Add opt-in Agent Mode action publishing (GA-AUD)

[Chance-Konstruktion]

Summary

Implements GA-AUD — the Home Assistant side of Agent Mode. The integration can now publish controller actions to the executor shipped in #92 (GA-109), so Agent Mode works end-to-end.

When enabled, each analyzed frame additionally:

1. asks the LLM for one controller action (PromptBuilder.build_action),
2. validates it against the action schema + a button whitelist (parse_action),
3. publishes it to gaming_assistant/{client_id}/action for the executor.

The normal tip pipeline is unchanged — this is an additional inference per frame (the chosen "tip + action" behavior).

Safety model

• Opt-in master switch CONF_AGENT_MODE (default False), exposed via a new gaming_assistant.set_agent_mode service and an Agent Mode switch entity.
• Runtime-only / resets to OFF on every restart — the coordinator is built from entry.data only, so I deliberately did not persist agent mode. The AI never controls inputs unless deliberately re-enabled.
• Isolated — action generation runs behind asyncio.wait_for + a broad try/except, so a bad model reply / timeout can never disrupt the tip pipeline.
• Whitelist enforced both in the prompt and in parse_action; no_op/invalid output is dropped (never published).
• The executor still enforces its own whitelist, --dry-run, and emergency stop. Recommended first run: executor in --dry-run.

Changes

• const: CONF_AGENT_MODE, DEFAULT_AGENT_MODE, MQTT_ACTION_TOPIC, AGENT_VALID_BUTTONS.
• coordinator: agent_mode flag + set_agent_mode() setter, async_publish_action(), and the isolated _maybe_publish_agent_action() hook after a successful tip.
• image_processor: generate_action() — build_action → LLM → parse_action; returns None on invalid/no_op output, never raises.
• service set_agent_mode (+ services.yaml), Agent Mode switch entity, and agent_mode strings in EN/DE/ES/FR/IT.
• tests: tests/test_agent_mode.py — behavioural tests for generate_action + static wiring contracts (the coordinator can't be imported without a real HA env, mirroring test_coordinator.py).
• Docs: CHANGELOG, README (enable example + safety note), ROADMAP (GA-AUD → 🟡, per-action confirmation still open).

Test plan

• python -m unittest discover -s tests → 313 tests, all green (+15 new).
• CI coverage gate locally: pytest --cov=custom_components/gaming_assistant --cov-fail-under=45 → 48.30%.
• py_compile + JSON validation for all touched files.
• Live end-to-end on real hardware (vgamepad/ViGEmBus) — not possible in CI.

Follow-up

Per-action confirmation UI (the remaining half of GA-AUD) is intentionally out of scope.

https://claude.ai/code/session_01VaYnzPqHF5kxB3stM73g5i

---

Generated by Claude Code