fix(miles-pipeline): narrow startup-validation import guard to ModuleNotFoundError

[TianyeGGBond]

Context

MilesPipeline._validate_topology() runs the F10 startup topology validation before any GPU allocation, so a misconfigured run fails fast with a clear message instead of OOM-ing mid-init.

It imports the validator lazily and wraps the import in a guard so a machine without miles installed (test/dev fixtures) can still construct the pipeline:

try:
    from miles.utils.rlix_validation import assert_rlix_topology
except Exception as exc:  # noqa: BLE001
    logger.warning("... unavailable; skipping startup validation ...")
    return

Problem

except Exception is too broad. It catches any failure while importing the validator — not just "miles not installed" (ModuleNotFoundError), but also SyntaxError, AttributeError, or a broken transitive import inside rlix_validation itself (ImportError). In all those cases the code logs a warning and silently skips every C1–C23 check, which defeats the whole point of the fail-fast guard: a real bug in the validator would let a misconfigured topology through to a mid-init OOM.

In production miles is always importable, so the only legitimate skip reason is ModuleNotFoundError (package genuinely absent). Anything else is a real bug and should propagate. This also matches the existing narrow pattern in coordinator.py (except ModuleNotFoundError on dynamic pipeline_cls import).

Change

• Narrow the guard from except Exception to except ModuleNotFoundError.
• Drop the now-unnecessary # noqa: BLE001 (no longer a blind except).
• Tighten the log message.

ImportError / SyntaxError now propagate and fail fast; the miles_args is None test-fixture path (a few lines up) is unaffected.

🤖 Generated with Claude Code

[JunzheJoe]

This is the right boundary for the F10 guard. except Exception swallowed every failure mode of the validation import — including a syntactically-broken or partially-installed miles — turning the whole C-check chain into best-effort. ModuleNotFoundError keeps exactly the legitimate skip (miles absent in test/dev fixtures, which is how the repo's own tests run) while a name-level failure (from ... import assert_rlix_topology with the module present) raises plain ImportError and now fails fast, as does any SyntaxError.

One boundary case worth knowing (fine as-is): a missing transitive dep inside rlix_validation's import chain also surfaces as ModuleNotFoundError and takes the skip path — that's the partial-dev-env case and matches the documented test-fixture intent. In production miles + deps are installed, so the guard only ever skips when miles genuinely isn't there. Closes the F10-FAILOPEN review finding. LGTM.