Fix reviewer bot never re-approving after blockers are addressed

[awinogradov]

Context

The reviewer bot could never lift its own CHANGES_REQUESTED after a PR author addressed a blocker via discussion — the react-mode verdict re-evaluation shipped in #275 never fired in production.

Root cause

Three gh pr view calls in code-review-action/action.yml run before actions/checkout, in a workspace that has no .git yet. Without -R, gh pr view resolves the target repo from the local git remote, so each call fails with fatal: not a git repository.

Confirmed on a live react run — both pre-checkout reads failed in the same job, before checkout:

Run gh pr view <n> --json reviewDecision
  fatal: not a git repository (or any of the parent directories): .git
  ::warning:: treating bot as not blocking (issue #275 re-verdict will not arm)
  bot_blocking=false
Run gh pr view <n> --json headRefOid
  fatal: not a git repository ...
Checkout repository (actions/checkout@v7)   <- .git first exists here
classify: NEEDS_REVERDICT=false             <- because bot_blocking=false
pr-answer: "updatedVerdict": null           <- reply only; verdict never re-evaluated

Because bot_blocking is always false in react mode, authorAcknowledgesWhileBlocking in classifyReaction.ts can never arm the re-verdict, so CHANGES_REQUESTED is never lifted.

What / Solution

Pass the repo explicitly to the three pre-checkout gh pr view calls:

• reviewDecision (ctx step) — the damaging one; its failure disarmed the re-verdict. Adjacent comment refreshed to explain the pre-checkout -R requirement.
• headRefOid (resolve-sha step) — was silently degrading the react checkout to the merge ref.
• isDraft (ctx step) — was failing open to "not draft".

Post-checkout gh pr view calls already have a git remote and are unchanged.

Scope

• The fail-open direction (a genuine query failure → bot_blocking=false) is intentionally kept; with the repo resolved, the deterministic pre-checkout failure no longer occurs, so no retry is added.
• The authorAcknowledgesWhileBlocking acknowledgement vocabulary (which phrasings auto-arm) is unchanged — a separate concern. An explicit "re-review"/"PTAL" arms regardless.

Testing

action.yml shell is not unit-testable (the composite action is not covered by actionlint, which is gated by the CI Validate Actions job). Verify on a live blocking PR: an author acknowledgement (or "PTAL") should now log bot_blocking=true, NEEDS_REVERDICT: true, and a non-null updatedVerdict, with no "re-verdict will not arm" warning.

Related to #275

[symbiot-bot]

✅ No issues found.

---

Review run summary 🤖

Metric	Value
Mode	review
Model	claude-sonnet-5
Model time	122.7s
Tool round-trips	20
Assistant turns	21
Tokens in / out	1489727 / 9499
Cache read / write	1392385 / 97313
Cost (USD)	$0.93