perf(tbr): finish T-300 — incremental rescore for TBR-rerooting accepts

[ms609]

Fixes #38

What was actually left

Most of #38 was already done. T-300 landed its SPR half long ago (f531bbcd EW, 014ccdea NA), so the issue text — and its line numbers (ts_tbr.cpp:1134) — describe code that no longer exists. The residue was accepts that rerooted the clipped fragment: those still fell through to full_rescore.

apply_tbr_move reverses the parent/child links along clip_node..reroot_parent, so every node on that path gains new children — and after the reversal that path is clip_nodes rootward chain. One extra dirty seed at clip_node therefore covers the rerooting case exactly, with no new machinery: the NA kernel already took a third seed for precisely this region (it backs exact_verify_sweep), so this adds the same parameter to the EW/IW kernel and drops the is_spr gate at the accept site.

Read this before merging: the lever is spent

dataset × mode	accept branch, % of tbr_search wall
Vinther2008 EW / IW	0.51 % / 0.43 %
Agnarsson2004 EW / IW	0.18 % / 0.22 %
Zanol2014 EW	0.22 %

Post-T-300 the whole accept branch is 0.18–0.51 % of tbr_search wall (in-DLL na_t_accept_ms, 15 paired cells, arms interleaved because the box is shared). This patch cuts that slice to ×0.901 median (faster in 12/15 cells) ⇒ ≈0.03 % end-to-end — undetectable.

Rerooting accepts are ~18 % of accepts, so that ceiling is real, not thin coverage. The case for merging is codepath unification, not speed — one fewer special case, and the accept path now shares its dirty region with exact_verify_sweep. If you would rather not touch a correctness-critical hot path for no measurable gain, closing this unmerged is a reasonable call; the measurement is the durable part and is recorded either way.

dev/profiling/findings.md and focus-areas.md now mark area 4 AT-LIMIT, and flag that the T-300 rows "19.2 % of NA-path DLL CPU" is the pre-fix number so it is not re-read as a live target.

Correctness

An earlier incremental attempt shipped a systematic delta = -3 (b7303ee5), so this is evidenced rather than argued:

• TS_TBR_ACCEPTCHK — new env-gated in-flight oracle: cross-checks every incremental accept against full_rescore and aborts on drift. Clean over a 70-cell battery (1322 accepts) and over 624 assertions spanning the constrained, pool, IW, XPIWE, sector and drift paths.
• Byte-identical trajectories — scores and accept counts match the pre-patch library in all 15 A/B cells and all 70 battery cells.
• Non-vacuous regression test — the four existing tests assert score identity but cannot tell whether the rerooting arm was ever entered, so they would pass just as happily if it were dead. n_reroot_accepts (src/ts_data.h) is asserted positive for each of EW / IW / NA / NA-IW, so losing coverage fails the test instead of silently voiding it.

CI

ubuntu-arm64 green (full R CMD check). windows fails, and it fails identically on cpp-search — I dispatched a control run (30909722952) to check rather than assume: same job, same step (Set up R dependencies, Cant find package called MaxMin), R CMD check skipped. Pre-existing infrastructure breakage, nothing to do with this branch.

🤖 Generated with Claude Code

[ms609]

🟢 on /code-review