Support delta-only routed experts replay

[S1ro1]

Summary:

• Preserve dense routed experts while transferring only delta routed-experts sidecars with explicit start offsets.
• Reconstruct routed prefix chunks for branched/compacted trajectories, then finalize trainer samples with dense routed experts.
• Match the longest active prefix in interleave_rollout so regenerated/compacted branches do not get folded into shorter active samples.
• Pin deps/verifiers to merged main with routed_experts_prompt_start support and vllm-router to the released v0.1.26 wheel.

Validation:

• uv run pytest tests/unit/orchestrator/test_trajectories.py tests/unit/inference/test_serving_tokens.py -q
• uv run ruff check src/prime_rl/inference/vllm/routed_experts.py src/prime_rl/inference/vllm/serving_tokens.py src/prime_rl/orchestrator/trajectories.py tests/unit/inference/test_serving_tokens.py tests/unit/orchestrator/test_trajectories.py

---

Note

Medium Risk
Changes affect MoE training signal assembly and multi-step rollout merging; incorrect routing alignment would skew RL updates, though behavior is heavily asserted in unit tests.

Overview
Adds delta-only routed experts with an explicit start offset on compact sidecars: inference serializes/captures using routed_experts_prompt_start, and the orchestrator rebuilds dense per-token routing for training by stitching prefix chunks, enforcing delta alignment on extensions, and padding to full sequence length.

interleave_rollout now picks the longest matching active prefix when several samples overlap (e.g. compaction/rollback), with a warning on ambiguous matches, so regenerated branches are not merged into shorter stale samples.

Dependency pin: vllm-router 0.1.25 → 0.1.26 (lockfile updated).

^{Reviewed by Cursor Bugbot for commit cda9b16. Bugbot is set up for automated code reviews on this repo. Configure here.}