fix: FUNCTION_INVOCATION_TIMEOUT on big sorts — time-budget the pipeline

The sort route died at Vercel's 120s limit on slow-API days: the Anthropic
client had NO timeout (SDK default is 10 minutes, plus 2 silent SDK-internal
retries), so one stalled call could eat the whole function budget, and the
pipeline had no notion of a deadline.

- sort route maxDuration 120 -> 300 (publish already runs at 300)
- every pipeline call: 60s per-call timeout, SDK retries off (the pipeline's
  own retry loop is the single retry layer)
- SORT_TIME_BUDGET_MS (250s) threaded through group/verify/merge: retries and
  calls stop when the budget is spent, so a slow run returns a usable partial
  sort instead of being killed by the platform
- merge-check route budgets 25s under its 30s cap
- new tests: per-call timeout/maxRetries wiring, zero-budget skip, partial
  results when verify/merge budget runs out