feat: retry Iron Loop executor on API overload (529) with configurable backoff

[davidbijl]

Fixes #6.

Summary

• Layer 1 — executor agent (agents/iron-loop/iron-loop-executor.md): new API Overload (529) Handling section instructs the executor to distinguish pre-write overloads (safe to retry) from mid-write overloads (human gate required) and write the appropriate .status file.
• Layer 2 — state layer (background.js, actions.js, state.js): overload-retry and overload-partial added to the status enum; cleanupStaleInProgress skips overload plans; startAgent resumes an overload-retry plan in-progress instead of picking a new todo plan and blocks with a human-gate error for overload-partial; getAgentStatus surfaces overload states when no lock is held.
• Layer 3 — dashboard (menu-screens.js): AGENT section shows ⏳ retry in Xm — <plan> for scheduled retries and ⚠ partial write — review: <plan> for mid-write overloads.
• Config (settings.js): new retry category with overloadIntervalSeconds (default 600 s / 10 min).
• Tests (tests/overload-retry.test.js): 9 unit tests covering all three layers (icon enum, writeStatus fields, cleanup skip logic, startAgent resume/block paths, dashboard labels, config schema).

Answers to the open questions in #6

1. Preferred layer for retry logic: the executor agent writes the status (Layer 1) and exits; the state layer drives resume/block on the next startAgent call (Layer 2). No ScheduleWakeup dependency — the operator restarts via the menu when ready, or the executor can call ScheduleWakeup if it's available in its context (the agent instructions mention it as optional).
2. Step-level resume vs full restart: full restart from the beginning of the current plan. The plan's completed [x] checkboxes are on disk, so the executor can fast-forward past already-done steps. No separate step-marker mechanism is needed for a first pass.
3. ScheduleWakeup availability: treated as optional in the agent instructions — if available, use it; if not, exit cleanly. The dashboard indicator and the menu's Start Agent button serve as the manual resume path.
4. Scope: all three layers are included, but the changes are minimal and additive — no existing behaviour is modified except cleanupStaleInProgress (now skips overload plans) and startAgent (now checks in-progress before picking todo).

Test plan

• Run node --test tests/overload-retry.test.js — 9 tests, 0 failures
• Run node --test tests/*.test.js — existing suite passes (the 1 pre-existing failure in update.test.js is unrelated to this PR and was failing on main before these changes)
• Manually: simulate an overload-retry status file in a plan under plans/in-progress/, open /ctoc:menu, confirm the dashboard AGENT section shows ⏳ retry in Xm
• Manually: simulate an overload-partial status file, confirm ⚠ partial write — review
• Manually: click Start Agent with an overload-retry plan present, confirm the executor resumes that plan rather than picking a new todo plan

🤖 Generated with Claude Code