Imprevista

Reliability fixes from review: trials keying, validator dedupe, crank tolerance, concurrency (#43)

- allocator: trials latestById used wrong key (trial_id vs id) - probesActive was blind - validator: validateAll now dedupes latest-per-id and treats non-pending verdicts as terminal (stops ~5 junk re-verdict records/day; trials.jsonl had 68 copies of trial_0001)

• allocator: trials latestById used wrong key (trial_id vs id) - probesActive was blind
• validator: validateAll now dedupes latest-per-id and treats non-pending verdicts as

terminal (stops ~5 junk re-verdict records/day; trials.jsonl had 68 copies of trial_0001)

• crank source: tolerate malformed/bad-date lines with warning + count instead of crashing

all daily validation

• allocator: UTC isoWeek, atomic writes, --apply+--product no longer clobbers full report,

promote check needs 9 buckets (5 baseline minimum)

• workflows: shared learnings-ledger-push concurrency group + push rebase-retry

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Files Changed