feat(review): drop the default reviewer timeout — reviewers run until done

Reviewers were killed at a wall-clock default (10m, later 20m) that conflated 'slow because working' with 'hung'. Measured: a reviewer at ~90% completion was killed with all work discarded, and legitimate review subagents run 12+ minutes with zero parent output — so no inactivity watchdog can save this either (a headless child offers no liveness signal that distinguishes a quiet working subagent from a hang). A controlled A/B also showed entire review is not slower than the same skills invoked directly in a session (9m57s plain vs 9m08s through entire, same branch, same skills) — the duration is the task, and killing it at an arbitrary ceiling only discards paid-for work.

Reviewers now run until they finish, exactly like a directly-invoked skill. --timeout remains as an explicit opt-in hard cap (CI, cost bounds); its default is 0/none. The judge keeps its own 5m default regardless — it is a single text-generation call, and the reviewer no-cap sentinel must not leak into its bound (judgeTimeoutArg).