fix(pr-review): robustly extract JSON from codex CLI transcript output

The codex reviewer (gautam-chatgpt) intermittently failed with "No valid
JSON found in response" on gbharg/pi-mono PRs #28-32. Root cause: `codex
exec` can frame its answer with a session transcript, a leading `codex`
marker line, and a trailing token-usage epilogue, and that surrounding
prose often contains stray `{`/`}` (shell snippets, file trees, JSON
examples in skill docs). The old parser used a naive first-`{`/last-`}`
slice, which captured the prose braces and threw — explaining why the
same PR sometimes parsed and sometimes did not.

parse-response.js now:
- scans for balanced `{...}` spans with a string-aware stack (immune to
  stray/unbalanced braces in surrounding prose), and
- prefers the LAST review-shaped object (codex prints the answer last), and
- falls back to a quote-insensitive rescan when prose has an unmatched `"`
  that would otherwise hide the real object inside a phantom string.
- adds parseReviewWithRepair(): one repair re-ask demanding pure JSON
  (faithful conversion, no fabricated findings) as a last-resort safety net
  for truly truncated/empty output.

review-handler.js calls parseReviewWithRepair (shared by codex/gemini/
claude). gemini/claude emit clean or fenced JSON and are unaffected.

Verified against the real captured codex transcript (full 19KB session
log on stdout) — now extracts the correct review. Added tests for the
transcript/fenced/stray-brace cases and the repair retry. 95/95 green.

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