fix(pr-review): reconcile live probot deploy drift into monorepo

The live gemini/codex/claude reviewers run on the VM from non-git deploy
copies that carried production drift never committed back to this repo.
PR #804 (repo-context fix) was merged here AND grafted into the deploy, so
the deploy now = #804 + drift. This commits the drift so the repo matches
what is actually running live (modulo secrets/env).

Reconciled (deploy -> repo):
- pr-review-shared/lib/ask-ai.js: 600s timeout (was 180s; ~12% of claude
  reviews hit SIGTERM/143 on large diffs), killSignal SIGINT for partial
  flush, explicit killedByTimeout timer + clearer "timed out" rejection,
  richer non-zero-exit error (binary/code/signal + both streams), and a
  best-effort /tmp/last-prompt-<bot>.txt dump for post-mortems. Layered on
  top of the #804 sandbox/cwd isolation, which is preserved intact.
- pr-review-shared/lib/__tests__/ask-ai.test.js: assert the deploy error +
  timeout message formats; all #804 sandbox tests kept.
- gemini-pr-reviewer/index.js: model gemini-3.1-pro-preview.
- codex-pr-reviewer/index.js: route via codex-stdin-shim.sh (codex takes the
  prompt as a positional arg, not stdin; askAI pipes via stdin).
- codex-pr-reviewer/codex-stdin-shim.sh: new shim (reads stdin, execs codex).
- claude-pr-reviewer/index.js: add --tools '' to disable tool calls that
  caused "Reached max turns (1)" exit 1 on large diffs.

Not committed (deploy is BEHIND main here, not drift): check-resolution.js,
post-review.js, review-handler.js — main already has the per-bot
checkResolution filter, the postStartedReview dedup, and the mbp-tmux author
notification; the deploy copies predate those. Skipped: .env*, private-key.pem,
start.sh, gemini api/, .gitignore (secrets/runtime/local cruft).

Tests: pr-review-shared vitest 86/86 green; node --check on all changed JS.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>