Hook: cost a root's kept build with the model's kept-start cell

[teamleaderleo]

cmux's warm-distance model now keeps the p50 compile per tier and start kind (tiers_by_start, with counts; manaflow-ai/cmux ci/warm-model-selfcal, PR linked below). Every root this hook compares starts from its kept build, and kept starts compile faster than the tier's mix of kept and seed starts. Over 243 owned admissions:

tier	tier p50	kept p50 (n)	seed p50 (n)
near	105 s	91 s (44)	157 s (54)
far	206 s	121 s (10)	247 s (12)
rebuild	373 s	355 s (76)	442 s (47)

• read_warm_model adds kept, the seconds per tier from tiers_by_start[tier]["kept"]. A cell counts only once it has ≥5 compiles (WARM_MIN_START_ROWS, the same rule as cmux's predict()) and a positive finite non-bool p50. A sparse or bad cell is skipped on its own, so it does not discard the whole model. A model without tiers_by_start, and the default model, have kept: {} and cost the tier exactly as before.
• warm_root_costs costs a stamped root at kept.get(tier, tiers[tier]). The unknown and cold costs are unchanged.
• Parity test in test_warm_root_costs_rank_roots_by_predicted_compile, mirroring cmux's test_predict_uses_a_cell_from_five_rows_and_the_tier_below:
  • a 5-row kept cell is used, and a 4-row one falls back to the tier;
  • p50 of -1, a bool, missing, a string n, a list or null all fall back to the tier, for that tier only;
  • a model with no tiers_by_start falls back to the tier.

python3 scripts/test-glaeda-cmux-runner.py -k warm passes.

Backward compatible in both directions: this hook with an old model, and an old hook with the new model, which only adds a key.

🤖 Generated with Claude Code