generic-http AI gate: schema-invalid output goes red, prose-factcheck misses PR text, wiring tests are per-mutation pins #260

Context

PR #259 (#252) added a generic-http runner for the AI gates. It only runs once vars.AI_GATE_HTTP_BASE_URL is registered, so everything below is dormant today. The integration review and the test-quality audits of that PR left these gaps.

1. Schema-invalid model output turns an advisory check red

response_format: json_object (scripts/ai_gate_http_runner.py:73) guarantees JSON, not schema conformance. Valid JSON that fails the verdict schema, or names the wrong gate, makes scripts/ai_gate_verdict.py exit 2 (l.66-74). The 판정 반영 step in .github/workflows/_ai-gate.yml (~l.451) turns exit 2 into a red job regardless of blocking: false. With a non-compliant model, every PR would show four red "게이트가 고장났다" checks. docs/CI-GATES.md (l.60-65) lists the generic-http failures that fall to unavailable but not this one.

Proposal: map exit 2 to unavailable for the generic-http runner (or document it explicitly), with a test.

2. prose-factcheck on generic-http never sees the PR text

build_messages sends the prompt and the diff only (scripts/ai_gate_http_runner.py:69). .github/ai-gates/prose-factcheck.md:3-4 defines the gate's subject as the PR title/body and commit messages, so prose-factcheck-generic-http (.github/workflows/ai-gates-quality.yml:88) can only judge prose inside the diff. expect_nonempty does not catch this, because examined.files comes from the diff.

Proposal: write the PR title/body and commit messages to a second $RUNNER_TEMP file and pass it to the runner; or document the limitation.

3. Workflow wiring tests are per-mutation pins

Three test-quality audits in PR #259 each found new workflow mutations that changed CI behaviour while all tests stayed green (step order, shared-step if: lines, defaults injected by the tests themselves). Each was fixed with a verbatim pin. A structural test would end this whack-a-mole: build the step graph from _ai-gate.yml and assert data flow (every file a step reads is written by an earlier step in the same runner path; every shared step runs for both runners).

Also noted (low): an extra --timeout 1 in one test, the == true skip line not counted, a duplicate gate_name key in a fixture.

Done when

• Items 1 and 2 decided and implemented or documented, each with a test.
• The structural workflow test replaces or backs the verbatim pins.