R3 delta replay picks (no configs)

[samsja]

Rebased version of #2647 on top of main, with unrelated changes removed.

Commits

• Implement routed experts delta replay (with branched deltas) — routed_experts.py / serving_tokens.py / trajectories.py + verifiers bump + router pin
• fix(orchestrator): match longest active prefix in interleave_rollout
• warn on ambiguous prefix match in interleave_rollout

Removed from the original PR

• The 4 configs/debug/qwen3_30b_a3b_pd_*_router_replay*.toml debug configs.
• Handle token export mkdir races commit — unrelated concern, belongs in its own PR.
• skills/configs/SKILL.md RLM-SWE note — unrelated.
• rlm-swe / wordle workspace entries in pyproject.toml — unrelated.
• uv.lock churn from the above (reverted to main).

Dependencies

• deps/verifiers bumped to d39cc5876 (as in R3 delta replay picks #2647).
• vllm-router pin switched to a local-path wheel that contains the router-side start field plumbing (counterpart to Merge routed experts deltas with start offsets router#37). This needs to flip back to a release URL once router#37 is merged and a wheel is published.

---

Note

High Risk
Changes how training samples merge tokens, masks, and MoE routing tensors; wrong prefix or delta stitching would silently corrupt RL training data, though behavior is heavily covered by new unit tests.

Overview
Adds routed-experts delta replay end-to-end: compact payloads now carry a start index (from routed_experts_prompt_start at inference capture through orchestrator merge), so multi-step and branched rollouts can stitch partial vLLM expert tensors instead of mis-aligning rows.

interleave_rollout is updated to extend the longest matching active token prefix (fixes compaction/rollback where a shorter prefix wrongly absorbed a longer sample’s completions) and to warn when multiple prefixes match. New-branch steps can replay routed-expert chunks from a matching prior prefix state when start > 0.

Dependency: vllm-router is wired from a local wheel under third_party/router/dist/ instead of a release URL.

Unit tests cover longest-prefix matching, multi-step/branch expert alignment, and the start field on serialized experts.

^{Reviewed by Cursor Bugbot for commit fc9fbaf. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9a99ff5. Configure here.}