Preserve first-captured MoE routes for extending prompts

[FurtherAI]

Summary

• Make overlapping MoE route assembly first-write-wins for multi-call trajectories with extending prompts.
• Fill only route positions that have not already been covered.
• Leave token identity, route shape, expert bounds, and downstream prefix-tree packing validation unchanged.

Why

vLLM retains routed experts for full prefix-cache blocks, but an unaligned suffix can be recomputed when a later request extends the prompt. Those tokens may move from decode execution to prefill execution and legitimately select different experts. A request that crosses an in-flight policy update can also recompute prompt positions under the new policy.

The earlier capture remains the replay authority for already-recorded tokens. Rejecting a later disagreement therefore aborts a valid trajectory instead of preserving its original routes.

Validation

• Real captured three-call trajectory: one 4,184-token sequence, 704 intended loss tokens, exact first-capture route ownership, unchanged tokens, masks, logprobs, offsets, and serialized trajectory.
• Focused MoE routing path: 14 passed.
• Adjacent preprocessing and prefix-tree tests: 44 passed, 2 environment-dependent Gemma tests deselected because the local CPU environment lacks the Transformer Engine shared library.
• Scratch overlap controls cover partial, empty, repeated, uint8, and uint16 inputs.
• Ruff, formatting, type checking, lock sync, and git diff checks pass.