feat(xla): support Molmo sparse additive visual embeddings

[inureyes]

Parent and dependency

• Parent epic: epic: OpenXLA multimodal / VLM architecture support (Qwen2-VL, Gemma3n, Phi4MM, Molmo, and more) #566
• Blocked by test(xla): validate a LLaVA reference architecture end to end #862
• Existing MLX implementation: src/vision/molmo_vl.rs, src/vision/encoders/molmo.rs, src/vision/processors/molmo.rs, and src/models/molmo.rs

Problem

Molmo v1 does not replace placeholder embeddings like LLaVA. Its processor supplies image_input_idx; negative entries are skipped, and selected projected feature rows are added to the existing OLMo text embeddings at the remaining target positions. The vision path also includes multi-crop patch masks, selected ViT features, attention pooling, and a SwiGLU projector.

Scanning for an image token or replacing embedding rows would implement a different merge while still producing shape-valid inputs.

Goal

Support a pinned Molmo v1 checkpoint end to end on XLA with IREE vision execution, processor-supplied sparse additive merge, OLMo language prefill/decode, CLI, and continuous batching.

Required implementation

• Reuse/share MolmoProcessor crop, patch, image_masks, and image_input_idx generation. Preserve index order and negative-sentinel semantics in an owned prepared payload.
• Define static maxima/buckets for crop count, patches per crop, active feature rows, and expanded text length. Validate allocation arithmetic and reject overflow before tensor construction.
• Emit the Molmo CLIP-style ViT, configured intermediate-layer selection/combination, patch-mask behavior, attention pooling, and SwiGLU projector through StableHLO/IREE with explicit weight-layout checks.
• Keep mask semantics distinct: image patch coverage controls the vision encoder/pooler, while the language model uses its normal causal mask.
• Implement sparse additive merge from (feature_row, target_position) pairs. Require target positions to be in range and unique unless the reference explicitly defines repeated accumulation; require feature rows to be in range and preserve their mapping after sentinel removal.
• Cast projected features once to the post-scale text-embedding dtype, then add rather than replace. Document any OLMo embedding scaling at the feat(xla): add a prefill-from-embeddings StableHLO entry #858 boundary.
• Include crop/patch config, selected vision layers, pool/projector config, merge mode, and OLMo text architecture in artifact compatibility identity.
• Integrate filtered loading, capabilities, CLI/server batching, cancellation, and no-media behavior without retaining a duplicate full MLX decoder.

Non-goals

• Molmo2 token-scanned merge or pooled-patch index contract.
• Molmo Point coordinate output.
• General arbitrary sparse writes supplied by API clients.

Validation

• Pin one Molmo checkpoint/revision and compare processor crops/patches/masks/indices, selected vision states, pooled/projected features, sparse active pairs, merged embeddings, prefill logits/KV, and greedy tokens against MLX and an independent oracle.
• Cover negative sentinels, non-contiguous target positions, one and multiple crops, masked patches, empty active features, invalid/duplicate indices, capacity overflow, no-image behavior, cancellation, and slot reuse.
• Add a negative fixture that performs replacement instead of addition and prove the merged-embedding/logit oracle catches it.
• Run CLI and mixed continuous-batch server tests on a production-relevant IREE target, plus test(xla): validate a LLaVA reference architecture end to end #862 and text-only OLMo gates.

Acceptance criteria

• Processor-supplied image_input_idx is authoritative; no token scan substitutes for it.
• Negative sentinels and feature-row/target-position correspondence are preserved exactly.
• Vision/pooling/projector intermediates and sparse additive merge match both oracles.
• Deterministic output is token-exact on the pinned fixture.
• Index/capacity errors fail before native execution and do not affect other slots.
• No full decoder duplication or silent MLX vision fallback remains.
• Text-only OLMo behavior remains token-exact.