Bug: Implementation node retry loop fills /tmp with workspace directories on persistent failure

[JoshSalomon]

Bug

When the implementation node encounters a persistent error (e.g., trying to checkout a non-existent fork branch), it retries in a tight loop without:

1. Cleaning up failed workspace directories — each retry creates a new /tmp/forge-{TICKET}-{random}/ directory
2. Backoff or delay — retries happen as fast as the loop can run
3. Retry cap — the loop continues until the disk fills up or an external limit is hit

Observed behavior

Setting fork_owner/fork_repo in the checkpoint to values where the branch doesn't exist on the fork causes:

git checkout -b forge/aisos-2119 fork/forge/aisos-2119
fatal: 'fork/forge/aisos-2119' is not a commit and a branch 'forge/aisos-2119' cannot be created from it

The error is non-recoverable (the branch will never appear), but the node retries immediately. In ~30 seconds this created 50,230 workspace directories in /tmp, consuming ~13GB and filling the tmpfs.

Root cause

In src/forge/workflow/nodes/implementation.py, the error handler returns:

return {
    **state,
    "last_error": str(exc),
    "current_node": implementation_node,
}

This routes back to the same node, which the graph re-enters immediately. The prepare_workspace() call creates a new directory on each attempt but never cleans up on failure.

Expected behavior

• Failed workspace directories should be cleaned up before retry
• Exponential backoff or a retry limit should prevent tight loops
• After N consecutive failures (e.g., 3), the node should escalate to blocked status instead of retrying indefinitely

Reproduction

1. Set fork_owner and fork_repo in a ticket's checkpoint to valid values
2. Delete the corresponding branch from the fork
3. Trigger the implementation node
4. Watch /tmp fill with workspace directories

Environment

• Forge commit: main branch (observed on forge/aisos-741 feature branch)
• OS: Fedora 43, tmpfs at /tmp