fix: increase max_auto_steps

[micpapal]

mas-ctl run-mas runs could silently truncate mid-turn on multi-agent workflows. Symptoms:

• The run exits with RUN_END status="success" and no exception/traceback.
• The last recorded event is a wm_append of tool_call: for a tool the model requested — immediately followed by turn_commit and the run ending. The tool call is never actually dispatched (no tool_call_start, no governance/contract check, no further LLM call).
• Non-deterministic: only reproduces on some runs, with no consistent trigger.

Root cause

KernelDriver._feed_bounded (runtime/src/mas/runtime/driver/driver.py) bounds a turn's internal processing loop with max_auto_steps (previously set to 64). If auto_steps reaches max_auto_steps before that next pass runs, the while loop simply exits with the item still sitting in the local queue. No exception is raised, nothing is added to trace.rejected_ingress, and the turn is reported as a normal success.

Fix

In runtime/src/mas/runtime/driver/driver.py:

1. max_auto_steps: 64 → 512.
2. Fail loudly instead of silently. If the step budget is exhausted with items still queued, the driver now appends a RaiseBoundaryError(code="STEP_BUDGET_EXHAUSTED", recoverable=False, ...) to trace.boundary_errors instead of returning a clean-looking trace.

[jordanauge]

lgtm