Add ecm_late.ep_ (a 25-position slow-to-confirm ECM subset) and wire run_tests.sh to run it sn-only, as a search/extensions-quality metric distinct from the curated suites' solve-count gate.

Positions are ECM problems where the correct move only becomes -- and stays -- the reported best move well into the search (5+ seconds, confirmed stable for 2+ further iterations, non-zero final score) across a majority of 8 independent full-suite runs; derived from mining 8 full 879-position ECM runs at 20s/move for exactly this pattern, then filtering out the "engine got lucky on the last iteration" and "drawn by repetition" false positives that a naive late-solve detector picks up (cross-checked against tests/ECM.md's existing flaky-position list).

sn-only (no sd sweep) because a fixed depth on 1 thread, with no cutoff-sharing across threads to shrink the tree, can blow up unpredictably on the sharpest of these positions; a fixed 15M-node budget keeps wall-clock bounded (~5min/suite) while staying in the right ballpark of the per-thread node share implied by the 8-thread discovery data.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

0 files changed, 0 insertions, 0 deletions