fix: carry source_split through validation, so split preservation actually runs

[Rebreda]

Split preservation shipped in 0.2.0 and never worked. My bug, and my verification was what let it through.

$ listenr build-dataset --manifest <mdc> --preserve-splits
ERROR: --preserve-splits was requested but no entry has a usable source_split.

on a manifest whose every record has source_split.

Cause

validate_entry rebuilds each record from a whitelist, and source_split was not on it. assign_splits runs on validated entries, so by the time the preserve logic looked, the field was always gone and the labelled count was always zero. Both branches were unreachable: --preserve-splits always errored, and the auto-detect default always fell through to reshuffling.

Reproduced on the first 50 records of the real manifest:

raw records with source_split: (50, 0)
after validate_entry:          (0, 50)

Why I did not catch it

My verification for #24 called assign_splits directly on raw manifest records, so it exercised a path no user could reach and printed a convincing result. The unit tests had the same gap: they passed dicts that still carried source_split. Everything agreed with everything else, and none of it went through the pipeline.

Fix

Carry the field through validation. Also write it to the CSV, so a built dataset records whether its splits are the corpus's own or a reshuffle. That distinction is the difference between a comparable result and a speaker-leaked one, and it was not previously recoverable from the output.

write_csv uses extrasaction="ignore" and the HF writer just gains a column that prepare_example already strips, so nothing downstream chokes.

Tests that would have caught it

Six end-to-end tests that write a manifest file to disk and run the real CLI, asserting the resulting split sizes. Against the 0.2.0 behaviour four of the six fail; with the fix all pass.

--- against shipped 0.2.0 behaviour ---
FAILED test_corpus_splits_survive_the_whole_pipeline
FAILED test_preserve_splits_flag_does_not_error_on_a_real_manifest
FAILED test_unlabelled_records_go_to_train
FAILED test_source_split_is_recorded_in_the_output
4 failed, 2 passed
--- fix restored ---
6 passed

Verification against the real corpus, through the full pipeline this time

validated 2194   corpus-labelled 2104   unlabelled 90
preserved -> train 1472, dev 385, test 337

dev and test match the corpus exactly. The 90 unlabelled go to train, so dev and test keep only the corpus's own speakers.

364 tests pass on 3.13, 342 passed and 12 skipped on 3.11.

Wants a 0.2.1 release once merged, since the headline fix of 0.2.0 is inert without it.