Inline raw image refs for multimodal rendering

[eligotts]

Summary

Intermediate multimodal storage mode — the inline bundle. Keeps this repo's raw multimodal contract (renderer computes token layout via pure geometry math, no image processor at render time; refs + descriptors instead of processed tensors) but images travel inline as data:image/...;base64 URLs instead of offloaded file:// run assets.

• mm_store: the mmraw ref envelope carries raw_image_data (the inline data URL) instead of raw_image_uri. offload_image_to_run_assets, run_image_dir, and VF_RENDERER_IMAGE_OFFLOAD_DIR are gone. New decode_data_image_url() is the single supported source decode.
• Qwen3-VL / Qwen3.5 / Kimi-K2.5 raw layout describes decode the inline source once for dimensions + sha256 hash; descriptors embed the source.
• client.generate() serializes the inline source into every image slot's ref — every request to /inference/v1/generate carries the full inline payload for every image in the prompt (current and prior turns), and the prime-rl serving layer dedups processor work by content hash.

Relationship to the offload bundle

This PR is cut from the offload branch (#89) and is the storage-layer intermediate: same processor-ownership architecture, no shared image directory. The offload bundle (renderers #89 / verifiers #1746 / prime-rl #2836) layers content-addressed file:// offload on top of this contract.

Companion inline PRs: verifiers feat/v1-inline-raw-images, prime-rl feat/v1-inline-mm.

Validation

• Full suite: 2153 passed, 121 skipped, 1 xfailed (processor-parity tests included, run against the real Qwen3-VL-4B and pinned Kimi-K2.5 processors).

Note

Send raw image refs to vLLM by default for multimodal rendering

• Introduces a multimodal_output config field (default "raw") on BaseRendererConfig; renderers for Qwen 3.5, Qwen3-VL, and Kimi K2.5 now emit raw image reference items instead of processed pixel tensors unless multimodal_output="processed" is set.
• Adds a new renderers/mm_store.py module with primitives for building, encoding, and decoding raw multimodal item references consumed by vLLM.
• Layout metadata (grid dimensions, token counts) is now computed deterministically from a baked spec via describe_qwen_image_layout / describe_kimi_image_layout without requiring an HF processor, enabling token counting without loading model weights.
• Processor loading is now lazy across all renderers; constructors no longer accept an injected processor, and the image_cache_max config field is removed from all renderer configs.
• renderers/client.py serializes raw mm refs into vLLM feature payloads via _build_vllm_mm_features, replacing the prior processed-tensor dispatch path.
• Behavioral Change: image parts in multi_modal_data now carry raw reference descriptors by default; callers expecting processed pixel tensors must set multimodal_output="processed".

^{Macroscope summarized d54e7fc.}