fix(channels/discord): surface image attachments to text-only providers

[benhoverter]

Summary

Fixes #1142.

Discord image attachments were silently dropped before reaching the model on text-only providers. Captioned images had attachments discarded (caption-wins path); bare-image messages were dropped entirely (early content.is_empty() return). Result: hallucinated acknowledgements of unseen images, or no response at all.

This PR rewires the inbound attachment path end-to-end so attachments survive parsing, dispatch coherently across vision-capable and text-only providers, and render as a clear marker on text-only providers instead of being silently elided.

Changes

• openfang-channels types: new ChannelContent::Multipart(Vec<ChannelContent>) variant for caption + attachment(s) as sibling blocks. Nesting forbidden (doc + debug_assert).
• Discord parser: rewritten to handle bare-image and captioned-image shapes correctly. MIME classification with filename-extension fallback for payloads missing content_type. 5 MB vision-size cap matching Anthropic's image-block limit; over-cap images classify as File. Emits Multipart whenever text + attachments coexist or multiple attachments are present.
• Bridge: flat-maps Multipart in both dispatch paths — Vec<ContentBlock> for multimodal-capable providers, newline-joined text descriptor for text-flatten providers.
• Telegram channel: exhaustive-match parity for the new variant; defensive outbound flatten.
• claude_code driver: renders Image blocks as [attachment: <mime> image, ~N KB — not viewable on this provider] instead of dropping them. The model still cannot see the image, but it can acknowledge it coherently rather than confabulating.

Out of scope (follow-ups): persisting attachments to disk; vision-provider dispatch refinements beyond exposing existing image bytes; non-Discord inbound attachment classification; handling for files larger than the 5 MB vision cap.

Testing

• cargo clippy --workspace --all-targets -- -D warnings passes
• cargo test --workspace passes
• Live integration tested (if applicable)

Live smoke test on a local-main daemon build, both shapes:

• Captioned image ("Deployed..." + PNG) → daemon log shows Multipart([Text, Image{...}]); model prompt contains caption + [attachment: image/png image, ~3172 KB — not viewable on this provider]; LLM acknowledgement coherent.
• Bare image (PNG, no body) → daemon log shows image dispatch; model prompt contains [attachment: image/png image, ~2774 KB — not viewable on this provider]; LLM acknowledgement coherent.

Added 9 parser unit tests covering all (text-empty, n-attachments) shapes plus MIME edge cases (HEIC, oversize, missing content_type), and 2 driver unit tests covering captioned and bare-image marker rendering.

CI note: a pre-existing cargo fmt --all -- --check failure on main (introduced in da6b567a and earlier; tracked by #1121) is inherited by this branch. None of the failing fmt lines are in files this PR modifies. Happy to address as a separate fmt-only PR if maintainers prefer.

Security

• No new unsafe code
• No secrets or API keys in diff
• User input validated at boundaries
• No personal agent names in diff, description, tests, or fixtures

[benhoverter]

On the Security Audit failure: this PR's Cargo.lock is byte-identical to RightNow-AI/openfang's main.

git diff --stat rightnow-ai/main...discord-file-sharing -- Cargo.lock        # empty
git diff --stat rightnow-ai/main...discord-file-sharing -- '**/Cargo.toml'   # empty

Zero deps added, removed, or bumped. The same advisories fire on main's latest CI run (25124418579, head 15ed29c = bump v0.6.2), so the failure is pre-existing on main rather than introduced here.