Harden mlx-lm capability probing and argument classification

[acari-git]

Problem

The mlx-lm capability probe is functional but still has several edge cases from the original audit.

Current verification (2026-08-15)

Already fixed:

• Cache invalidates when the executable at the same path changes.
• Unavailable probes are retried instead of cached.
• Rapid-MLX does not use the mlx-lm probe.
• Repeated Start requests are guarded.

Still open:

• Concurrent requests for the same executable can start duplicate --help probes because the actor does not share an in-flight task.
• requestedOptionalFlags derives flags from the flattened argv; a value beginning with -- can be misclassified as a flag.
• Help parsing accepts any --word in output and does not require canonical flags such as --model, --host, and --port, so diagnostic prose can cause false positives.
• Capability probing runs before the fresh port check, so an occupied port can still pay the probe delay.

Desired behavior

• Coalesce concurrent probes per executable identity.
• Classify flags from structured launch-option sources rather than arbitrary argv values.
• Treat help output as valid only when expected core server flags are present.
• Perform the fresh port check before an optional runtime capability probe where that preserves existing safety semantics.
• Add regression tests for all four cases.