Bounded implement-gate-fix loop with 30-min checkpointing and stuck-agent guard

[Ryan-Atkinson87]

Implement the implement → gate → fix inner loop with long-session checkpointing — Spec §4.2 steps c–d plus §18.9. This is one in-flight issue's path from a fresh feature branch to passing gates (or a parked blocker).

Spec §4.2 (steps b–d), §18.9 (long-session checkpointing + stuck-agent guard), §15/§16.3 limits.test_fix_attempts (default 3), §9 (park as blocker).

Acceptance criteria

• On a per-issue feature branch (Feature-branch lifecycle: per-issue branch, review diff, serialised merge, delete #11), invoke the AI implementor session (adapter Implement Claude provider adapter via Claude Agent SDK #17, rendered instructions Provider-format renderer (compose personas to CLAUDE.md + skill files) #22) to write code against the issue's acceptance criteria, writing new tests as part of implement (Spec §17.3 implement stage).
• Run the full test + lint + typecheck gate (Test/lint/typecheck gate runner with structured per-repo results #12) after the session (Spec §4.2 step d).
• On gate failure, feed structured results back to a bounded number of AI fix attempts (limits.test_fix_attempts, default 3, from project.yaml project.yaml Pydantic schema and fail-fast loader #4); still failing after the bound → park the issue and record a structured blocker (§9), do not loop forever.
• 30-minute soft checkpoint (§18.9): when a session crosses 30 min wall-clock, signal a stop at the next safe boundary (after the current tool call/edit completes and changes are committed), log progress to the state marker (Per-issue state markers and resume-or-reset crash recovery #24), commit WIP on the feature branch, then start a fresh session that resumes the same issue (not the loop).
• Checkpoint stop only ever happens at a committed point, so it never triggers the discard-and-restart path (§18.4) — verified.
• Stuck-agent guard (§18.9): if the same issue hits the checkpoint 3 times without producing an internal PR, park it as a blocker and notify, instead of a fourth session. The checkpoint count is per-issue and resets when the issue produces a PR or is parked.
• State markers (Per-issue state markers and resume-or-reset crash recovery #24) updated at each step boundary so a crash mid-loop recovers correctly.
• Tests: gate pass first try; gate fail then fix within bound; exceed test-fix bound → parked blocker; checkpoint at 30 min commits + resumes same issue; third checkpoint without PR → parked.

Notes

Backend API-contract verification (§4.2 step e) runs after this loop and before review; it is wired by the driver (#29). Notify-on-park uses the escalation path (Phase 5) — emit the structured blocker record here; channel delivery is Phase 5.

Depends on: #11, #12, #17, #22, #24