Resolve training runs by the sampler-path run ID

[kevintli]

Summary

Sampler paths are minted as tinker://<model_id>:train:0/sampler_weights/<name> (_sampler_model_path), but checkpoint entries are keyed by the bare model_id. When a user calls the stock cookbook's checkpoint_utils.get_renderer_name_from_checkpoint_async(service_client, sampler_path), it goes through rest.get_training_run_by_tinker_path → GET /api/v1/training_runs/<model_id>:train:0, and that request 404s:

training run not found: ae716ee9...:train:0

The cookbook only logs a warning and falls back to the configured renderer, so this does not break anything. It does mean renderer metadata never round-trips for sampler checkpoints (seen with sec-rl eval --checkpoint <sampler_path> in the sec-search-rl repro).

entries = await self.checkpoints(training_run_id)
+ if not entries and training_run_id.endswith(":train:0"):
+     entries = await self.checkpoints(training_run_id.removesuffix(":train:0"))

This PR is stacked on #7. It is not deployed to the repro frontends, because redeploying would restart their control planes in the middle of the runs.

Link to Devin session: https://modal.devinenterprise.com/sessions/f53cfabb210146de8f0338fe388d7973
Open in Devin Desktop: https://modal.devinenterprise.com/desktop/session/f53cfabb210146de8f0338fe388d7973?variant=devin
Requested by: @kevintli

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring