fix: keep intent classification available under worker load

[giscebot]

Summary

• preserve actionable work intent while normalizing classifier write permissions
• report intent-classifier subprocess timeouts without persisting a truncated command and prompt
• document the OpenClaw concurrency headroom required by enqueue-time classification and feedback learning
• recommend agents.defaults.maxConcurrent=8 for the standard four-worker bridge deployment

Why

Intent-classifier calls share OpenClaw's global concurrency with long-running bridge jobs. With four bridge workers and only four OpenClaw slots, classifier calls can remain queued until they time out and silently fall back to the deterministic parser. Operators need an explicit sizing invariant: OpenClaw concurrency must exceed bridge worker concurrency.

Validation

• pytest -q tests/test_intent_classifier.py tests/test_queue.py — 30 passed
• git diff --check main...HEAD
• production configuration validated with openclaw config validate

[giscebot]

Superseded by #213, rebuilt from current main. This PR reused the old #149 branch and therefore contained stale history and conflicts.