Make speaker rename actually land in every transcript

[execsumo]

The popup save and the Speakers settings tab both relied on a single
directory scan to rewrite Speaker_<hex> placeholders into real names,
and the popup's "specific transcript" fallback only ran when the
transcript's directory differed from the configured output dir. That
made the rename quietly skip the file whenever a URL-comparison quirk
(symlink resolution, trailing-slash normalization, etc.) made the two
URLs compare unequal in one direction but equal in the other — leaving
Speaker_ABC123 in the saved .md even after the user named the speaker.

• Centralize the rewrite in rewriteSpeakerAcrossTranscripts, used by
  the popup save, the Speakers tab inline rename, and the merge action.
  In addition to scanning the output directory, also walk every
  transcriptPath tracked by the pipeline queue so a moved output
  folder doesn't strand old transcripts.
• In saveSpeakerName, always rewrite candidate.transcriptPath
  unconditionally instead of only when the directory diverges from
  outputDir. The redundant pass is harmless thanks to the new no-op
  short-circuit below.
• In TranscriptWriter.renameSpeaker, skip the disk write when the
  rebuilt content matches what's already on disk. The defensive scans
  now hit many transcripts that don't contain the placeholder, and
  rewriting unchanged files just churns the FS.
• Add tests that pin down the production Speaker_<hex> placeholder
  format (the existing tests only covered "Speaker 7" with a space) and
  verify the file-targeted rename + idempotency.