factory: reviewer verdict — structured output + review against the issue, not just the diff

[ALLiDoizCode]

Gives the factory reviewer a verdict channel and the missing Spec axis. Port of toon-protocol/toon-meta#299 to buzz (toon-meta#275, epic toon-meta#270).

What changes

• Structured verdict, enforced. Every reviewer run must end with <review>{"verdict":"clean"|"blocking","blockingFindings":[{file,line,summary,why}]}</review>. The schema is declared as sandcastle.Output.object({ tag: "review", schema }); extraction/validation is performed by the new .sandcastle/review-verdict.ts because — verified against the @ai-hero/sandcastle@0.12.0 dist — the engine's structured-output surface exists only on top-level run(), and the sandbox.run() the runners use silently ignores an output option. Extraction mirrors the engine's semantics exactly (last tag wins, fence-aware unwrap, JSON parse, zod schema validation).
• Malformed verdicts fail the run. One retry via SandboxRunResult.resume() (the engine's own structured-output retry mechanism, exactly one iteration, with engine-style error feedback); if the verdict is still malformed the runner throws and the Actions job goes red — a missing verdict is never mistaken for a clean one.
• Spec axis. The implement runner forwards ISSUE_NUMBER/ISSUE_TITLE via promptArgs; the standalone agent:review runner resolves the issue from the PR body's Closes #n (no reference → Standards-only review). The prompt instructs the reviewer to gh issue view the target FIRST and review the diff against its acceptance criteria. buzz's repo-specific review steps and gate commands (just fmt-check/clippy/test-unit, just desktop-*/web-*) are preserved.
• Blocking side effects. Findings are posted as a PR review (event COMMENT) and needs:human is applied — pure REST via gh api (porcelain gh pr edit is broken org-wide by the projectCards GraphQL deprecation). In auto-merge mode a blocking verdict refuses to merge and falls back to PR mode.
• Standalone runner unbroken. Adopts connector#634's proven first-live-run fix: agent-review.yml now checks out main (git refuses one branch in two worktrees — sandcastle checks the PR head out in its own worktree under .sandcastle/worktrees/), and the runner materialises the PR head as a local branch so the engine's worktree add -b … HEAD fallback cannot silently review an empty diff.
• main.ts (the reserved autonomous loop) passes ISSUE_NUMBER/ISSUE_TITLE to its reviewer so the shared prompt's placeholders always resolve; it does not yet consume the verdict (auto-merge wiring is toon-meta#270 work).

Forensics steps (toon-meta#278) and the ~200k context-budget prompt wording (toon-meta#273) are preserved byte-for-byte. zod was already a root devDependency (^3.25.0, v3) — no dependency or lockfile change.

Verification

• Extraction smoke-tested in-repo via tsx (valid clean/blocking, fenced JSON, last-tag-wins, missing tag, invalid JSON, bad enum, blocking-without-findings, clean-with-findings, null line) — all 10 pass; temp script deleted before commit.
• Both runners import-smoke-tested to their env guards (fail on SANDCASTLE_ISSUE_NUMBER/SANDCASTLE_PR_NUMBER, nothing else).
• The CI that gates these paths (the changes job's release/mobile/desktop contract scripts + file-size ratchet tests) run green locally; actionlint green on the workflow. Biome findings in .sandcastle/ are pre-existing on main (the directory is not format-gated); no new format errors introduced.
• Live proof (clean + blocking on real PRs) comes from the first agent:review run after merge, since the label→runner executes main's code.

Part of toon-protocol/toon-meta#270
Part of toon-protocol/toon-meta#275

🤖 Generated with Claude Code