RT 2026-08-05 - area 15 - opus (Opus 5) - yield 26

[ms609-agent]

opus (Opus 5) | effort: default | 2026-08-05 | second pass — opus arm of the tier-economics experiment

Yield: 26 candidates, 26 confirmed, 0 refuted. Four sev:high.
Filed as 7 grouped issues: #136, #137, #138, #139 (sev:high), #143, #144 (sev:med), #131 (sev:low). One defect fixed inline.

---

THE TIER EXPERIMENT IS ANSWERED — DO NOT RE-RUN IT

This arm ran over the identical scope with the sonnet arm's entire yield in its do-not-re-investigate block, and still returned 26 candidates including 4 sev:high against sonnet's 5 and zero.

	sonnet	opus
candidates	5	26
sev:high	0	4
finder tokens	137k	180k

The cheap pass removed no work from the expensive one — opus still read all 2,185 loc — so C_sonnet + C_opus > C_opus, exactly as the cost argument predicts. The cheap rung was not even much cheaper on this scope.

The severity split is the real result: sonnet found broken documented contracts; opus found silent wrong answers. Different class, not different amount.

Standing conclusion: start_tier: sonnet on a never-visited area is not a saving. The genuinely cheap lever this round exposed is verifier routing, not finder tier — the haiku batch cost 41k against the finder's 180k and confirmed 13 rows.

(Caveat added after the fable round: the sets are not nested in either direction, so "higher tier finds a superset" is false as stated. The missing control — a second opus pass with no exclusions — is #148.)

The four sev:high, all silent-failure class

#136 — Ratchet(stopAtScore=) returns the input tree carrying the improved score. edgeList <- candidate sits at R/Ratchet.R:190, below the break at :178. Measured on Lobo: returned tree RF-0 from the input, attr(,"score") 223, independent TreeLength() 230. Two more symptoms share the cause — the target-meeting tree is never added to forest so the success path errors "No trees!?", and the already-met early return omits the score attribute and kills MultiRatchet().

#137 — Ratchet() never forwards TreeScorer to its Bootstrapper, and ... cannot rescue it because TreeScorer is a named formal (R/Ratchet.R:82), so argument matching consumes it. Instrumented run: the custom scorer called 0 times during both bootstrap phases against 391 overall. The perturbation and search phases optimise different objectives — in the one function whose reason to exist is a custom criterion. Jackknife() forwards it correctly.

#138 — TreeSearch()'s default RootedTBRSwap holds the root fixed. The root split is a hard invariant (one distinct root bipartition over 20,000 calls), so a 150k-step walk reached 45 of 945 7-tip topologies: a mid-rooted start is confined to ~4.8% of tree space, permanently. EdgeListScore verified root-invariant (230 across rootings), so the loss is pure.

#139 — fractional weights truncate to zero before bootstrap resampling. original_weight is filled by as.integer() truncation (R/PrepareData.R:76) while scoring uses .ScaleWeight(). Uniform fractional weights floor to all-zero, every character resamples to weight 0, every tree ties at 0, and R/CustomSearch.R:74 accepts equal scores — so BootstrapTree()/JackknifeTree() degenerate into a random walk, with zero errors, warnings or messages (confirmed via withCallingHandlers).

VERIFICATION CORRECTED THE FINDER THREE TIMES

Do not treat a finder's numbers as the record, even at a 100% confirm rate.

1. One sev:high downgraded to med — a loud immediate error, not the silent one claimed.
2. A15-06: TreeSearch()'s default swapper holds the root fixed, confining a mid-rooted start to ~5% of tree space (needs a design decision) #138's mechanism is worse but its motivating example was wrong. The finder wrote "NJTree()/rtree()/read.tree() all yield mid-rooted trees"; NJTree() is tip-rooted (measured 1|47 on Lobo), so the package's own ?TreeSearch example is in the benign 78–81% regime. ape::rtree(30) (24|6), a balanced Newick (4|4) and BalancedTree(20) (10|10) are the genuinely confined cases.
3. Three findings came back WORSE. The NNISwap self-hit rate is 2/(nTips−2) not 1/(nTips−2) (40% at 7 tips, not 19.8%) because a root with two internal children makes both root-adjacent edges self-hitting — a fixer following the finder's number would have excluded one edge and left half the bug live. NNI(-1) is missing one internal edge's entire pair of true neighbours (6 of 8 reachable), not merely padded with duplicates. And TBRSwap() on a trifurcating root does not misbehave — it HANGS INDEFINITELY (killed after 2+ min), which is why A15-09: exported swappers' argument validation is unreachable or absent — TBRSwap() HANGS on a trifurcating root #144 is med not low.

Verifier methodology worth reusing, not re-deriving

Both peer verifiers built independent unrooted-neighbourhood enumerators and validated them against the closed forms 2(n−3) for NNI and 2(n−3)(2n−7) for SPR before using them as ground truth (n = 5–10, several shapes). That validation is what licensed them to contradict the finder rather than defer. Rebuild trap: tied n/2 | n/2 splits need lexicographic, not size-based, tie-breaking.

Fixed inline, not filed

R/SuccessiveApproximations.R:111-119 — the "Stability not reached" branch emitted its message regardless of verbosity while "converged" was gated, so a default verbosity = 0 call was noisy. Both now inside one gate. An identical ungated message() survives at R/Ratchet.R:291 in MultiRatchet(), left for #131's sweep.

Ruled out — do not redo

morphy-deprecated.R is clean: all eight shims delegate to native replacements, none reaches a removed MorphyLib path, none loads a DLL, UnloadMorphy() returns invisible(0L), .GapHandler()'s pmatch accepts legacy "inapp". Only cosmetic residue (nTaxa/gap accepted and not forwarded). Also: the 1:n idiom survives nowhere else in scope; RearrangeEdges()'s double TreeScorer() call is correctness-safe for deterministic scorers and a hazard only for stochastic ones; Ratchet()'s accept criterion is monotone on the normal exit path; .NonDuplicateRoot() (R/SPR.R:42-54) is correct including its asymmetric branching — do not "simplify" it; TBRSwap() has full neighbourhood coverage (64/64, 106/106).

Not chased

TBRSwap()'s residual 1.7–2.65% identity rate — mechanism unidentified, not the brokenRootDaughters bug and not a warning path. Whether SPRSwap() can reach nCandidates == 0 and die on parent[[integer(0)]] (R/SPR.R:231-248). PrepareDataSA() reachable only through the broken SuccessiveWeights(), so entirely untested. NNISwap()/RootedNNISwap() derive rootNode <- nTips + 1L rather than reading it from the edge list.

Seam status: still yielding, abundantly. Highest-value next work-shape is not another general finder: audit which advertised custom-criterion entry points work end to end (#137 + #126 + dead SuccessiveWeights() are one decayed-façade story), and port the validated neighbourhood enumerator to the C++ all_spr/all_tbr paths.