fix(runtime): unwrap claude CLI JSON envelope in Start() Raw

[jerryfane]

Trace (issue #721)

First live gitmoot skillopt synth run (weak=codex, strong/challenger/judge=claude) rejected all 3 items bad_rubric, weak/strong scores exactly 0.0 every round, empty stderr — 9/9 silent failures.

Root cause:

• ClaudeAdapter.Start runs claude -p --output-format json and returned StartResult{Raw: result.Stdout} — stdout is claude's CLI result envelope ({"type":"result",...,"result":"<answer>"}), not the answer text (internal/runtime/adapter.go:637-642).
• realSkillOptABDeliver returns started.Raw verbatim on the forked-session path (empty RuntimeRef → Start) (internal/cli/skillopt_ab.go:163).
• extractSynthJSONObject then grabs the FIRST balanced {...} = the envelope; parseSynthGeneratedItem finds no context/question/rubric → bad_rubric, continue, no stderr. Judge verdicts fail identically.

Contrast: kimi's Start already unwraps (StartResult{Raw: content}, kimi.go:56); codex returns prose stdout. Claude was the only enveloped Start.

Fix

At the adapter (right altitude). ClaudeAdapter.Start now reuses the existing parseClaudeJSONResult helper — the same one the Deliver path uses to build Summary — to extract the envelope's result field into Raw. Fail-open: falls back to raw stdout when stdout is not the envelope, has no result field, or result is empty (e.g. an older CLI's plain-text --output-format fallback), so unwrap never errors. No new parser was added. RuntimeRef behavior is byte-identical. kimi/codex untouched (kimi already unwraps; codex is prose).

Consumers audited (Start().Raw)

The only reader of Start().Raw is realSkillOptABDeliver's forked-session path (skillopt_ab.go:163). Everything downstream routes through it:

• skillopt synth — challenger and judge (skillOptSynthDeliver = realSkillOptABDeliver). JSON-parsed → was the hard failure. Fixed.
• skillopt ab --judge (skillOptABJudgeDeliver, both single + jury paths). JSON-parsed. Fixed.
• skillopt ab variant deliver + skillopt binary. Answer used as prose; envelope leak was cosmetic/quality, now clean.
• live-AB challenger (agent_dispatch_live_ab.go) — presented to the user for a pick as prose (not JSON-parsed); previously showed the raw envelope, now shows the answer.

Non-empty RuntimeRef paths already went through Deliver → result.Summary (already unwrapped) and are unaffected.

All other adapter.Start call sites — agent restart, agent_dispatch instance start, daemon temp/ephemeral workers, skillopt agent start — read only started.RuntimeRef, never .Raw. Unaffected by the unwrap.

How tested (deterministic, no LLM)

1. Unit (TestClaudeStartUnwrapsJSONEnvelope, internal/runtime): fake subprocess runner returns a realistic claude JSON envelope → Start().Raw == inner result text; plain-answer envelope → the answer; non-JSON stdout → falls back to full stdout; empty result → fallback; envelope without result → fallback. RuntimeRef asserted identical in every case.
2. Consumer regression (TestRealSkillOptABDeliverUnwrapsClaudeEnvelopeForSynth, internal/cli): drives the REAL realSkillOptABDeliver + a real ClaudeAdapter wired to a fake runner (via replaceRuntimeFactory) on the forked-session path, then asserts the delivered answer is the unwrapped inner JSON and that parseSynthGeneratedItem succeeds on it — the exact bad_rubric parse that failed live.
3. Both tests were verified to fail on the pre-fix code (leaked envelope) and pass after.

Gates (go1.26.4, GOTOOLCHAIN-pinned, pipefail + explicit exit checks)

• go build ./... → 0
• go vet ./internal/runtime/ ./internal/cli/ → 0
• go test -count=1 ./internal/runtime/ → ok 0.084s
• go test -count=1 ./internal/cli/ → ok 208.926s
• go test -race -count=1 ./internal/runtime/ → ok 1.114s
• gofmt -l on all changed files → clean

Closes #721

🤖 Generated with Claude Code