Fixed
- Research verification recognizes section labels such as "(Finding 4)" and
references such as "See conflicts 1 & 2". Quantities and thresholds still
need evidence. Numbers written separately in totals remain checked,
including lists with a comma before "in total".
Maintenance
- Council deadline tests measure prompt preparation separately from worker
startup, preventing false timing failures on busy runners. The watchdog
limit and checks for cancelled children and late output are preserved. - Focused unit CI retains timeout logs and cleans up owned processes before
the job deadline, matching the full unit lane. Interrupted steps preserve
collected output alongside the cancellation message. - Full unit shard weights include hosted macOS timings to spread process-heavy
suites across the four shards.