Add TNT-style transient autoconstraint (default OFF, opt-in)

[ms609]

Summary

Adds transientAutoconst (+ transientAutoconstMinReps) to SearchControl() — a per-replicate, TNT-style transient autoconstraint, distinct from the existing rigid consensusConstrain. Default OFF, additive, flag-gated.

Each replicate:

1. builds its Wagner start unconstrained;
2. forms a constraint = strict consensus of the previous replicates' best-score pool trees ∩ this replicate's own Wagner-start splits (TNT's "previous and wagner");
3. enforces it only over the early stages (initial TBR + first sectorial pass);
4. releases it before ratchet / drift / later-sectorial / final-TBR.

The first replicate of each hit (rep 0, or any rep immediately after a new best) is left unconstrained. Because the constraint refreshes every replicate and is never permanently locked, it can guide early search without entrenching a basin — unlike the rigid path, which can permanently lock splits that forbid the optimum.

Constraint-mapping fix (valid-by-construction)

map_constraint_nodes matches a rooted node subtree exactly, so a split whose tip-0-free clade is not a rooted clade of the current tree fails to map (constraint_node == -1) and would then spuriously reject moves. Such splits are dropped and the constraint rebuilt from the mapped subset, so every enforced split maps to a real node. The unmappable fraction scales with tree depth (tip-0's ancestor count), so it is ~19% at 62 tips but only ~1% at 482 tips.

Safety / no-op-when-off

• Default OFF; mutually exclusive with the rigid path and with a user constraint.
• Env kill-switch TS_NO_TRANSIENT_AUTOCONST.
• Off-path byte-identical to the prior engine (verified: identical replicate trajectories / candidate counts vs the pristine tip).
• Existing constraint + SearchControl test suites pass unchanged.

Validation (why it's opt-in, not a default)

Built and A/B-tested on project5432 (482 tips) — the corpus's lone never-reached floor (optimum 1943; TS plateaus ~1948–1949). Matched engine, matched seeds, exact thorough + sectorMax45 + tbrMaxHits20 config:

• Engages healthily: median 124 enforced splits/rep, enforced/kept ≈ 0.99 — a large, real constraint, so the result is interpretable (not a frozen artifact).
• Clean null: no dose-response (a seed with 35 engaged reps returned the worst ON result; a seed with 13 returned the best), and nothing came within 4 steps of 1943; no matched-seed effect.
• Mechanism of the null: the first-rep-of-hit rule makes it engage post-convergence, so the consensus it builds is the already-converged 1950-basin consensus — releasing late drops the search straight back into that basin.

Kept as a documented default-OFF opt-in knob (same disposition as other built-and-A/B'd TNT features that came up null).

🤖 Generated with Claude Code