Regression triage gives different buckets per run: parallel tasks sharing a reference run race on its temp list file

Symptom: rerunning the same regression triage on an identical, deterministic store with N worker threads gave different failure signatures and first-divergence cycles each time (20-55 of 130 failures changed; bucket split 119/6/5, then 115/10/5, then 107/13/10). With 1 worker it was identical across runs. Cause: each task wrote a decoder input list to a path derived from the *reference* run's id (triage/<ref>.catlist). Many failures share one passing reference seed, and each task listed a different signal, so tasks overwrote each other's list while a decoder was reading it. Fix: per-call unique temp files (mkstemp) removed after use. How to catch it: regenerate the phase with identical code and params and diff every per-failure field, then rerun at jobs=1 twice and at jobs=N twice. Deterministic inputs plus nondeterministic outputs that vanish at jobs=1 means shared mutable state. Corollary: a replay/equality check that compared zero chunks must not report REPRODUCED. Here a bogus cycle-0 divergence gave a 1000-cycle window with no full chunks, and the check correctly reported it as not reproduced.