chore: bump verifiers to main (v1 seq_len fix)

[hallerite]

What

Bumps the deps/verifiers submodule e1d4f259 → 05c66c23 (current verifiers main).

The headline reason is the merged v1 token-usage fix (verifiers#1525): for v1 environments, token_usage.final_input_tokens/final_output_tokens were always 0 because the trajectory's responses are serialized to plain dicts and compute_context_token_metrics gated on isinstance(response, Response). That zeroed our wandb seq_len/* and progress/tokens metrics. v1 now records those metrics at write time from the live Response.

The bump also pulls in the other commits merged to verifiers main since e1d4f259 (Scale-SWE taskset, v1 sandbox runtime scaling, Multi-SWE patch alignment, unbounded harness max_turns, port-rebind fix).

Validation

Ran configs/debug/reverse_text_v1.toml end-to-end on GPU against the bumped submodule:

metric	before	after
seq_len/all/mean	0	160.1
progress/tokens	0	21013

Full inference + orchestrator + trainer loop completed cleanly on the new main. uv.lock unchanged (verifiers is a workspace/path dep).

---

Note

Low Risk
Dependency-only submodule pointer update with no in-repo logic changes; validated on a debug v1 training run with improved metrics reporting.

Overview
Updates the deps/verifiers git submodule from e1d4f259 to 05c66c23 (current upstream main), so this repo picks up verifiers’ v1 token-usage fix: trajectory responses serialized as dicts no longer leave token_usage.final_input_tokens / final_output_tokens at zero, which restores meaningful seq_len/* and progress/tokens reporting (e.g. wandb) instead of flat zeros.

The same bump also brings in other recent verifiers main work (Scale-SWE taskset, v1 sandbox runtime scaling, Multi-SWE patch alignment, unbounded harness max_turns, port-rebind fix). uv.lock is unchanged because verifiers is a workspace/path dependency; validation noted a full reverse_text_v1 debug run with non-zero sequence-length and token progress metrics.

^{Reviewed by Cursor Bugbot for commit db56c75. Bugbot is set up for automated code reviews on this repo. Configure here.}