feat(v1): carry paired routing weights in single-turn traces

[faresobeid]

Summary

Add the verifier transport and trace layer for Total Router Recall (TRR): immutable paired expert IDs and captured FP32 coefficients, with token-row validity. Keep IDs-only routing supported.

• Validate the versioned HTTP payload before graph mutation. Preserve coefficient bytes and expert-slot order.
• Carry paired routing through graph node slices, branch concatenation, and raw-byte msgpack serialization.
• Accept complete capture or one unforwarded final sampled token. Mark that terminal placeholder invalid with zero weights.
• Reject full-coefficient delta capture, reused-prefix continuation, and shared-prefix forks. This MVP supports independent single-turn calls only.
• Extend existing graph/trace tests. No standalone routing test suite is added.

Validation

Current-main CPU validation, using the prepared renderer companion source:

• uv run pytest tests/v1/test_graph.py tests/v1/test_trace.py -o addopts= -q: 37 passed.
• Audited standalone routing suite, copied unchanged to an ignored external QA location and not included in this PR: 89 passed.
• Ruff lint/format, git diff --check, and uv run pre-commit run --all-files: passed.

A broader local-only contributor run was attempted before the final test fix: 79 passed, 15 failed, 1 collection error. Eleven failures and the collection error require omitted example/optional task packages. Four failures came from a new test fixture that dropped the transient routing field; the focused rerun above confirms its fix. The full suite is not claimed as passing.

The original audited implementation reported 111 verifier routing/graph/trace passes. That is prior evidence, not a result from this rebased branch. No live provider, sandbox, GPU, training, throughput, or numerical-equivalence result is claimed. Model expert-count validation remains the trainer loader's responsibility.

Companions

• Companion renderers draft: feat(client): preserve paired routing weights for Total Router Recall renderers#154
• Companion prime-rl draft: feat(rl): add experimental Total Router Recall prime-rl#3509
• vLLM coefficient capture is prepared separately; upstream submission is held pending the required human review, testing and DCO sign-off.

Related #2544 backfills IDs-only turn boundaries; #2408 handles opaque ID decoding. Neither adds coefficient transport or full-coefficient replay.

AI assistance

AI agents assisted with implementation, review, tests, and preparation of this draft.

Note

Add RoutingData carrier for paired routing weights in single-turn traces

• Introduces an immutable RoutingData dataclass in routing.py carrying FP32 weights, expert IDs, and per-row validity, with wire encode/decode and model-dependent validation.
• Updates graph.py _commit_turn to decode and validate full routing captures before mutation: rejects malformed coverage, unsupported reused-prefix/delta captures, and invalid message-span metadata; valid P+C-1 captures receive a synthetic invalid terminal row.
• Changes MessageNode.routed_experts and Branch.routed_experts to accept either legacy ID arrays or RoutingData, enforcing single-sampled-node replay and sampled-terminal semantics for full mode while preserving legacy None behavior.
• Adds HTTP/base64 and graph-wire decoders for legacy IDs-only and versioned full routing envelopes in routing.py; exports RoutingData from verifiers/v1/init.py and adds weight typing to types.py.
• Risk: _commit_turn now rejects reused-prefix continuations when existing prefix nodes already carry full routing and rejects mixed full/legacy routing evidence in concatenate_routing; callers passing legacy ID arrays must still use supported uint8/uint16 dtypes.

^{Macroscope summarized 036c828.}