fix(review): claude reviewer path actually works (#51 C2/C3/H3)

[huiyu]

Campaign-2 fix #2 — CRITICAL C2+C3 (re-opens #51 B4). From the independent re-verification: #84 fixed only the codex direction; the entire claude reviewer path (Codex→Claude) was dead.

Bugs

• C2 claude's --json-schema takes inline JSON, not a file path (verified against claude 2.1.241 — a path exits 1 with "not valid JSON"). The code reused codex's schemaPath for both.
• C3 claude -p --output-format json returns an array of stream events, not {findings}. The runner treated the event list as the findings list → 9 phantom empty findings (even once C2 is fixed).

Fixes

• C2 --json-schema JSON.stringify(FINDINGS_SCHEMA) on the claude branch; codex's --output-schema keeps the file path.
• C3 extractClaudeStructuredOutput() unwraps the terminal result event (or {result} envelope) to the schema-conforming payload before normalizing; a non-conforming reviewer degrades to invalid, never phantoms.
• H3 (meta-lesson) the tests were falsely green — fabricated {findings} claude never emits + name-only smoke test. Now: the invocation test asserts --json-schema is inline JSON not the path (deterministic, always runs); the runner test drives the real event-array shape and asserts one real finding not one-per-event; a live binary-gated test actually runs claude with our inline schema and asserts the flag-parse error is absent.

Verification

• test:cross-review: 23/23 — including the live claude smoke test (ran claude 17.7s, inline schema accepted).
• Adapters regenerated — the codex mu-review mirror carried the identical bug, and it is the copy that runs under the Codex host.

🤖 Generated with Claude Code