fix(ai-review): review exact changes with complete PR context

[absorbb]

PR reviews could inspect files from a different ref than the requested head and compare against the base branch tip, including unrelated changes when a branch was behind. Check out the exact head and derive the PR diff from its merge base; commit reviews retain parent-to-commit comparison.

Prepare PR title/body and independently paginated review threads/comments before analysis. Tighten the prompt to validate concrete failure scenarios, omit speculative comments, consider intentional tradeoffs, and avoid repeating findings already covered by the discussion. review_instructions now supplements the core rules.

Intentional decisions and tradeoffs

• This is PR 1 of the review-quality work: model (gpt-5.6-terra), effort, CLI/action pins, and direct APPROVE posting remain unchanged.
• Context preparation stays inline because reusable workflows execute against the caller checkout, which does not contain this repository's helper scripts.
• A PR-head mismatch during context preparation fails the run rather than mixing a newer PR context with an older reviewed head.
• The optional decisions section is documented for caller PR descriptions/templates; consumers are not modified in this PR.
• Structured publishing, incremental review, and model/runtime comparisons are deferred to subsequent PRs.

Validation

• python3 -m unittest discover -s tests -v: 5 regression tests passed (branch behind main, parent-to-commit review, incorrect checkout, changed PR head, prompt/checkout contract, and multi-page API response handling).
• actionlint 1.7.7: passed.
• Executed the exact context API code against a real PR: fetched its title/body, one resolved thread, and both comments including the dismissal.
• git diff --check: passed.

Live Codex analysis was not invoked; this PR validates workflow plumbing and prompt generation without posting test reviews.

[jitsu-code-review]

Reviewed the exact-head checkout and merge-base diff preparation, PR context/history pagination, prompt contract, and regression coverage. No high-confidence defects found in this change set.