feat(review-tutor): add claude-code connector

[ElbertePlinio]

Changed

• add the flag-gated Claude Code connector for discovery, isolated read-only invocation, stream parsing, usage, and mapped failures
• keep assistant answers in the runner-owned sink and clear connector terminal/model state after every completed run
• omit usage when Claude Code does not report it and simplify the Claude Code 2.1+ version gate
• document that the server enforces thinking levels through model membership
• register unavailable harness state without exposing models
• document the connector credential environment boundary
• add fake/fixture coverage for discovery, parsing, sequential-run isolation, runner isolation/cancellation/redaction, and /api/state

Tested

• cd packages/review-tutor && npm run typecheck
• npx vitest run packages/review-tutor/test — 302 passed
• bun run check — 593 passed, 25 skipped; lint, coverage, and builds passed

Not tested

• live unauthenticated ask path
• macOS and Windows

Risk

• Risk class: serious backend
• Child-process isolation and credential forwarding are the main risk; exact argv, environment allowlist, init tool/permission checks, cancellation, and redaction are pinned by tests.
• Stacked on the hardened connector boundary in PR feat(review-tutor): add harness connector boundary #76 and rebased onto fb6a9cc.

Real smoke evidence

• Flag: reviewTutorHarnessConnectors
• Prompt via stdin: Reply with the single word ok.
• Exact argv: claude -p --output-format stream-json --verbose --include-partial-messages --model sonnet --effort low --tools Read,Grep,Glob --permission-mode dontAsk --strict-mcp-config --setting-sources '' --disable-slash-commands --no-session-persistence --max-turns 8
• Exit code: 0; answer: ok
• Sanitized real init: {"subtype":"init","model":"claude-sonnet-5","permissionMode":"dontAsk","tools":["Glob","Grep","Read"],"claude_code_version":"2.1.245"}
• The real tools are a strict subset of {Read, Grep, Glob}, so the strict subset assertion remains in place.
• No configured secret pattern or [redacted] marker appeared in the captured events.

Refs #63

[ElbertePlinio]

Ready for human review at 089ff84 (stack: #76 → #77). Panel: Opus correctness (P1 evidence, 2×P2, P3s → fixed and verified), Grok security/process clean. Real smoke: sonnet/low → "ok", init tools Glob/Grep/Read, dontAsk. Deferred P3s (message quality, test strength) tracked in a follow-up on #63. CI green.