fix(go): fall back to the harness when the .ai() gates are unavailable

[vbp1]

Problem

intake_phase and coverage_gate in the Go port ask the AI seam for a response_format schema and propagate every failure out of the reasoner. Two real configurations hit this:

1. A node whose harness is not OpenRouter-backed. node.BuildAgent attaches AIConfig only when OPENROUTER_API_KEY is set (go/internal/node/node.go), so running with, say, PR_AF_PROVIDER=claude-code and no OpenRouter key fails the entire review at the first phase:

   review execution failed: AI not configured for this agent; set AIConfig in agent Config
   

2. A provider that rejects structured output — the case fix: fall back to harness when .ai() structured output is unsupported #60 fixed on the Python side.

The Python implementation already handles both: harnesses.py wraps each .ai() call in try/except and falls through to the harness (commit 3c8eca9). The Go port does not, so the two implementations diverge on a failure path that takes down the whole pipeline.

Change

Restores parity with harnesses.py, nothing else:

• intake.go — an unusable AI seam leaves the gate unconfident, so control falls through to the harness classifier already sitting below it (Python's except -> gate_result = None).
• coverage.go — retries the same coverage prompt through the harness, returning {} when the harness result fails to parse (Python's gate.parsed.model_dump() if gate.parsed else {}).

Behaviour is unchanged whenever the AI seam works: the fallbacks only run on the path that previously raised.

Tests

Two new cases in reasoners_test.go:

• TestIntakePhaseAIUnavailableFallsBackToHarness — table over a nil AI seam and a seam that rejects structured output; both must invoke the harness exactly once and return the full intake key set.
• TestCoverageGateAIUnavailableFallsBackToHarness — asserts the harness receives the same coverage prompt, plus the parse-failure path returning {}.

go build ./... && go vet ./... && go test ./... all pass; gofmt clean.

Verification

Ran the full pipeline against a local repo/branch review with PR_AF_PROVIDER=claude-code, PR_AF_MODEL=sonnet and no OPENROUTER_API_KEY:

• before: fails immediately in intake_phase;
• after: HTTP 200 in 611s, 22 agent invocations, all phases completed (intake → anatomy → meta_* → review_dimension → evidence_verifier → compound_finder → synthesis), 7 findings on a diff with 3 planted defects.

Note, out of scope for this PR

Under claude-code the budget accounting stays at zero (total_cost_usd: 0, all per-phase entries 0), because the cost the CLI reports is not the OpenRouter spend PR-AF meters. The USD ceiling is therefore inert for that provider and only the wall-clock cap applies. Worth a separate issue; this PR deliberately does not touch it.

[CLAassistant]

All committers have signed the CLA.