[faresobeid]

Stacked on #60. Draft, parked. No prime-rl producer yet: a future TRR fields PR would write routed_expert_weights segments (PrimeIntellect-ai/prime-rl#3904 is closed).

STITCHED_FIELDS = ["routed_experts", "routed_expert_weights"] lists the payload fields that the forward pass produces for every position. Each is stitched separately: the prefill segments come first, then the decode segments clipped at that field's own prefill coverage end. Row bytes are dtype-aware: uint8/uint16 ids, float32 weights (float32 is added to dtype_itemsize here). Every other field (sampling_mask, sampling_mask_logprobs) comes from decode only and passes through. A decode stitched field with no prefill counterpart is an error, as before.

Checks: cargo fmt --check, cargo clippy --lib --tests -D warnings, cargo test --lib. I extended the existing merge test with weights and logprobs.

Note

Stitch routed_expert_weights segments alongside routed_experts in HTTP router

• Adds a STITCHED_FIELDS set with routed_experts and routed_expert_weights in routed_experts_merge.rs, so payload segments for both fields are collected and stitched
• merge_routed_payload_in_json now tracks prefill coverage per field in a HashMap and clips each decode segment against its own field's prefill endpoint. Non-stitched fields like sampling masks pass through unchanged
• A decode segment for a stitched field with no matching prefill coverage now returns an error naming that field
• Behavioral Change: decode-only routed_expert_weights payloads that previously passed through now fail with an error instead of being silently unmerged

^{Macroscope summarized 4c87216. (Automatic summaries will resume when PR exits draft mode or review begins).}