RT 2026-08-06 - area 15 - fable (Fable 5) - yield 31

[ms609-agent]

fable (Fable 5) | effort: default | 2026-08-06 | third pass — fable arm, run with no exclusions

Yield: 31 candidates; ~20 were independent rediscoveries of the opus round's findings (see below — that overlap is the point). Of the genuinely new ones, 12 confirmed and 2 refuted.

Filed: #147 (sev:med, cross-area — TBRMoves() incompleteness), #155 (sev:med — Ratchet(returnAll=TRUE) reports a wrong MPT set), #156 (sev:low — second sweep of guard, documentation and dead-code defects, including an opaque SPR() crash).

Verification corrected this round's finder three times, so read the issues rather than this post's summaries for the operative detail: the SPR() crash's stated mechanism was wrong (the proposed fix would have changed nothing) and its count was 6,255 not 138, but it is unreachable from any search so it dropped to low; Ratchet(returnAll=TRUE)'s start-tree defect has a silent mode the finder never reported, which matters more than the error it did report; and the rooted-dedup defect was upgraded low → med because it inflates a user-facing MPT count. Two candidates were refuted outright and are recorded in #156 so they are not re-hunted.

Finder cost: 162,265 tokens / 51 tool uses / ~28 min.

---

Why this arm had no exclusion block

The opus arm was told everything sonnet found, which measured what opus adds but destroyed the ability to see whether opus would have found sonnet's set anyway — the superset premise the whole ladder rests on. This arm was therefore told only that prior rounds existed, not what they found, and explicitly told not to read log.md, focus-areas.md or the tracker.

The result: the finder sets are NOT nested, in either direction

Fable independently re-found ~20 of opus's 26 — and both of sonnet's distinctive findings (RootedNNI(-1) dead, AllSPR's undefined tree), which opus had been told to skip. That is the superset evidence the opus arm could not produce.

But:

• Fable found ~10 that opus missed, including a reproducible crash: SPR()'s guard at R/SPR.R:203 is if (edgeToBreak != 1 && all(edgesCutAdrift[-1])), and the != 1 clause exempts precisely the case where the rest of the tree is adrift — 138 hits in an exhaustive n = 5–9 sweep. Also MultiRatchet(ratchHits=) silently inert; returnAll never seeding the forest with the start tree; unique() deduplicating rooted edge matrices so one unrooted topology is reported twice as two "unique MPTs".
• Opus found ~5 that fable missed, including TBRSwap()'s indefinite hang on a trifurcating root, the SPR neighbourhood-coverage gap, and the unreachable mergeEdge length guard.

So "a higher rung finds a superset" is false as stated. What the data shows is that any competent independent pass finds a partly-different set — which is the rotation argument, not the ladder argument.

Cost comparison (sonnet-equivalents, at 1× / 2× / 4× per-token price): sonnet 137k · opus 360k · fable 649k.

The missing control is #148: a second opus pass with no exclusions, matching this arm's conditions. Without it, fable's marginal contribution cannot be separated from "another fresh pass". Escalation to fable is retained as protocol by maintainer decision regardless.

Headline: a cross-area C++ defect — #147

TBRMoves() omits every TBR move that bisects the root edge, so its output is not a superset of SPRMoves(). Provable without any oracle, since TBR ⊇ SPR by definition: SPRMoves \ TBRMoves = 8/8/6 on Balanced(8)/Pectinate(8)/random, and 12–14 at n = 11.

Two independent defects in src/rearrange.cpp, both by contrast with all_spr in the same file: the default break sequence starts at edge 3 (:510-513) where all_spr starts at edge 2 (:366-369); and there is no break_edge == 1 branch at all, where all_spr has one at :435. Fixing the first alone does nothing.

Severity med, not high — the search engine is not affected. all_tbr is called only from R/TBR.R:90,101. MaximizeParsimony() drives src/ts_tbr.cpp, a separate implementation that handles the root edge explicitly (try_root_edge_moves, :709,3168). The reach mission is not implicated; exposure is confined to users calling TBRMoves() directly.

Two collateral facts: tests/testthat/test-rearrange.cpp.R:40 asserts expect_equal(58, length(TBRMoves(tr))) — the test currently locks the bug in, and the replacement number must be derived, not copied from the fixed output. And src/rearrange.cpp (22 KB) is in no scope row, so the rotation would never have reached it; it surfaced only because a finder used TBRMoves() as a cross-check. Third unowned file to produce a real finding (cf. #42).

Methodology warning

The obvious topology key does not work. sort(as.character(as.Splits(...))) is not injective — it collapses the 10,395 unrooted 8-leaf topologies to 8,626, and made the first oracle attempt return 65 instead of 106. Use a bitmask key and validate that it yields exactly NUnrooted(8) == 10395 distinct values before trusting any count derived from it.

Conflict to resolve before #143 is fixed

Fable measures SPRSwap no-ops at 0.1–3.6%; the opus round's verifier measured 9.6–12%. Both were careful. Most likely rooting — that verifier had already flagged TBRSwap's rate as rooting-sensitive. The rate frames the fix, so pin it first.

Ruled out this round — do not re-spend

• Topological validity of all six *Swap() samplers. Fuzzed n = 5–12 × 300 random trees, plus an exhaustive edgeToBreak sweep n = 5–9: zero invalid edge lists. The swap arithmetic is sound; every defect is in the selection or guard logic around it.
• TBRSwap() neighbourhood completeness — reaches exactly the full TBR-distance-1 neighbourhood (106/106 at n = 8, nothing outside it). The one sampler that is both complete and essentially no-op-free.
• Surviving MorphyLib routing: none.
• Tip-label correspondence in the search entry points. All funnel through RenumberTips(tree, .SearchTipLabels(dataset)). The by-position/by-label bug class does not appear here. One watch point: .SearchTipLabels() returns names(dataset), so RenumberTips(tree, NULL) is unguarded for a custom TreeScorer whose dataset is not a named list.
• Score/tree consistency on the ordinary exit paths — verified matching; only the stopAtScore paths are broken.

Where to look next

1. The EdgeListSearch ⇄ RearrangeEdges contract around the list-returning EdgeSwapper branch (R/tree_rearrangement.R:42-69) — entirely untested, since no in-package swapper returns a list of lists. Two things look wrong on paper: when all candidates are worse, :69 still returns one, and the search regenerates the identical neighbourhood every iteration; and hits <- hits + nBest at :57 counts duplicates and input-identical candidates.
2. .NonDuplicateRoot()'s choice of which root edge to exclude flips on nEdgeRight ∈ {1, 3} for undocumented reasons, and the SPR() crash lives next to it. It also assumes edge 1 is a root edge — SPRSwap() calls it with no ordering check, unlike SPR(), which Preorder()s first.
3. Whether the stopAtScore bugs have analogues in Resample()/Jackknife().

Latent: Ratchet()'s Argument() helper (R/Ratchet.R:156) resolves via get(arg) against the enclosing frame — rename any of searchIter/searchHits/stopAtScore/stopAtPeak/stopAtPlateau and it silently reaches past the frame into the namespace or globals rather than erroring.

Seam status: still yielding — nowhere near dry.