fix(recognize_image): restore vision fallback in remote workspace mode

[cursor]

Bug and impact

When a session uses a remote workspace and the main LLM lacks vision, Feishu image attachments synced via attachment_sync are downgraded to text placeholders that point at .macchiato/inbox/... paths. Calling recognize_image to understand those images always failed because resolve_media_to_content_item rejected all local path resolution whenever remote mode was active—even though attachment sync keeps a daemon-local copy specifically for vision hydrate.

Impact: image understanding is completely broken for the common path of remote workspace + non-VL main model + Feishu attachments.

Root cause

1. resolve_media_to_content_item had a blanket early return for remote workspaces before checking whether the file exists locally.
2. recognize_image expected image_url content items but resolve_media_to_content_item returns media_ref, so path-based fallback never produced a usable data URL.
3. _lookup_unseen_media did not match remote_path, so models following prompt hints could not resolve synced attachments.

Fix

• Allow daemon-local files in remote mode when the resolved path exists on disk.
• Add resolve_media_path_to_data_url helper and use it from recognize_image.
• Teach recognize_image to match remote_path in unseen media and fall back to file_blob_read for remote-relative paths.

Validation

• uv run pytest tests/test_recognize_image_tool.py -v (10 passed)
• New tests cover local daemon copy in remote mode and remote-relative blob reads.