[chmjkb]

Description

privacy-filter had two MoE paths. The old one read the router logits back to the CPU, picked the top 4 experts per token in JS, grouped tokens by expert and ran one matmul per expert. Nothing has called it since the GPU path (moeUp/moeDown) landed. This PR deletes it and makes the GPU path the only one.

• Delete routing.ts (routeTop4, groupByExpert, parseTopk) and the old forward, expertMatmul and gateColumnShape in moeMlp.ts.
• Rename EagerMoeMlp.forwardGather to forward. model.ts calls it.

The code that runs at inference is unchanged, so there is no performance change and no benchmark. Weights are unchanged: no parameter was added, removed or renamed, and stateDictHooks.ts is untouched. The hosted f32, f16 and int8 files load as before.

Prepares the privacy-filter migration in #70: one MoE path to move to the new weight format instead of two.

Introduces a breaking change?

• Yes
• No

EagerMoeMlp and routing.ts are privacy-filter internals, not exported from runntime/zoo.

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Performance (kernel tuning, graph building, anything related to performance)
• Other (chores, tests, code style improvements etc.)

Testing instructions

1. pnpm typecheck && pnpm lint && pnpm --filter runntime test:unit
2. pnpm --filter runntime test:gpu: same results as main (1439 passed, 16 failures that also fail on main).

Not checked: the full hosted model end to end in the Redact demo.

Related issues

#70

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings