RT 2026-08-05 - area 13 - opus (Opus 5) - yield 12

[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: 13 (Constrained search correctness) — opus finder sweep, directed round
reviewed_by: opus finder ac6ece01b95dec79c + opus verifier ad74e7246bc17d134 (A13-01/A13-02) + haiku verifier a67f772b67755c4a0 (A13-03..A13-12 batch)
date: 2026-08-05
tier: opus (Opus 5)
yield: 12 filed (#61-#72: A13-01/A13-02 sev:high, A13-03/A13-04/A13-05 sev:med, A13-06..A13-12 sev:low) + 2 doc fixes applied inline; 0 refuted — all 12 candidates confirmed REAL
notes: Dispatched directly at the user's request ("sounds like we're ready for a proper area 13 sweep now"), immediately after the same-day directed harness re-run (see the entry below) closed out area 13's standing "not a finder" blocker.

Two high-severity findings, both independently reproduced end-to-end. A13-01 (#61): .PrepareConstraint's joint-satisfiability gate (R/MaximizeParsimony.R:156-181, T-329's fix) is pairwise four-gamete only, which does not imply joint displayability for partial splits (taxa excluded from both groups) — unlike for complete splits, where pairwise ⇒ joint holds by Buneman. Verifier independently enumerated all 105 six-taxon trees and confirmed a 3-split counterexample ({a,b}|{c,d}, {a,b}|{c,e}, {a,d}|{b,e}) that is pairwise-compatible on every pair but displayed by 0 trees; an end-to-end run then accepted the constraint with no error/warning and returned trees that violate it (score 67 vs unconstrained 65). Aggravating factor found during verification: only consSplitMatrix crosses to C++, never consZero, so C++ actually enforces a different, stricter proposition than what the R gate validated — in the repro this makes the C++-level splits mutually incompatible, which is why the search freezes on top of the original silent-violation risk. A13-02 (#62): map_constraint_nodes (ts_constraint.cpp:207,:217) iterates tree.postorder, confirmed internal-nodes-only (ts_collapsed.h:23), so a single-tip 1-group constraint split can never map unless that tip happens to be a root child — constraint_node[s] stays -1 forever, freezing every phase touching that split. .PrepareConstraint doesn't filter size-1 rows. Verifier's control repro (25 taxa) isolated a clean 12-step cost for an otherwise-free vacuous single-tip constraint, and corrected the finder's "totally frozen" framing to "measurably degraded, not fully frozen" (1068 < 1090 Wagner baseline) — kept the finding, tightened the claim.

Ten more filed at med/low, all confirmed by the haiku batch verifier against current HEAD line numbers. A13-03 (#63, med): loose (violates_constraint_posthoc+has_posthoc) vs strict (constraint_node[s]<0+active) gates diverge across phases, full call-site inventory across Wagner/sector/TBR/drift/spr_search/prune-reinsert/nni_perturb/fuse/parallel-fuse — including a genuine serial-vs-parallel-fuse mismatch (parallel-fuse trigger is loose, its own verify half is strict). A13-04 (#64, med): consensusConstrain auto-constraints (has_posthoc=false by construction) are invisible to the sector phase's gate but the sector phase's post-accept resync fires regardless, freezing the rest of the replicate — opt-in only (consensus_constrain defaults false), distinct from the already-ruled-down T-13-B. A13-05 (#65, med): tree_fuse has no ConstraintData parameter at all — constraint-blind internally, makes fuse near-inert under a constraint (relies entirely on caller-side repair/discard). A13-06 (#66, low): anneal_search has no post-move constraint verification, unlike every sibling phase — mitigated because the pre-move screen is argued sound for its SPR-only move set and annealCycles defaults 0L, so filed as defence-in-depth not a proven hole. A13-07 (#67, low): impose_one_pass computes move_out_roots/move_in_roots once per pass but consumes them stale across re-anchored moves within that pass — silently ineffective repair, not corruption (T-333 still catches actual invalidity). A13-08 (#68, low): topology_spr's root-child no-op bail is still counted as a successful move by try_move, defeating impose_constraint's moves==0 convergence check. A13-09 (#69, low): drift's RFD re-apply failure path is the one exit skipping the constraint resync every sibling exit performs — assessed practically unreachable, filed for consistency. A13-10 (#70, low): spr_search leaves cd stale on return, safe today only because every downstream caller happens to re-init. A13-11 (#71, low): extract_consensus_splits dedups by 64-bit hash alone with no bitset comparison — theoretical collision risk feeding the same freeze mechanism as A13-04. A13-12 (#72, low): expand_and_reinsert takes a ConstraintData* cd parameter it never reads — dead/misleading, and a natural spot to close part of A13-03's gate inconsistency if enforcement is ever added here.

Q3 (false-negative direction of regraft_violates_constraint) answered as a resolved negative — do not re-hunt. Traced every ClipZone×constraint_complement combination including the below==cn boundary exception and the cn==parent(clip) splice-out case. The straddle→UNCONSTRAINED branch is sound for SPR (the clip's internal induced bipartitions are invariant under relocation) and is only unsound for TBR (rerooting moves the attachment leaf, changing which clip-internal edges induce which whole-tree bipartition) — which is exactly why the TBR (ts_tbr.cpp:2867) and drift (ts_drift.cpp:741) backstops exist, and both are correctly placed with correct resync on every reject path. The residual exposure from this line of attack is A13-06 (anneal_search's missing backstop), not a hole in the predicate itself. One over-rejection case noted but not filed separately (when the clip subtree is cn, every regraft is rejected even though it's harmless) — that's issue #54's territory, not a new bug.

Q4 (laminar/nested consistency across TBR/Wagner/sector) answered: the divergent axis is loose-vs-strict (A13-03), not nesting. Every path tests each constraint split independently and none assumes or mishandles laminarity — confirmed by direct code read, not just absence of a counterexample.

Ruled out this round, do not re-trace: constraint-metadata resync on TBR / prune-reinsert / ratchet / sector revert paths (all correct — A13-09/A13-10 are specifically about narrower exit paths, not a general resync failure); per-worker ConstraintData copies in parallel mode (T-336 still holds); split canonicalisation agreement between ts_splits.cpp::canonicalize_split and build_constraint post-T-384 (now largely cosmetic for mapping, polarity handled symmetrically); structurally_valid() (T-333) remains a complete check as documented.

Not yet examined, flagged as the obvious next seam: Resample.R/SuccessiveApproximations.R also call .PrepareConstraint (R/Resample.R:311, R/SuccessiveApproximations.R:71) and route to ts_resample_search/ts_parallel_resample/ts_successive_approx — not traced this round for whether they carry the fuse/sector/pool gates at all. Also latent, not yet a bug: impose_constraint's repair always targets making the canonical side the clade regardless of which side is cheaper to realise (ts_constraint.cpp:741-753), which combined with A13-07 probably explains part of the heuristic's weak convergence on nested/complex constraints.

Doc fixes applied inline (not filed, comment-only): src/ts_constraint.h — regraft_violates_constraint's doc no longer references a nonexistent above parameter and now states explicitly it's screening-only (every caller must re-verify the applied move); impose_constraint's doc replaced the false "after return, all constraint splits are displayed" contract with the actual heuristic contract (bail-out cap, per-move revert, return value doesn't distinguish repaired-from-gave-up).

Seam status: still yielding (12 filed from 12 candidates, 0 refuted) → next area-13 visit, once rotation reaches it, should stay opus with a fresh agent, targeting the Resample.R/SuccessiveApproximations.R seam named above first.

---