fix(core): forward task images into the first LLM turn (multimodal passthrough)

[louisss1016]

Part of #813 (item A-1 — multimodal images silently dropped). Follow-up to #822 (A-4/A-5).

Problem

put_task() has always accepted images and every frontend passes them (fsapp downloads Feishu images to disk and forwards the paths; desktop_bridge and tui_v3 forward paths too), but agentmain.run() never handed them to agent_runner_loop — the first user turn stayed plain text and attachments were silently dropped. The desktop bridge papers over this with a backend.ask monkey-patch; fsapp and tui have no such workaround.

Even when image blocks do reach the LLM layer, two more drop points exist:

1. _to_responses_input() only understood OpenAI-style image_url blocks — Claude-style {"type": "image", "source": {...}} blocks vanished on the Responses API path.
2. NativeToolClient.chat()'s whitespace filter ([c for c in combined_content if c.get("text", "").strip()]) dropped every block without non-blank text, image blocks included.

Fix

• agentmain.py: new _multimodal_initial_content() builds the first turn's content blocks (text + base64 image blocks); run() passes them via agent_runner_loop(initial_user_content=...). Gated on NativeToolClient — only those backends understand image blocks (native Claude API takes them as-is; the OAI paths convert them), so other clients keep plain-text behaviour. Non-vision formats (e.g. SVG) degrade to a text path reference, and unreadable files surface as an error text block instead of vanishing silently.
• llmcore.py: _to_responses_input() converts Claude-style image blocks (base64 and url sources) to input_image, mirroring _msgs_claude2oai().
• llmcore.py: extract _drop_blank_text_blocks(); only blank text blocks are dropped now, image/tool_result blocks survive.
• desktop_bridge.py: remove the _patch_chat_for_images monkey-patch — with the core plumbing fixed it would double-inject every image. Attachments flow through put_task() instead. The guard tests in test_bridge_submit.py are repointed at the core mechanism (test_core_forwards_images_to_agent_loop, test_run_agent_turn_does_not_monkeypatch_images).

Tests

New frontends/tests/test_multimodal_image_passthrough.py (21 cases, red before the fix):

• _to_responses_input: base64/url image blocks -> input_image, OpenAI-style image_url regression guard, assistant role and malformed blocks ignored
• _drop_blank_text_blocks: blank text dropped, image/tool_result blocks kept
• NativeToolClient.chat end-to-end: image blocks reach backend.ask, blank text still dropped
• _multimodal_initial_content (extracted via ast, no import): no images / non-native client -> None, text+image blocks with verified base64 bytes, unknown extension -> image/png, SVG -> text reference, unreadable path -> error text, dict-shaped entries
• run() wiring: agent_runner_loop receives initial_user_content fed from task.get("images")

Full suite: 294 passed, with the same 3 pre-existing environmental failures as the untouched baseline (272 passed + symlink x2 + release packaging) — zero regressions, +22 new passing tests.