…budget

The Required OpenCode Review check timed out red on every gyeot PR
(ContextualWisdomLab/gyeot #4, #5, #6 — including a one-file docs PR),
so the model-pool step is structurally, not transiently, broken.

Root cause (from the job logs): the 3-4 front-loaded mini models
(o4-mini, gpt-5-mini, gpt-5-nano) each hang/rate-limit and are killed
at the 90-min (5400s) per-model `timeout` (exit 124). Three hangs burn
~270 min before the pool can reach any high-quota fallback
(deepseek/mistral/llama, candidates #6-11), so the 350-min hard step
timeout kills the job first — opaquely, with no recorded review_status.

Fix:
- OPENCODE_RUN_TIMEOUT_SECONDS 5400 -> 1800 (30 min): still ample for a
  deep tool-using review (5x the 600s that was too short), but low
  enough that the full 13-candidate list is traversable inside the
  350-min step even if every leading model hangs, so a responsive
  fallback is reached.
- OPENCODE_TOTAL_RETRY_BUDGET_SECONDS 0 -> 20400 (340 min): a clean
  internal deadline 10 min under the step timeout. The pool already
  caps each run at min(run_timeout, deadline-now) and exits with an
  explicit "retry deadline elapsed" reason + recorded status, so a
  genuine exhaustion now prints its cause instead of GitHub's opaque
  "timed out after 350 minutes".

Also updates the strix quick-gate assertions that pin these two values.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RTAMs4bpSZS77Xe3RQjv9P