harness-followup(mofa-podcast): preserve per-segment WAVs on success (#59)

* harness-followup(mofa-podcast): preserve per-segment WAVs on success

Refactor `SegmentDirCleanup` from "always wipe on Drop" to
"wipe-on-failure-only" so the octos harness can run a future
`PerFileNonSilent` validator against `<output_dir>/segments/seg_*.wav`
after `generate_podcast` returns. A TTS dropout on one segment is
otherwise padded to silence and slips past the whole-file
`AudioNonSilent` check, leaving the user with a mid-conversation
dropout.

Contract:
- Success path (Ok return): `mark_success()` flips an AtomicBool on
  the guard; Drop becomes a no-op and dialogue/placeholder WAVs
  survive.
- Failure path (Err return, panic, SIGTERM cancel): flag stays false
  and Drop wipes the scratch directory — matches the prior behaviour
  so failed runs don't leak partial state.

Filename shape is unchanged (`seg_{NNN}_{voice}.wav` for dialogue;
`pause_after_*`, `pause_line_*`, `bgm_placeholder_line_*` for the
intentionally-silent placeholders) so the future octos validator can
glob `**/segments/seg_*.wav` and skip placeholders.

Tests: +5 unit / integration-style tests around the cleanup contract
(default-wipe, mark-success preserve, idempotency, missing-dir
graceful, full happy-path file shape). Also stabilises three
pre-existing flaky tests that shared a second-resolution timestamp
with each other and could collide under parallel `cargo test`.

Manifest bumped 0.4.5 -> 0.4.6 so the octos harness can key
`PerFileNonSilent` on it once the octos-side PR lands.

cargo check / cargo clippy -- -D warnings / cargo test all green
(matches CI). The `--all-targets` clippy errors are pre-existing on
origin/main and out of scope.

* mofa-podcast: wipe stale segments at run start; correct cancel docs

Address codex review P1s on #59:

P1.1 — Stale-segment reuse (functional regression introduced by this
PR's preserve-on-success behaviour). Previously the always-wipe Drop
guaranteed `<output_dir>/segments/` started empty. Now that the
success path preserves files, a subsequent run reusing the same
output directory could pick up a stale `seg_NNN_voice.wav` from a
prior successful run via the assembly loop's `seg_path.exists()`
check (line ~1640) when its own TTS call fails for that slot.
Failure would be silently masked: errors are recorded but the
consistency check `assembled_dialogue_segments != dialogue_count`
passes against stale audio.

Fix: clear `<output_dir>/segments/` at the start of every
`generate_podcast` invocation, before constructing the guard.
Combined with the in-run failure-path Drop, the directory only ever
holds files from the current run on disk after return.

P1.2 — Documentation inaccuracy. Both `check_cancel()` and the
SIGTERM handler call `std::process::exit(130)`, which bypasses RAII
Drop. The PR's comments overclaimed that Drop runs on SIGTERM —
corrected to explicitly note partial state may survive a cancel
until the next run wipes on entry (long-standing pre-existing
behaviour, out of scope to refactor here).

Test coverage:
- +1 unit/integration: `generate_podcast_wipes_segments_directory_on_error_return`
  plants a stale file, calls `generate_podcast` with an Err path, and
  asserts both the start-of-run wipe and Err-path Drop removed it.

Integration shell: `scripts/test-integration.sh` 5.5 inverted —
previously asserted `segments/` was cleaned after success; now
asserts `seg_*.wav` are preserved (matching the harness contract).

cargo check / clippy / test / fmt all green; 54 tests (was 53 in
the first commit on this branch).

Author: ymote

mofa-podcast/Cargo.lock

@@ -1,6 +1,6 @@
 [package]
 name = "mofa-podcast"
-version = "0.4.5"
+version = "0.4.6"
 edition = "2021"
 description = "Multi-speaker podcast generation with TTS voice synthesis and audio assembly"
 authors = ["hagency"]

mofa-podcast/Cargo.toml

@@ -1,6 +1,6 @@
 {
   "name": "mofa-podcast",
-  "version": "0.4.5",
+  "version": "0.4.6",
   "author": "hagency",
   "description": "Multi-speaker podcast generation with TTS voice synthesis and audio assembly",
   "timeout_secs": 1800,

mofa-podcast/manifest.json

@@ -226,11 +226,15 @@ if $OMINIX_AVAILABLE; then
         skip "5.4 Audio duration check" "ffprobe not available"
     fi
 
-    # 5.5 Segments cleanup
-    if [[ ! -d "$TMPDIR/output1/segments" ]]; then
-        pass "5.5 Segment files cleaned up after concat"
+    # 5.5 Segments preserved on success (harness PerFileNonSilent contract).
+    # The success path now keeps `<output_dir>/segments/seg_*.wav` on disk so
+    # the octos harness can validate each dialogue segment is non-silent.
+    if [[ -d "$TMPDIR/output1/segments" ]] && \
+       compgen -G "$TMPDIR/output1/segments/seg_*.wav" >/dev/null; then
+        pass "5.5 Per-segment WAVs preserved after success (for PerFileNonSilent)"
     else
-        fail "5.5 Segment files cleaned up after concat" "segments dir still exists"
+        fail "5.5 Per-segment WAVs preserved after success (for PerFileNonSilent)" \
+            "segments/ missing or has no seg_*.wav files"
     fi
 
     # 5.6 Stderr shows correct phase ordering