prep_wake_samples leaves stale clips when re-run into a populated output dir

[jamditis]

Found during PR #4 review (Codex, prep_wake_samples.py:218).

Problem

Re-running prep into an existing samples/positive or samples/negative directory after a recording is shortened, re-segmented, or replaced with fewer utterances overwrites only the current indices. Older higher-numbered WAVs from the previous run remain. Downstream training globs the whole directory, so the stale clips are silently included even though process() reports only the newly written count.

Why it matters

This is in the normal iterate-on-recordings workflow (re-record, re-run prep), so it bites during routine use, not just an edge case. The training set silently accumulates orphaned clips.

Approach

Before writing, clear prior outputs for the same class (with a confirmation or a --clean flag to avoid surprise deletion), or write to a fresh run-stamped subdir, or maintain a manifest and prune files not in the current run. Document the chosen behavior.

Acceptance criteria

• Re-running prep does not leave clips from a previous run in the output dir (or the behavior is explicit and documented).
• The reported count matches the files present after the run.
• A test covers the re-run-with-fewer-clips case.

[jamditis]

Second half of the fix opened as #86 (stacked on #85).

#85 gave --clean the stem match, which covers the case the issue leads with: re-record a take with fewer utterances, and its now-unused higher-numbered clips go. #86 covers the two orphans that match cannot structurally reach, both of which come out of the same iterate-on-recordings loop:

• an input dropped between runs, whose clips no rerun ever overwrites
• a basename collision re-run with one of the pair, stranding the disambiguated take-1_*.wav

Prep now records what it wrote in .prep-manifest.json and uses that record to reach them, gated so a mistyped --output still cannot delete another dataset: provenance says prep made a file, not that the file belongs to the set being refreshed.

Acceptance criteria against the two PRs together: re-running no longer leaves prior-run clips (criterion 1), the reported count matches what is on disk (criterion 2 — total is len(written), counted from files actually written, and the orphans that would inflate the directory past it are now removed), and the re-run-with-fewer-clips case is covered plus nine more (criterion 3).

wake-20260723T1705-0bf0ee