RT 2026-07-24 - area 3 - opus (Opus 4.8) - yield 1

[ms609-agent]

opus (Opus 4.8) | effort: unrecorded | 2026-07-24

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.

⚠ Scoped round. This was one of the 2026-07-24 directed passes over what was new in
MaximizeParsimony since the previous inspection — not the area's full file list. The
seam verdict it records is about that surface. Read the verdict in the notes rather
than inferring one from the yield in the title: a low number here can mean the new
code was clean, which says nothing about the parts of the area it never opened.

---

area: 3 (Ratchet & perturbation — DIRECTED round: "what's new in MaximizeParsimony since last inspection")
reviewed_by: opus finder (ace557dc) + haiku verifier (a057639d) + orchestrator (advisor-briefed scoping); pinned @ ec8edd5
date: 2026-07-24
tier: opus (Opus 4.8)
yield: 1 (T-335 P3 latent tripwire; 1 candidate RT-A3-02 REFUTED as inverted)
notes: First area-3 review since 2026-05-26 (yield 0). The prior 3 dry opus-depth rounds (RT12-02 flagged
them) ALL predate this new surface → fresh-surface opus pass, not a dry re-mine. New commits reviewed:
sector in-sector drifting (1333c9e2, opt-in), RSS size-scaling (1dcd5b1e, default-OFF), per-sector
column-axis reduction (830b8cc3 TS_SECT_COLREDUCE, default-OFF), prune-reinsert exact directional scorer
(4007f8e8, opt-in), fuse per-round reroot + replace_subtree guard (ac8e808a, default-ON). All 4 in-scope
files verified byte-identical to the pinned tip ec8edd50 (concurrent session actively editing
ts_sector/ts_fuse — cited lines re-checked against the pinned hash to avoid phantom findings). Advisor
reframed the round: the T-330 HSJ/XFORM analogy is a severity TRAP, not a multiplier — T-330 was P1 only
because it is default-ON TERMINAL output with no downstream recompute; the sector family is SEARCH-INTERNAL
heuristic + mostly opt-in. CORE DELIVERABLE — Target #1 full-dataset-rescore firewall: YES. Every
splice-back accept is gated on a rescore of the FULL tree against the COMPLETE dataset: rss `reinsert_sector

• build_postorder + score_tree(tree, ds)(ts_sector.cpp:1734-1736, accept/revert :1752-1772), xss identical (:1941-1977), css never builds a reduced dataset (runs tbr_search against fullds, :2028-2099), prune-reinsert whole-tree backup + full rescore gate + full restore on reject (ts_prune_reinsert.cpp:601/656/662), sector-gocomb + top-level pool fuse both rescore via score_treebefore pooling (ts_sector.cpp:1351; ts_driven.cpp:1058,1135). **HSJ/XFORM sectors are SKIPPED entirely** (rss:1522-1525, xss:1888-1891) — the old T-303 mitigation ("guard rss/xss/css for HSJ/XFORM") is now IMPLEMENTED, so the HSJ/XFORM-blind reduced-dataset field-copy is moot for those modes. **Consequence: the HSJ/XFORM-blind reduced dataset AND the new TS_SECT_COLREDUCE are HEURISTIC-QUALITY only (worst case a mis-ranked/missed candidate = reach, P3), never a wrong returned score.** The one memo that could have leaked reach across the reduced↔full boundary — theexact_verifyNA-optimum cache (advisor-directed check) — was TRACED to a confirmed NON-bug:evs_false_cache/evs_last_fpare per-DataSet-INSTANCE members (ts_data.h:211-212), reduced sectors use a fresh isolatedrd.data(never copiesevs_false_cache), and the key tree_topo_hash ^ ds_fingerprint ^
  weight_fingerprint(ts_tbr.cpp:864) distinguishes both reduced-vs-full and base-vs-perturbed regime axes (the 78b74147 fix, present at tip). **T-335 (P3, verified REAL by haiku) FILED:**reduce_sector_columns_ew(ts_sector.cpp:407-411) rewritespattern_indexto identity but leavesn_patterns/min_steps/pattern_freq/ precomputed_stepsat full size/content — INERT today (hard EW-only gate at :315-317:scoring_mode==EW &&
  all_weight_one && no upweight/inapplicable; the invoked EW scorers read only block structure, never those 4 arrays) but a live wrong-score + OOB tripwire if the gate is loosened toward IW. Filed defensively because the concurrent mission-b work is actively on ts_sector. **RT-A3-02 REFUTED (not filed):** the finder INVERTED the divergent-field claim — BOTH sector build_reduced_datasetfns DO copyprecomputed_steps(ts_sector.cpp:605,810); the true (opposite) divergence is prune-reinsert OMITTINGinfo_amounts/ info_max_steps(ts_prune_reinsert.cpp:298-321), which is PROFILE-guarded (prune-reinsert skips PROFILE/HSJ/ XFORM) and harmless. Carried as a corrected note here, NOT a findings row (matches the 2026-05-26 precedent of logging-not-filing the divergent-fns footgun). **Targets #2-#4 CLEAN:** (#2 composition) each sector pick builds a freshReducedDataset rddestroyed at pick-end; drift/TBR/fuse operate onrd.subtree/rd.data; the sector-level revert is topology-only (save_left/right/parent+original_score) which SUFFICES because reinsert_sectorreads only topology and the full-tree layer is restored byrestore_clade+ fullscore_tree— drift's torn state arrays are never read post-reject → structurally immune to cross-pick contamination. (#3 fuse per-round reroot, DEFAULT-ON)reroot_at_tip0early-return invariant (ts_fuse.cpp:23) relies on incoming valid postorder, verified on all in-scope call paths; PRE-EXISTING invariant unchanged byac8e808(which only ADDS reroot calls — strictly safer). (#4 prune-reinsert exact scorer) full backup/rescore/restore, no T-235-class stale-array leak. **Ratchet:**ts_ratchet.cppdid NOT exist pre- 2026-05-26 (extracted by 78b74147 on 2026-06-19, behaviour-preserving) — NOT re-audited-old-code; masks/ pattern_freq restored + flat_blocks synced every cycle before the accept check.0fb6de9thread_local probe accumulators are behind#ifdef TS_SCOREAPPROX_PROBE(not compiled in production). **Default-state map (from code, not commit titles):** default-ON = between-replicatetree_fuse (fuse_interval=3); everything else (intra_fuse, sector drift/comb, collapse_target, TS_SECT_COLREDUCE, RSS size-scaling, prune-reinsert prune_reinsert_cycles=0) is opt-in / default-OFF. **Seam status: RAN DRY on the new surface (evidenced, not assumed) — i.e. dry at opus-4.8** → *(as recorded 2026-07-24: "per doctrine next area-3 visit escalates to FABLE")* — **RE-TARGETED 2026-07-27 by the opus version bump: the next area-3 visit is opus(Opus 5) with a fresh-angle brief, NOT fable** (version bump before rung bump). BUT the residual is THIN: the whole new surface is opt-in/heuristic behind the full-dataset firewall, so a correctness bug is structurally hard to hide here. **Recommend area 3 as a DOWNTIER/retire candidate at the next area-12 meta-review** — this is now the 4th dry round, and the firewall explains WHY (RT12-02 already flagged the 3-dry-round tension). **Caveat added 2026-07-27:** three of those four dry rounds are **pre-tier and version-unrecorded**, so only *one* (this one, opus-4.8) counts toward the two-consecutive-versions dormancy bar — a retire decision should wait on the Opus 5 re-visit. Residual if pursued: exact numeric correctness ofcompute_from_above_for_sector/build_ras_sector`
  heuristic proxies (firewall-protected → reach-only, not a returned-score risk). NEXT ROTATION AREA = 4
  (Parallelism & RNG).