Fix crash when --omit_probs is combined with 3Di masking

[gbouras13]

Fixes the --omit_probs crash that threw away 20 minutes of MI250X compute on 161,473 sequences.

The defect

--omit_probs sets output_probs=False, which makes the inference engine store None in place of the per-residue probability array (inference.py:174-178). write_predictions() then indexed all_prob[0] unconditionally and died with

TypeError: 'NoneType' object is not subscriptable

at write time — after the entire GPU prediction had already completed.

The masking loop also ran when masking was disabled: mask_threshold=0 makes the comparison all-False, so the loop did nothing but still required the array.

Worth noting: --mask_threshold defaults to 25 at the CLI (only the function signature defaults to 0), so plain phold predict --omit_probs with no --mask_threshold at all was also a guaranteed crash — not just the --mask_threshold 0 invocation from the report.

The fix

Fail fast, plus defensive guards, since they cover different callers:

• validate_mask_options() in phold/utils/validation.py, called in run, predict and proteins-predict immediately after begin_phold — before the database check and before ProstT5 is loaded. Verified end to end: proteins-predict --omit_probs now exits in ~1s with a message naming both escape routes, and --omit_probs --mask_threshold 0 proceeds normally.
• write_predictions() skips the masking loop entirely when mask_threshold is 0, and warns-and-continues rather than crashing when all_prob is None (library callers of get_embeddings bypass the CLI check).
• subcommand_predict() gets the same guard on the amino-acid masking path. Passing None to mask_low_confidence_aa reaches np.asarray(None, dtype=np.float64) — a silent array(nan) on numpy 1.x, a TypeError on numpy 2.x that the (KeyError, IndexError) handler doesn't catch. Proteins whose prediction genuinely failed still become "X" * len as before; a merely absent all_prob leaves the sequence unmasked rather than blanking a good prediction.

Not affected

predict_3di_12st.py hardcodes output_probs=True and doesn't pass mask_threshold. write_probs() only touches all_probs when output_path_all is non-None, which is itself gated on output_probs.

Testing

tests/unit/test_predict_3di_masking.py — 13 tests covering mask_threshold=0, all_prob=None, the mixed case, proteins-mode headers, the zero-length drop (#47), masking when probabilities are present, and the CLI validator. Reverting only predict_3Di.py makes 5 of them fail with the exact TypeError from the traceback.

Full unit suite: 95 passed, 18 skipped. Ruff: new file clean, src/phold/ count unchanged.