feat: Render per-group video timestamps for the Qwen3-VL family #2402

Problem / Background

The reference Qwen3-VL processor renders a video as one <t.t seconds> text marker plus its own <|vision_start|>...<|vision_end|> block per temporal patch group, and splits each video_grid_thw row (t, h, w) into t rows of (1, h, w) for MRoPE. mlxcel renders one flat, timeless video block. Upstream found that without the markers the model reads a clip as a spatial collage (a moving square described as "several squares"); Neywa-Labs/mlx-vlm#2142 added the per-group rendering and Neywa-Labs/mlx-vlm#2413 took the timestamps from the frames actually sampled rather than an assumed 2 fps. Reference: _video_group_timestamps and _timestamped_video_placeholder in https://github.com/Blaizzy/mlx-vlm/blob/main/mlx_vlm/models/qwen3_vl/processing_qwen3_vl.py, and the get_rope_index split in https://github.com/Blaizzy/mlx-vlm/blob/main/mlx_vlm/models/qwen3_vl/language.py.

Current Behavior

• insert_qwen_vl_media_tokens (src/multimodal/qwen_vl.rs:558-700) expands each video placeholder into one run of grid_t * (h/m) * (w/m) video tokens framed by a single start/end pair. It works on token ids and has no tokenizer.
• qwen_vl_mrope_positions_from_tokens (qwen_vl.rs:949-1073) treats each video run as one (t, h, w) grid and advances current_pos by max(t, h/m, w/m).
• The server video path (src/server/model_worker.rs:3440-3492, also used by mlxcel generate on the engine) calls video::load_video_source_with_policy (src/multimodal/video.rs:1350-1384), which computes uniform_indices(meta.total_frames, nframes) and meta.fps but returns only Vec<DynamicImage>, so the timing is lost. Qwen2VLMediaInput::Video (src/vision/processors/qwen2_vl.rs:138-141) carries frames only.
• The same processor and expansion serve Qwen2-VL and Qwen2.5-VL, whose reference processors do not emit timestamps.

Proposed Solution

1. Timing. Add pub struct SampledVideo { frames: Vec<DynamicImage>, frame_indices: Vec<usize>, source_fps: f64 } and load_video_source_sampled_with_policy in video.rs; make load_video_source_with_policy a thin wrapper returning .frames. The worker uses the new function and keeps the timing beside the frames.
2. Timestamps. Port _video_group_timestamps: pad frame_indices by repeating the last index to a multiple of temporal_patch_size (the processor's field, qwen2_vl.rs:32), frame time is index / source_fps, group time is the mean of the group's first and last frame time. Error if the padded count differs from grid_t * temporal_patch_size or source_fps is not positive and finite.
3. Family flag. Add video_timestamps: bool to QwenVlmPromptInfo (qwen_vl.rs:28-34), true for Qwen3-VL, Qwen3-VL-MoE, Qwen3.5-VLM and Qwen3.5-MoE-VLM, false for Qwen2-VL and Qwen2.5-VL (their output stays byte-identical).
4. Expansion. compute_qwen_vl_media_embeddings (src/multimodal/vlm_runtime.rs:942) gains video_timings: &[Vec<f64>] (group seconds per video) and encode_text: Option<&dyn Fn(&str) -> anyhow::Result<Vec<i32>>>; the worker passes its tokenizer's encode without special tokens. When video_timestamps is set, each video placeholder (including its surrounding start/end pair) becomes, per group g, encode_text(format!("<{:.1} seconds>", t_g)), start, (h/m)*(w/m) video tokens, end. A timestamp family with encode_text == None is an error, not a silent fallback. The already-expanded-prompt validation path must accept this shape.
5. MRoPE. For timestamp families, expand each video grid (t, h, w) to t grids (1, h, w) before qwen_vl_mrope_positions_from_tokens, so each group run is positioned on its own and the timestamp text between runs gets text positions. The adjacent-run check already permits runs separated by framing tokens.

Scope

In scope: video.rs, qwen_vl.rs, vlm_runtime.rs, the worker video path, the four Qwen3-family runtimes' prompt_info, tests beside each.

Out of scope: Qwen2-VL / Qwen2.5-VL second_per_grid_ts behavior, Qwen3-Omni (separate runtime), image paths.

Acceptance Criteria

• _video_group_timestamps port test: indices [0, 15, 30, 45, 60] at 30 fps with temporal_patch_size = 2 give [0.25, 1.25, 2.0].
• Expansion test with a stub encode_text: one video with grid_thw = (3, 4, 4), merge 2, yields three <t seconds>/start/4 tokens/end blocks in order; a Qwen2.5-VL prompt_info yields the old single block unchanged.
• MRoPE test: the split grids produce temporal positions that restart per group after the timestamp text, and total length matches the token count.
• Integrated: mlxcel generate -m mlx-community/Qwen3-VL-4B-Instruct-4bit --video <clip of one moving square> -p "Describe the motion" describes a single moving object; post the before/after outputs in the PR.

Verification

cargo test --profile test-fast -p mlxcel qwen_vl
cargo test --profile test-fast -p mlxcel multimodal::video
cargo test --workspace --profile test-fast --features metal,accelerate
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --check