fix: vision/audio argv builders silently drop or duplicate transform flags

[musicalplatypus]

Summary

Two independent bugs in the vision/audio ai_modules' train/test argv builders — both variations of the same underlying defect: the train-time and test-time argv construction for the same parameter disagreeing with each other.

image_base.py: train/test disagree on data_proc_transforms's wire format

Train argv rendered it through an f-string (stringified list, e.g. "['BINARIZE']"), while test argv passed the real list object.

prepare_transforms() (tinyml-tinyverse's train_base.py) only combines data_proc_transforms into args.transforms when isinstance(args.data_proc_transforms, list) is true, with no else branch — so the stringified form silently skipped it during training while testing still applied it. Pure train/test skew for anyone customizing this parameter, with no error or warning.

timeseries_base.py already passes the raw list on both sides correctly; image_base.py's train side now matches it. feat_ext_transform/augmentation_transform don't have this defect — they're parsed via _normalize_transform_list's literal_eval fallback, which tolerates both string and list forms — and were left untouched (not a real bug, no need to change working code).

audio_base.py: --data-proc-transforms/--feat-ext-transform declared twice

Both the train and test argv builders declared these two flags twice: once correctly as a raw list, and again later as a stringified f-string. argparse's last-occurrence-wins semantics meant the later, stringified declaration always won — making the earlier, correct declaration dead code. Same underlying defect as above, just duplicated instead of split across two methods.

Removed the dead duplicate declarations in both _build_common_train_argv and _build_common_test_argv, keeping the single correct raw-list version each already had.

Verification

Unit tests against the argv-building methods directly (constructed via unittest.mock.MagicMock for self — these methods only read self.params, no full ModelRunner instance needed). All 4 tests fail on the unmodified code with the exact symptoms described above (stringified list where a real list was expected; --data-proc-transforms appearing twice in the argv) and pass after the fix.

🤖 Generated with Claude Code

[musicalplatypus]

An independent peer review caught that this PR's own fix broke every image-classification training run. Pushed a follow-up commit (72196d6):

Making --data-proc-transforms a raw list (correctly, per the original fix) exposed a latent issue: prepare_transforms() in tinyml-tinyverse does args.data_proc_transforms + args.feat_ext_transform whenever data_proc_transforms is a list. --feat-ext-transform (and --augmentation-transform) were still stringified in the train argv builder, so this became list + str, raising TypeError: can only concatenate list (not "str") to list on every training run -- not conditional on any user setting, since the default data_proc_transforms=[] still enters that branch.

My earlier comment claiming feat_ext_transform "tolerates both string and list forms via _normalize_transform_list" was true of a different consumer (image_dataset.py's Dataset construction, later in the pipeline) but not of prepare_transforms(), which reads it directly with no such tolerance and runs first.

Fixed by passing both flags as raw lists too, matching _build_common_test_argv (already correct) and timeseries_base.py.

Added a test that drives the real prepare_transforms() against the argv builder's actual output rather than only checking argv shape -- confirmed to fail pre-fix with the exact TypeError above and pass post-fix.