fix: stamp last_active before LLM call to prevent mid-iteration heartbeat timeouts

[Fail-Safe]

Problem

Agents running slow local models (e.g. 27B quantised MLX models running on-device) can take 3–4+ minutes per loop iteration. The heartbeat monitor's default inactive_secs timeout is 180s. Because last_active was only stamped at the end of an iteration — after the LLM call returns — the heartbeat would fire mid-call and flag the agent as unresponsive, triggering crash/recovery while the loop was still running correctly.

Root Cause

last_active updates happen as a side-effect of state/data mutations in AgentRegistry (e.g. set_state, update_model). There was no mechanism for the agent loop to signal liveness during a long-running LLM call — the only updates were at iteration boundaries, not within one.

Fix

Add a lightweight touch_agent callback through the existing KernelHandle trait so the agent loop can refresh last_active without mutating any real state:

1. AgentRegistry::touch(id) — updates last_active only, no other side-effects. Silently ignores unknown IDs (non-blocking).
2. KernelHandle::touch_agent(&str) — new trait method with a default no-op, so all existing mock/test implementations require zero changes.
3. OpenFangKernel::touch_agent — parses the UUID string and delegates to registry.touch().
4. agent_loop — calls kernel.touch_agent(agent_id) at the top of each iteration, immediately before call_with_retry, resetting the inactivity clock at the start of every LLM call rather than only at completion.

Impact

• Remote API providers (OpenAI, Anthropic, Groq) are rarely affected since calls complete in seconds; this is a no-op overhead for them.
• Local model users (mlx-lm, llama.cpp via OpenAI-compat, Ollama) benefit directly: a 4-minute 27B model call no longer triggers false crash detection.
• The KernelHandle default no-op means downstream forks/embedders that implement the trait don't need any changes.

Testing

Verified on a local 27B 4-bit MLX model. Before the fix, agents were killed and restarted mid-loop. After the fix, inactive_secs resets to ~0 at the start of each iteration and the loop completes without interruption.

• cargo build --workspace --lib passes
• cargo clippy --workspace --all-targets -- -D warnings passes (zero warnings)
• cargo test --workspace passes
• Live integration tested with slow local LLM (27B MLX model, ~3–4 min/iteration)

[jaberjaber23]

Real bug, clean fix, not slop. But needs tests before merge.

Stamping last_active before LLM call correctly prevents false Crashed state on slow local models. The fix only fires at loop iteration start so it won't mask genuinely stuck agents.

Blocking issue: no tests. AgentRegistry::touch() is trivially testable — register agent, set last_active to 5 min ago, call touch(), verify last_active is now recent. The project has 2074+ tests; this should meet that standard.