feat: re-pin verifiers for per-call trace records + generation time split

[mikasenghaas]

Summary

Companion to the merged verifiers trace work: per-call records (#2061, RES-1085) and the model/harness generation-time split (#2060, RES-1084).

• Re-pins deps/verifiers to d5320edcb (both merges): Trace.calls per-call ModelCall records (node, model, sampling, endpoint, finish_reason, usage, time, error; upstream status on Error.status_code), with finish_reason/usage moved off MessageNode onto ModelCall (trace schema v2); and Timing.generation becoming a GenerationSpan with model/harness TimeSplits.
• tests/unit/orchestrator/test_advantage.py: the rollout builder attaches provider usage via a per-call vf.ModelCall instead of the removed node field.
• Exports the generation split to wandb: TimingMetrics gains generation/model and generation/harness (per-rollout timing.generation.model.duration / .harness.duration), emitted as {prefix}/{subset}/timing/generation/model/* and .../generation/harness/* alongside the existing timing/generation/*. The env server stamps split_generation in Rollout.run, so the split rides the wire already populated.

No other prime-rl behavior change: GRPO length penalties and token metrics consume the trace-level num_*_tokens, which verifiers still derives (branches read usage off their attached calls); calls and the generation split ride the existing records additively.

Breaking

• Inherited from the verifiers pin: trace schema v2 — previously persisted traces (rollouts/step_N/**/traces.jsonl from older runs) no longer validate under the new verifiers, since their nodes still carry finish_reason/usage. Read old dumps with the old verifiers.

Verification

• tests/unit/orchestrator + tests/unit/utils green at the new pin (159 passed); test_metrics extended to assert the timing/generation/model + timing/generation/harness wandb keys.
• Reverse-text RL smoke on this box (2x RTX PRO 6000 Blackwell, subprocess env server, 3 steps): trains end-to-end (reward 0.12 -> 0.17, 0% error). Across the 552 saved rollouts the split is populated on every one - timing/generation/model/mean = 0.775, timing/generation/harness/mean = 3.209, timing/generation/mean = 3.984 (model + harness == generation exactly).

🤖 Generated with Claude Code