Stop the DeepLC workers filling the log with blank lines

[RobbinBouwmeester]

Reported from the desktop interface: predicting retention times produces a run log that is almost entirely empty lines.

Cause

DeepLC's progress writer emits a bare \r per update, which renders as nothing when stdout is not a terminal — so each update arrives as an empty line. The engine inherits a sidecar's stdout rather than capturing it (sidecar.rs::run_worker, deliberately, so hours-long progress reaches the user live), so those lines land in the terminal, in any redirected log file, and in the desktop run log, pushing the real messages out of view.

Measured before the fix:

log	total lines	blank
2.9M-peptide library prediction	5,743	5,697 (99%)
entrapment A/B, baseline arm	5,810	5,722 (98%)
entrapment A/B, multi-head arm	5,824	5,734 (98%)

Not new, but it now affects every run that predicts retention times — which, since multi-head calibration became the default, is every run with a DeepLC interpreter.

No upstream switch exists

deeplc.predict takes a predict_kwargs passthrough, but neither enable_progress_bar nor verbose reaches the writer — both raise TypeError. So the filtering has to happen in the workers.

The fix

Both DeepLC workers wrap their prediction in quiet_deeplc_progress(), which replaces sys.stdout with a proxy that drops only segments that are empty once carriage returns and whitespace are stripped.

• anything DeepLC actually says still comes through, in order;
• the workers' own per-chunk progress is unaffected;
• stderr is not touched at all, so a traceback is never swallowed;
• \r counts as a terminator as well as \n, so a writer that never emits a newline cannot accumulate in the buffer for a whole run;
• MUMDIA_DEEPLC_RAW_OUTPUT=1 restores the unfiltered output for debugging the worker itself.

Verified on the real workers

	before	after
deeplc_worker.py, 6,000 peptides	13 lines, 12 blank	1 line, 0 blank
deeplc_finetune.py, re-prediction mode	—	8 lines, 0 blank, every message intact

The surviving output is exactly what you want to see:

no-finetune: predicting with the DeepLC base model (seed ignored)
predicting 3 unique standard peptidoforms with the DeepLC base model (torch threads=8)
  3/3  0.0s for this chunk (76 peptidoforms/s, ETA 0.0 min)
prediction phase: 0.0s total
wrote library with re-predicted iRT (the DeepLC base model): ...
  rows=3000 repredicted=3000 retained_imported=0 (...)

Tests

Three: both workers route prediction through the filter, the escape hatch is present, and the proxy keeps content and ordering while dropping blanks. pytest tests/python — 83 passed in the DeepLC environment, and the new static checks also pass on a bare interpreter with no ML dependencies.

The filter is duplicated in the two workers rather than shared, following the precedent _check_deeplc_version already sets for the same reason: the sidecars are standalone files invoked by path.

🤖 Generated with Claude Code