fix: clear recovery state on requeue_task

[nikitaxru]

Summary

• requeue_task (supervisor MCP tool) removed the task from the dispatched set and approved preflight, but never cleared the recovery failure counter
• Tasks that hit max retries (retryable: false) could never be re-dispatched even after the underlying issue (e.g. missing repo config) was fixed
• Now calls recovery.clear() to reset the failure counter, allowing the task to be picked up on the next poll

Context

QR-273 failed 3 times with "No repos configured" (config issue since fixed). After requeue_task, the orchestrator still refused to dispatch it because recovery state had retryable: false.

Test plan

• Verify requeue_task clears recovery state by requeueing a previously-failed task
• Confirm task gets dispatched on next poll cycle

🤖 Generated with Claude Code

[cursor]

PR Summary

Low Risk
Low risk: a small, localized change to the supervisor requeue_task tool to reset retry state, plus accompanying tests; no changes to core dispatch logic beyond passing a new callback through constructors.

Overview
Fixes supervisor requeue_task so re-queued tasks can be dispatched again by clearing the per-task recovery/failure state (in addition to removing the task from the dispatched set and re-approving preflight).

Threads a new clear_recovery_callback from Orchestrator → SupervisorChatManager/build_supervisor_server, and adds focused tests covering both the callback-present and callback-absent paths. Also significantly condenses/updates CLAUDE.md guidance (style, quality, and testing rules) without runtime impact.

^{Written by Cursor Bugbot for commit 77bd03c. This will update automatically on new commits. Configure here.}