Feat: downstream perf improvements - bump .25

[S1ro1]

Note

Medium Risk
Medium risk because it changes the on-the-wire JSON schema for routed_experts (from base64-encoded .npy strings to {data, shape} objects), which can break compatibility with existing clients/servers if not coordinated. The merge logic now relies on shape-provided sizing and row offsets, so malformed shapes could surface as new validation errors.

Overview
Updates routed_experts prefill/decode merging to use a raw uint8 payload encoded as JSON { "data": <base64 bytes>, "shape": [seq_len, layers, topk] } instead of a base64-encoded NumPy .npy blob, removing dtype/header parsing and related validation.

Tightens merge validation by checking row_start <= seq_len, verifying byte length matches seq_len * layers * topk with overflow-safe math, and updates tests accordingly. Also bumps crate/Python package version to 0.1.25.

^{Reviewed by Cursor Bugbot for commit 510092f. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Replace NumPy .npy encoding with raw uint8 JSON objects for routed experts payloads

• Replaces base64-encoded .npy strings with a JSON object {data: <base64>, shape: [seq, layers, topk]} for all routed experts payloads in routed_experts_merge.rs.
• Removes all NumPy header/dtype parsing logic (parse_npy_payload, parse_descr, dtype_item_size); row size is now computed as layers * topk assuming 1 byte per cell.
• Shape validation now parses a JSON array of three non-negative integers instead of a NumPy header string.
• Risk: Breaking change — callers providing .npy-encoded routed_experts strings will receive errors; payloads must be migrated to the new object format.

^{Macroscope summarized 510092f.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR changes the wire format for routed_experts from base64-encoded NumPy .npy files to JSON objects with data/shape fields. This protocol change affects runtime behavior and compatibility with other components, warranting human review.

^{You can customize Macroscope's approvability policy. Learn more.}