdAt orders correctly against
the other areas' — the backfill ran oldest-first for that reason. All fourteen posting
timestamps are later than the rounds they describe; the real date is in the title and
in the date: field.

---

area: 1 (Fitch scoring correctness) — version-bump revisit
reviewed_by: opus finder a681ba267668c44d8 + haiku verifier a269ed170472c58bf (RT1-A1-01..08 batch) + sonnet verifier a1355d889b272f745 (RT1-A1-08 re-verify, overturning a flawed haiku refutation)
date: 2026-08-05
tier: opus (Opus 5)
yield: 8 filed (#76-#83: sev:med #77 (A1-02), #83 (A1-08, upgraded from finder's low); sev:low #76,#78,#79,#80,#81,#82) + 1 doc fix + 1 new regression test (65 assertions) applied inline; 0 refuted
notes: Rotation (13 mod 13)+1 = 1. Legend reconciled at round start: no rung has moved since 2026-07-27, so the routing was purely the standing version-bump-revisit record for this area (dry at opus-4.8, 2026-07-24; re-targeted to opus-5 with a fresh angle, NOT fable — version bump precedes rung bump). Brief explicitly chartered the ONE residual the 2026-07-24 round's relative TS_L3B_ORACLE-style checks structurally cannot see: a bug shared by BOTH the incremental and full-recompute paths, invisible to any check that only compares the two against each other. Pasted the prior round's full derivation as claims to break, per the version-bump-revisit protocol.

THE CHARTERED QUESTION WAS ANSWERED, AND THE ANSWER IS NEGATIVE — recorded as a genuine result, not an absence of one. The finder built two independent reference implementations sharing no code with src/ts_fitch* — a from-scratch Sankoff DP for EW, and a from-scratch brute-force applicability-enumeration + region DP for BGS inapplicable scoring — and cross-checked the absolute output of the real score_tree() against them across ~3000 datasets (multi-state, ambiguous, {-,X}/{-}/?-dense inapplicable, degenerate, deep-pectinate and shallow-balanced). 0 mismatches wherever the criterion is tie-break-free; the one-sided never-below-minimum bound held 1000/1000 including 500 {-,X}-dense cases; rooting-invariance held on every edge across 150+1500+576 test configurations; block-boundary additivity held across nChar ∈ {1,5,63,64,65,70,130}; IW/profile C++-vs-R agreement held 0/900. The shared-machinery correctness question for area 1 is now closed at this tier — a future revisit needs either a genuinely new angle (the finder's suggestion: cross-check vroot_cache/below_actives_cache candidate-scoring values against an independent rescore of the applied move, rather than the reconstruction layer already covered here) or an escalation to fable if a new delta lands that this round's coverage doesn't reach.

Incidental product of the chartered work: the BGS criterion this engine implements is now pinned down in writing for the first time in this repo. min over applicability reconstructions that are Fitch-optimal for the binary applicability character, of sum(within-region Fitch length) + (tip-bearing regions - 1), applicable-preferred tie-break on {-,X} tips. Two consequences that will look like bugs to a future reviewer and are not, now both encoded as assertions in the new tests/testthat/test-ts-na-oracle.R: (1) the engine can exceed the unconstrained minimum-cost reconstruction (17/500 definite-applicability cases, 30/500 with {-,X}) because Fitch minimises applicability changes, not regions — exactly what Goloboff et al. 2021 critique and what inapplicable = "xform" exists to address; (2) two applicable tips with the same state separated by inapplicable tips score 1, not 0.

Eight findings filed, none touching the shared-scoring correctness question itself — all either perf, latent-unreachable, or a documentation/invariant gap. RT-A1-02 (#77, med): the EW dirty-buffer allocation was static thread_local-optimised, then reverted by the emutls/MinGW fix (d6fa51293) for correctness, and the cost was never recovered on the NA hot path (exact_verify_sweep, ~97.7% of native-NA wall per this project's Mission-B profiling) — same class exists in 4 sites across ts_fitch.cpp/ts_fitch_na_dirty.h/ts_fitch_na_incr.h. RT-A1-08 (#83, upgraded to med): TreeLength() defaults to extended IW (XPIWE) but PrepareData()/EdgeListScore() — the documented default TreeScorer for TreeSearch()/Ratchet()/Jackknife() — always scores plain IW with no way to opt into XPIWE, so the two nominally-identical "score this tree at concavity k" paths silently disagree by default (600/600 measured mismatches, resolved by explicit extended_iw=FALSE). RT-A1-01 (#76, low): a detached-tip OOB guard present in fitch_na_score (added after a real ASAN abort) is missing from both incremental NA sibling loops — latent, currently unreachable. RT-A1-03/04/05/06/07 (#78-#82, low): residual per-call allocations in two "allocation-free"-labelled functions; a negative-delta path in precompute_profile_delta whose bounded-scorer-bail soundness rests on an unstated, unasserted invariant; an unproven (7500+ fuzz cases, 0 counterexamples, no proof) equivalence between two different ss_app derivations for EW vs IW/profile NA scoring; a diagnostic-only export whose own internal sum doesn't reconcile because it omits precomputed_steps; and two harmless dead/unreachable code fragments.

ONE HAIKU REFUTATION OVERTURNED ON RE-VERIFICATION — record the pattern, not just the outcome. RT-A1-08 was initially routed to the haiku batch (all 8 candidates were low/med severity, per this skill's severity-matched-verification rule — no sev:high this round to trigger automatic peer-tier routing). The haiku verifier refuted it by arguing extended_iw is a legitimate, documented feature — true, and not the claim. The actual claim was a default mismatch between two commonly-used paths for the same nominal computation. Per [[redteam-reverify-flawed-refutes]] and this project's standing note to route library-fact verdicts above haiku, re-verified with a sonnet pass, which traced the full default chain (R/tree_length.R → R/PrepareData.R → src/ts_rcpp.cpp's xpiwe=false C++ default) and confirmed REAL, overturning the refutation. Generalisable point for this severity-routing rule: "not sev:high" does not mean "safe to haiku" when the claim is a library/default-semantics fact rather than a straightforward code-pattern check — this round's median finding (allocation exists, guard missing) was well within haiku's competence, but the one finding requiring cross-file default-chain reasoning was not, independent of its filed severity.

Fixed inline, verified before committing. Corrected a stale comment in fitch_incremental_uppass (src/ts_fitch.cpp:244-251) that still described the static thread_local optimisation as current; it was reverted by d6fa51293 and the comment was never updated (this round's finder caught it while investigating RT-A1-02). Added tests/testthat/test-ts-na-oracle.R (new, 65 assertions, 2.3s), the independent-reference NA oracle the chartered work produced as a byproduct — built via the tarball recipe into .agent-rt1, ran with NOT_CRAN=true, confirmed 65/65 pass before committing; build artifacts (.agent-rt1/, src/*.o, src/*.dll) cleaned up after.

Seam status: shared-correctness question closed at this tier; perf/hygiene seam still yielding (8 filed, 0 refuted) → a future area-1 visit should target the candidate-scoring layer (vroot_cache/below_actives_cache vs an independent rescore of the applied move) per the finder's own suggested next angle, not another reconstruction-layer sweep.

---