fix(ambiance-agent): prevent LLM request floods (250+ req/session)

[Copilot]

The AmbianceAgent called the LLM on every UI frame while Execute was checked and state was READY. Errors returned the agent to READY instantly with no backoff, creating unbounded request loops.

Changes

• Error cooldown: LLM errors now enter COOLDOWN (15 s) instead of transitioning directly to READY, preventing tight retry loops on repeated failures
• Input deduplication: MD5 hash of prompt + aggregated_inputs is stored on each dispatch; if inputs haven't changed the LLM call is skipped entirely
• One-shot execute: The "Execute Agent" checkbox is automatically unchecked once a request is dispatched — re-enabling requires explicit user action
• Cooldown duration hygiene: _cooldown_s is reset to the success default (30 s) on every READY transition, so a preceding error cooldown cannot bleed into subsequent runs

# Before — on error:
self._state = 'READY'   # triggers a new call next frame

# After — on error:
self._state = 'COOLDOWN'
self._cooldown_start = time.time()
self._cooldown_s = self._error_cooldown_s  # 15 s