Fixes #6348.

Problem

On a channel bot with a coalesce window (message_coalesce_window_ms > 0), when an inbound batch contained an image together with other media (e.g. a photo + a voice note within the debounce window), the non-image media reached the agent as a bare text placeholder — [Voice message (Ns): url], [File (name): url] — instead of its transcript / extracted text.

The gap was narrow but real:

This is the same class of divergence #6321 / #6323 fixed for images — but only images were made path-agnostic. The debounce ingest pre-downloaded only images, and the debounced flush only ran prepend_image_descriptions; audio transcription and document extraction lived solely in dispatch_message's per-type arms and were never reached for a mixed batch.

Fix

Generalize the #6321 approach to all media types by funnelling both inbound paths through two shared helpers, so they cannot drift out of parity again:

• download_media_blocks(handle, adapter, content) -> Option<DownloadedMedia> — downloads any attachment (image / file / voice / audio / video) to disk and builds its blocks (caption / context headers, the ImageFile or [File: …] saved to … marker, and any document extraction already baked into download_file_to_blocks). It is I/O-bound only, so it is safe to run inline on the debounce ingest loop exactly as the original image-only pre-download was. It returns None for non-media content and for a failed download, so the caller falls back to the text placeholder — matching the immediate path's fall-through.
• enrich_media(handle, media) -> Vec<ContentBlock> — applies the deferred, LLM-backed enrichment recorded on the payload (image description via prepend_image_descriptions; audio transcription via maybe_transcribe_inbound_audio, inserted right after the context header) in the dispatch task, off the ingest loop.

Wiring:

• The debounce ingest now pre-downloads every media type (not just images) while the message's individual content is still available — after coalescing the debouncer only retains a merged text placeholder.
• The debounced flush enriches each coalesced attachment with the same enrich_media the immediate path uses, concatenates in arrival order, and prepends the coalesced text (which keeps media captions inside the sanitizer-checked text).
• dispatch_message's five per-type media arms collapse into one download_media_blocks + enrich_media + dispatch_with_blocks call, preserving their exact output: context headers, transcription inserted after the header (index 1), and — deliberately — no image description for image-typed documents (matching the old File arm).

Net effect: a mixed photo+voice+pdf burst reaches the agent as one turn with every attachment turned into text, identically whether or not it was coalesced — which also benefits text-only models.

Verification

• cargo check -p librefang-channels --lib — clean (no warnings).
• New enrich_media unit tests next to the existing maybe_describe_inbound_image / maybe_transcribe_inbound_audio fakes:
  • enrich_media_describe_image_describes_each_imagefile — DescribeImage inserts the description before the ImageFile.
  • enrich_media_none_leaves_blocks_and_skips_vision — None returns blocks verbatim with no kernel call (an image-typed document is not described).
  • enrich_media_transcribe_audio_inserts_after_header — TranscribeAudio inserts [Transcription: …] between the [Voice message …] header and the saved-path block (same order the immediate path's transcription_block_lands_between_header_and_file_path test pins).
  • enrich_media_transcribe_audio_noop_when_disabled — Ok(None) (feature off) leaves the [header, saved-path] pair unchanged.
• Existing debouncer drain tests updated for the new Vec<DownloadedMedia> return shape.
• clippy + full test suite (including the integration shards) run in CI.

Live LLM round-trip verification (a real coalesced photo+voice burst against a running daemon) is left for a maintainer per the repo's human-only integration-testing policy.