Node: report every radio heard on a call, not just the last (server-26#304)

P1 of the radio-ID plan. metadata_watcher records each transmission as
[radio, first seen, last seen] from channel rows and call_log grants in both
audio and console modes. Sightings of the same radio under 2 s apart are
merged, because the P0 probe found the ID in only 44% of polls during a
burst. call_end gains `radios: [{radio, start, end}]`.

main.py rebases the offsets into the trimmed audio file's frame, so
wall clock = audio_start_epoch + offset, the mapping C2 already uses. It
also sends the list on /upload (UploadMeta.radios), because C2 starts the
pipeline from /upload before call_end arrives.

srcaddr keeps its meaning (latest talker). Metadata only: no boundary,
timestamp or attribution change (drb-audio-review).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

drb-edge-node/app/internal/call_recorder.py

@@ -785,6 +785,7 @@ class CallRecorder:
         system_id: Optional[str] = None,
         audio_start_epoch: Optional[float] = None,
         audio_end_epoch: Optional[float] = None,
+        radios: Optional[str] = None,
     ) -> Optional[str]:
         """
         Upload once, inline. Returns the storage URL on success.
@@ -802,6 +803,7 @@ class CallRecorder:
         meta = UploadMeta(
             call_id=call_id, talkgroup_id=talkgroup_id, talkgroup_name=talkgroup_name,
             system_id=system_id, audio_start_epoch=audio_start_epoch, audio_end_epoch=audio_end_epoch,
+            radios=radios,
         )
         result = await self._post_inline(file_path, meta)
         self.spool.last_result = {"call_id": call_id, "kind": result.kind, "status": result.status,