Fix unnecessary deepcopy in validate_psm_list causing OOM on large PSM lists

[ArthurDeclercq]

Problem

validate_psm_list() unconditionally called copy.deepcopy() on the filtered PSMList on every call to predict(), calibrate(), and predict_and_calibrate(). This was flagged by a # TODO: Is deepcopy really necessary or can it be avoided? comment at the time it was introduced (d7a016a), but never resolved.

This surfaced as an OOM crash when running IM2Deep through MS²Rescore on a large timsTOF immunopeptidomics dataset (PSM list expanded by Mumble's mass-shift candidate search). The process was killed by earlyoom at ~78 GB RSS after several minutes of steadily ramping memory, while the equivalent DeepLC step on the same PSM list completed quickly.

Investigation

n PSMs	deepcopy time	extra RSS
50,000	3.0 s	+251 MB
200,000	12.7 s	+1,006 MB
446,902	27.5 s	+2,094 MB

Roughly linear, ~4.7 KB/PSM. On a Mumble-exploded PSM list in the millions, this alone accounts for tens of GB and the multi-minute "slow ramp" observed before the kill.

• Confirmed deeplc._parse_psms (DeepLC's equivalent entry point) does no copying at all when given a PSMList — this was the actual asymmetry, not the shared feature-encoding path.

Fix

The deepcopy is only structurally necessary because the needs_target=True branch mutates PSMs in place (psm.metadata["CCS"] = ...), and boolean-mask PSMList indexing shares PSM instances with the input rather than cloning them (verified directly). For the plain predict() path (needs_target=False, what MS²Rescore's feature generator uses), nothing downstream mutates any PSM, so the copy was pure overhead.

Made the deepcopy conditional on needs_target, matching what it actually protects against.

Caveat for reviewers

predict_and_calibrate() calls validate_psm_list(psm_list) with the default needs_target=False for its primary psm_list argument, then later does:

for idx, psm in enumerate(psm_list):
    ...
    psm.metadata["predicted_CCS_uncalibrated"] = predicted_ccs[idx]

Previously this ran against an isolated deep copy, so the mutation never reached the caller's original object. With this fix, predict_and_calibrate() will now mutate the caller-supplied psm_list in place (attach predicted_CCS_uncalibrated to metadata). This has no impact on MS²Rescore (it only calls predict(), never predict_and_calibrate()), but it is a visible behavior change for any other direct consumer of this function. Flagging so you can decide whether that's acceptable or whether predict_and_calibrate() needs its own explicit copy of psm_list restored.