Summary

Fixes #6321. Inbound channel images were only described on the immediate dispatch path; the debounced (message-coalesce) path never ran the description step, so text-only models lost vision on coalesced bots.

The kernel's image-description step (maybe_describe_inbound_image, gated by [media] image_description) was reachable only from dispatch_message (the message_coalesce_window_ms == 0 fast path). The debounced path (message_coalesce_window_ms > 0) assembles its image blocks in flush_debounced and calls the shared dispatch_with_blocks tail directly, which contains no description call. So [media] image_description was only ever consulted when coalescing was off.

On a text-only model the un-described ImageFile block is then redacted to the [image omitted: …] placeholder by the #6010 vision gate, so the failure is silent — no HTTP 400, no crash, the agent simply replies that it cannot see the image. Vision-capable models were unaffected (they read the raw bytes either way), which is why this only surfaced on text-only + coalesce setups.

Change

• Extract the per-image describe-and-prepend logic into one shared helper, prepend_image_descriptions(handle, blocks), placed next to maybe_describe_inbound_image.
• Call it from both dispatch_message (immediate) and flush_debounced (debounced) right before dispatch_with_blocks, so the description step lives in one place and the two paths cannot drift out of parity again — the root cause was exactly that divergence.
• The helper describes every ImageFile block, not just the first, inserting each description immediately before its image. This matters because MessageDebouncer::drain coalesces the image blocks of several buffered messages into one list (all_blocks.extend(...)), so a debounced batch can legitimately carry multiple images.

The step self-gates inside describe_inbound_image ([media] image_description), so it remains a cheap no-op when the operator has not opted in, and a failed/timed-out description still degrades to the existing opaque [Image description unavailable] note rather than dropping the image.

Behavior note (intentional)

On the immediate path the description used to be prepended at index 0 (before the caption); it is now inserted immediately before the image block (after the caption), matching the debounced path. The description is still adjacent to and precedes the image, so model behavior is unchanged; the ordering is now consistent across both paths. No test asserted the old absolute-front position.

Verification

• cargo check -p librefang-channels --lib — clean
• cargo clippy -p librefang-channels --all-targets -- -D warnings — zero warnings
• cargo test -p librefang-channels --lib — 510 passed, 0 failed
• cargo fmt -p librefang-channels -- --check — clean

New regression tests in bridge::tests::inbound_image_description (reuse the existing DescribeHandle fake):

• prepend_describes_every_imagefile_before_its_block — multi-image list gets a description before each image; both images hit the vision path.
• prepend_is_noop_when_description_disabled — Ok(None) (feature off) leaves blocks unchanged.
• prepend_leaves_non_image_blocks_untouched — a text-only list triggers no vision call.

The existing maybe_describe_inbound_image cases (success / disabled / provider-failure-doesn't-leak / timeout / non-image-mime) continue to pass.

Out of scope

Live LLM round-trip verification (real Groq vision call against a coalesced bot) requires a running daemon and is left for a maintainer per the repo's human-only integration-testing policy.