fix: exit non-zero when all slides fail in aggregate_slide_features_batch

[raylim]

Summary

Fixes a silent failure mode where extract_features (and TITAN in particular) would exit with code 0 even when all slides failed during batch aggregation.

Changes

• mussel/utils/feature_extract.py: Raise RuntimeError in aggregate_slide_features_batch when all slides fail (both the method-level and model-level aggregation paths).
• mussel/cli/extract_features.py: Add a defense-in-depth check after aggregate_slide_features_batch returns — raise RuntimeError if no output .features.pt files were created.

Motivation

When TITAN (or other models) hits a CUDA OOM or similar error on every slide, the batch loop catches per-slide exceptions, logs warnings, and continues — returning without writing any outputs but still exiting 0. This silently breaks downstream pipelines that depend on the exit code.

Testing

• Verified the new RuntimeError is raised when the failed-slides count equals total slides.
• Existing partial-failure behavior (some slides succeed, some fail) is unchanged — a warning is logged but no exception is raised.

[raylim]

The core intent is right and the fix is needed. Two instances of the same 0 == 0 false-positive exist — one in each of the two new guards.

---

The problem — both guards can trigger on empty input

The all-slides-fail condition len(failed_slides) == num_slides is mathematically true when both are zero. The two locations handle this inconsistently:

Location	Placement	Empty-list safe?
feature_extract.py non-model path (~line 1457)	inside if failed_slides:	✅ yes
feature_extract.py model path (~line 1688)	outside if failed_slides:	❌ no
extract_features.py defense check (~line 322)	bare len(missing) == len(output_pt_paths)	❌ no

For the model path, an empty slide list causes failed_slides = [], num_slides = 0, the if failed_slides: block is skipped (correctly logs "All 0 slides processed successfully"), and then 0 == 0 fires a RuntimeError("All 0 slide(s) failed…").

For the defense-in-depth check, the same thing: output_pt_paths = [] → missing = [] → 0 == 0 → spurious raise. This path is reachable in the non-model case because aggregate_slide_features_batch returns normally for an empty non-model input.

Both are triggered by an empty --slide-paths list reaching _main_batch. That's an operator mistake, but the error message "All 0 slides failed" is actively misleading and hides the real problem.

Fixes

# feature_extract.py model path — add the same guard the non-model path already has
if failed_slides and len(failed_slides) == num_slides:
    raise RuntimeError(…)

# extract_features.py defense check
if output_pt_paths and len(missing) == len(output_pt_paths):
    raise RuntimeError(…)

---

No other issues — the RuntimeError placement for the all-slides-fail case is correct for both non-empty paths; the model cleanup before the check is fine (happens in the same function scope); and the CLI exception propagation to a non-zero exit is handled by the existing top-level error handler.