fix(pipeline): audit round 12 — F1–F3

F1 step1_transcribe (L313–325):
  When w1 is invalidated by a transcript sha256 mismatch, also delete
  ep{N}_w2aligned.json and ep{N}_anchor_offset.json if they exist.
  These step-2 caches store the anchor_offset calibrated against the old
  w1 timestamps.  Without this cleanup, a fresh w1 (generated from a
  corrected transcript) would be fed into a stale step-2 cache — the old
  anchor_offset applied to new w1 timestamps produces silently wrong
  alignment with no warning.  Matches the chunk-JSON invalidation pattern
  added in round 11.  Each deleted file is printed for operator visibility.
  The deletion uses an os.path.exists guard (no unconditional remove).

F2 process_episode (L1223):
  Added -> None return type annotation.  Was the only public function in
  the file without a return annotation.

F3 find_post_ad_anchor (L713):
  Narrowed return annotation from bare 'tuple | None' to
  'tuple[int, int] | None'.  The function always returns either None or
  (cand_si: int, anchor_wi: int).  This was the last bare tuple annotation
  after round 11 parameterised step2_align.

Author: Zenflow

scripts/podcast_to_moshi_dataset.py

@@ -310,6 +310,17 @@ def step1_transcribe(ep_num: int, mp3_path: str, transcript_path: str) -> tuple[
             ):
                 os.remove(stale_chunk)
                 print(f"  step1: staler Chunk-Cache gelöscht → {stale_chunk}", flush=True)
+            # Also delete step-2 caches: w2aligned.json and anchor_offset.json
+            # are derived from w1 timestamps.  If w1 is regenerated, these are
+            # stale — the anchor_offset calibrated against the old w1 would
+            # produce a misaligned result when applied to the new w1.
+            for stale_step2 in [
+                os.path.join(WHISPER_CACHE, f"ep{ep_num}_w2aligned.json"),
+                os.path.join(WHISPER_CACHE, f"ep{ep_num}_anchor_offset.json"),
+            ]:
+                if os.path.exists(stale_step2):
+                    os.remove(stale_step2)
+                    print(f"  step1: staler Step-2-Cache gelöscht → {stale_step2}", flush=True)
             # Fall through to re-transcription below.
         else:
             print(f"  step1: w1 gültig — Transkription übersprungen ({w1_path})", flush=True)
@@ -703,7 +714,7 @@ def last_w1_end_before(t_exp: float) -> float | None:
         idx = bisect.bisect_right(w1_starts, t_exp, lo, hi) - 1
         return w1_ends[idx] if idx >= lo else None
 
-    def find_post_ad_anchor(si_e: int, t_exp_end: float) -> tuple | None:
+    def find_post_ad_anchor(si_e: int, t_exp_end: float) -> tuple[int, int] | None:
         """
         Findet den ersten zuverlässig bestätigten Anchor-Punkt in w1 für den
         w0-Inhalt nach dem END-Marker si_e.
@@ -1213,7 +1224,7 @@ def parse_episode_filter(spec: str) -> set:
     return result
 
 
-def process_episode(ep_num: int, transcript_path: str, mp3_path: str):
+def process_episode(ep_num: int, transcript_path: str, mp3_path: str) -> None:
     print(f"\n{'═' * 60}", flush=True)
     print(f"  Folge {ep_num}  —  {os.path.basename(mp3_path)}", flush=True)
     print(f"{'─' * 60}", flush=True)