Serve the RoboLab leaderboard pi05 through the env-server (droid_jointpos)

[vertix]

Summary

First policy served against the merged RoboLab env (#469): openpi's DROID jointpos model — the same recipe RoboLab's leaderboard runs for its pi0-family entries.

• droid_jointpos codec (vendors/openpi/codecs.py): droid_obs + absolute joint-position actions, gripper binarized at 0.5, no horizon — the chunk length is the open-loop cadence for every pi0-family variant (pi05 = 15, pi0 = 10), matching their client's open_loop_horizon = model action_horizon. The full-chunk cadence itself comes from the validity sentinel Close action chunks with a validity sentinel in codec decode #488 added for Trajectories never state how long their last action lasts — ChunkedSchedule gives it zero time #487, so this PR carries no timing mechanism of its own.
• server.droid_jointpos preset: config pi05_droid_jointpos, checkpoint gs://openpi-assets-simeval/pi05_droid_jointpos (the policies/pi0_family/README.md recipe). The config lives in our openpi fork: Add pi05_droid_jointpos config serving absolute joint actions openpi#7.
• test_full_chunk_executes_between_replans pins the cadence contract over the full wire stack: every action of every chunk lands on the wire, replans exactly chunk_len periods apart.

Differential eval

Same checkpoint, same serving process (our server hosts openpi's serve_policy.py on :8001 and speaks the positronic wire on :8000). Two harnesses, 10 episodes per task per arm, on one L40S (2026-07-19):

• ours — positronic eval run → env-server proxy → RoboLab env
• theirs — RoboLab's own policies/pi0_family/run.py --policy pi05 --remote-port 8001, no positronic code in the loop

Task	ours	theirs
BananaInBowlTask	10/10	10/10
RubiksCubeAndBananaTask	5/10	6/10
BlockStackingOrderAgnosticTask	6/10	7/10

Per-task rates match within n=10 noise (RoboLab has no seed hook, so both arms draw fresh randomized scenes). BananaInBowl our-arm was also 9/10 on 2026-07-07 with the pre-rebase tree.

The eval ran with the pre-#488 tree, where an interim hold_last guard produced the same full-chunk cadence the sentinel now provides universally; the wire test above proves the cadence is unchanged on the rebased tree.

Test plan

• pytest suite green locally (605 passed / 7 skipped); test_remote_env pins the full-stack cadence.
• Differential eval above; episodes on the eval box under ~/eval_out/2026-07-19/, RoboLab-side logs under output/differential_2026-07-19/.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8049ca9e9b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".