fix(forge): resolve abort race condition in agentic loop

[Mog9]

The Problem

The agentic loop's abort mechanism had a race condition that made the abort button silently fail. When a user tried to cancel a running kernel optimization, the orchestrator would overwrite the abort flag, preventing the run from stopping.

Before:

• User clicks abort → API sets abort_requested=true on disk
• Orchestrator finishes tool execution → calls save_state() with in-memory state
• In-memory state has abort_requested=false → overwrites disk state
• User's abort request is lost → run continues burning API credits

After:

• User clicks abort → API sets abort_requested=true on disk
• Orchestrator calls _sync_abort_flag() before every save_state()
• Sync function reads disk state → merges abort flag into memory
• Abort flag is preserved → run stops correctly

The Fix

Added _sync_abort_flag() function in app/services/forge/agent_runner.py that reads the abort flag from disk and merges it into the in-memory state before every save_state() call. This ensures the API's abort request survives the orchestrator's state saves.

    """Merge abort flag from disk into in-memory state before save_state()."""
    disk_state = load_state(run, repo_root)
    if disk_state and disk_state.abort_requested:
        state.abort_requested = True

Files Changed

• app/services/forge/agent_runner.py — Added _sync_abort_flag() and called it before every save_state() in the main loop
• tests/test_abort_race_condition.py — New file with 6 tests covering the fix