completed-tasks.md

Closed Tasks — Decisions Worth Not Re-Litigating

FROZEN 2026-08-04 — closed to new rows. Task state moved to agent-issues/TreeSearch issues; a closed issue (wontfix, or closed as not planned) is now where a terminal decision is recorded, and to-do.md has been deleted.

This file stays exactly as it is because its content is the one thing the tracker cannot replace: offline, grep-able reasoning for decisions whose rationale is not in any commit. Keep consulting it — the instruction below is still live — but do not append to it.

This is not a full archive of every completed task. Routine fixes live in git history and merged PRs; do not duplicate them here. This file keeps only the closures whose reasoning is not recoverable from a commit: not-a-bug determinations, superseded/ruled-out designs, and negative experimental results — the things an agent would otherwise waste budget re-investigating.

How to consult: before investigating a recurring symptom or reopening a closed T-nnn/<Letter>-nnn, grep this file (by ID or keyword) for a prior closure. Do not Read it whole. Old task IDs referenced here remain valid (see AGENTS.md) and need not be renamed.

How to add a row: only when a task closes without a routine fix — a negative result, a "not a bug", or a superseded decision. One row, terminal decision + a pointer to the write-up. Routine fixes get a one-line row in to-do.md's removal commit, not an entry here.

---

Not a bug / scoring-method confounds

ID	Topic	Decision
T-242	Agnarsson2004 IW "~2% hit rate"	Display bug only. ThreadSafePool::extract_into() reset hits_to_best to distinct-topology count, not replicate hits. Search algorithm unaffected; real hit rate ~60–67%. Fixed bc19667f2; regression test in test-ts-parallel.R.
T-247	XPIWE Vinther2008 score ≠ TNT	Not a bug. Discrepancy is entirely Brazeau three-pass vs standard-Fitch inapplicable handling. TreeSearch's tree (EW=79) is genuinely better under three-pass scoring. XPIWE uses eff_k in all paths — verified correct.
T-265 / T-249 / T-264	Per-replicate "regression" vs TNT	Scoring-method confound, not a regression. T-249/T-264 compared Brazeau-scored TreeSearch to EW-scored TNT (apparent gap +17.8 steps; real EW-vs-EW gap +2.2, 5/11 datasets at 0). Future TNT comparisons MUST use fitch_mode() for apples-to-apples.
T-211	Stale final_ in temper candidate scoring	Not worth fixing. Conservative-only: stale final_ biases Boltzmann screening but temper_full_rescore gates every accepted move. Fix cost (per-candidate rescore or full save/restore) exceeds negligible SA benefit.
T-325 / T-326	MPT-enum collapse-dedup "30 trees" red-team finding (2026-07-02)	Already fixed before filing — stale build, not stale source. Root defect (enum pool deduped on full-key add() instead of collapsed-key add_collapsed()) was fixed 8 days earlier by 0daea13f (2026-06-25): R-level collapse=TRUE default (ts_collapse_pool) + matching C++ enum change. The finder cited ts_tbr.cpp:2345,2377, which are that commit's pre-fix line numbers — the trace read a version of the file predating the fix even though HEAD had it. Verified by building both revisions fresh: 0daea13f^ reproduces 30 trees/n_topologies=30 exactly as reported; current HEAD returns 1. The decisive fix layer is the R-level collapse=TRUE default, not the C++ enum change alone (confirmed by disabling the latter via TS_ENUM_RESOLVED=1 — still 1). T-326's regression ask is already covered by test-MaximizeParsimony-features.R:402-421 (added in the same commit; re-ran green). Residual test-ts-collapsed.R:159-179 weak assertion is now redundant-weak, not dangerous-weak — not worth a dedicated task. Full writeup: dev/red-team/log.md area 11 post-hoc correction.

Superseded / ruled-out designs

ID	Topic	Decision
T-183	Pool-seeded Wagner / consensus backbone	Superseded by consensusConstrain (ts_driven.cpp), which constrains the whole replicate pipeline, not just Wagner. Marginal starting-tree value given the NNI→TBR pipeline.
T-198–201	Boltzmann parallel tempering	Ruled out by T-199: 0% cold↔warm swap acceptance across all datasets. PCSA component salvaged as T-207/PR #227. See pt-evaluation expertise note.
T-185	IQ-TREE acceleration ideas	Stochastic NNI-perturbation worth trying (→T-186, implemented). Batch NNI not worthwhile — see batch-nni expertise note.
VINE	Vine-style embedding + NJ-decoder search (bioRxiv 10.64898/2025.12.24.696405)	Ruled out for both candidate slots. (i) LeastSquaresTree(): the refit discards the decoder's branch lengths, so RSS-after-refit is piecewise constant in the input matrix (6 distinct values over 241 grid points); the only differentiable surrogate misranks topologies (Spearman +0.10–+0.36 among RF==2 neighbours, unchanged by scale/affine correction). An oracle hill-climb on the true objective never beat NNI+SPR. (ii) Sector starts (build_ras_sector): perturbed-NJ makes better individual starts than RAS Wagner (2.5 vs 4.0 excess) but loses on best-of-5 (1 win / 7 ties / 7 losses) because its restarts are half as diverse (CID 0.162 vs 0.328) — multi-start sector solving is diversity-limited, not start-quality-limited. Two unrelated leads recorded in the write-up: method = "ols" returns topologies worse than the NJ start when rescored under NNLS (5/5 targets), and ts_ls.{cpp,h} has no certify_unrooted equivalent. Full write-up: dev/benchmarks/vine-embedding-ls-search.md.
MAXINFO / AML	Maximum-information tree; concave-per-branch parsimony	Both ruled out. The published Nishimaki-Sato 2026 metric is topology-independent at realistic n (consensus collapse), inflatable by invariant sites, and label-invariant (a complement child scores a perfect 1.000). Its sound chance-corrected form collapses onto weighted parsimony. The surviving concave-per-branch exit is Ancestral Maximum Likelihood — NP-hard (Addario-Berry et al. 2004) and statistically inconsistent (shrinks internal edges to nothing as k grows, arXiv:0802.0914). AML is the transpose of implied weighting, which is exactly why IW is cheap and this is not. Only open question: AML vs IW at >=6 taxa. Write-up: dev/plans/2026-08-03-concave-branch-criterion-is-aml.md.

Search-tuning experiments — settled, don't re-run

Each row records a benchmark whose conclusion fixed a default or killed an idea. Detailed data is in the named dev/benchmarks/ write-up; re-running wastes Hamilton/GHA budget unless the underlying kernel has changed.

ID	Experiment	Conclusion
T-254	Drift cycles (0 vs 2)	Drift gives zero score/MPT/diversity benefit, costs 10–22% of reps → driftCycles=0 in default+thorough (T-255). drift_mpt_analysis.md.
T-256	Sectorial intensity	Doubling/tripling xss+rss rounds → no score gain. Current xss=3, rss=1 sufficient.
T-259	Ratchet cycle count	Reducing 12→8/6/4 is mixed-to-worse; default 12 justified (3-seed, directional).
T-274	NNI-perturb cycles (thorough)	59–69% overhead, ≤0.1-step benefit → nniPerturbCycles=0 in thorough. bench_t274_nni_perturb.R.
T-248	SA phase tuning (large)	annealCycles=3 no significant gain over AC=1 (p>0.5) → AC=1 in large preset (~6% faster).
F-029 (T-269)	Outer-cycle count	Higher outerCycles cuts replicate throughput with no score gain → outerCycles=2 optimal.
PA-002	XSS↔TBR cycling	Benefit scales with tree size, not scoring mode. ≤88t: pure overhead. 180t: −6.8 to −9.8 EW steps. No IW-specific treatment needed. expt_tbr_xss_v2_results.rds.
PA-003	Targeted post-clip sector search	NET HARMFUL — local sector refinement after each TBR move steers into worse basins (+17 to +34 steps at 180t). Confirms XSS-as-a-separate-phase-after-TBR is correct. Do not implement.
PA-001 → F-030	TBR clip ordering	PA-001's "tips-first falsified" was an artifact (clip_order reached only ~10% of TBR calls). F-030 with full propagation: TIPS_FIRST gives +8–13% throughput on 75–88t thorough preset. Default unchanged (clipOrder=0); PR #239.
T-289 / T-289f	Prune-reinsert polish (large)	TBR polish catastrophic at ≥206t (0 reps); NNI polish helps 131–180t → pruneReinsertCycles=5, pruneReinsertNni=TRUE in large preset, PR disabled elsewhere. t289f_pr_nni_polish.csv.
G-001 (T-290)	Brazeau-track phase profiling	Wagner is 3.6–5.2× costlier under Brazeau than Fitch; Fitch-tuned presets remain appropriate. wagnerStarts=3 justified (better starting topology dominates when TBR convergence > budget).
F-006 (T-253)	Gap characterization	ntax is the dominant predictor of TNT gap (ρ≈0.63); nchar matters only >2000. t253_gap_characterization.md.
F-004 (T-252)	MorphoBank baseline	≤35t converge at 30s; 66–135t still improving at 120s; project4284 (4062t) can't finish 1 replicate. CSVs in dev/benchmarks/.
T-251	TNT trajectory analysis	Drift 30–170× less efficient than the next-worst phase; TNT spends ~67% in sectorial search vs TS's single pass. tnt_trajectory_analysis.md.
T-250	TNT Fitch kernel disassembly	TNT is 32-bit scalar, no SIMD; TreeSearch has ~4× kernel throughput. TNT's convergence edge is strategic, not implementation. tnt_disassembly_analysis.md.
T-260	VTune TBR overhead	Non-scoring overhead = 37.8% of TBR (StateSnapshot 14.6%, reset_states 9.1%) → T-261/262/263 (done). vtune_tbr_analysis.md.