[merceod]

Two fixes in the robotics benchmark harness.

• run_robotics.py read the server's NDJSON with iter_lines() at its 512-byte default, which is roughly quadratic on multi-MB lines. Reading with a 1 MiB chunk takes a 4-rollout V-JEPA2-AC run from 10.0 s to 0.29 s of client time, latents bit-identical.
• DROIDDataset read tasks_df["task"], but droid_100 is a lerobot v3 dataset whose tasks.parquet keeps the instructions in the index. The lookup warned could not load tasks.parquet: 'task' and every pi0.5 request got the prompt "manipulate the object" (V-JEPA2-AC sends no text, so it was unaffected). It now reads the index, where lerobot v3 keeps the strings: requests 0 and 1 become "Put the marker in the pot" and "Put the candy bar on the left side of the first shelf". pi0.5 numbers from before this PR used the fallback prompt, so they aren't comparable with new ones.

Tests: new test_droid_dataset_tasks.py (1 passed). The eval and bench test modules show the same two box-specific failures as main when run serially. e2e --tier smoke pi05 passes; run_robotics against servers: vjepa2_ac 4 + 16 + 4 rollouts, pi05 eager 4/4 with the real prompts. ruff clean.

[mstar-review]

Automated review

🟡 No findings on b1fa740, light review

Description

• The 10.0 s → 0.29 s client-time figure doesn't say how it was measured (which timer, what hardware, or the server config and command used for the 4-rollout run), so it can't be reproduced.
• "The same two box-specific failures as main" doesn't name the two failing tests, so a reviewer can't check that the set of failures matches main.

What the review checked

• _task_strings uses the task column when it exists and otherwise falls back to the index, so the older lerobot layout gives the same dict as the inline code it replaces.
• benchmark/dataset.py only imports stdlib, numpy, requests and other benchmark.* modules at top level, and other test/modular modules already import benchmark, so the new test can run in CPU Core; it also skips when pandas is missing via importorskip.
• Both iter_lines call sites in run_robotics.py (_send_pi05 NDJSON branch, _send_vjepa2_ac) now pass the same chunk size, and the line-splitting behaviour and the JSON-decode skip are unchanged.
• The change is limited to the benchmark harness and touches no engine, worker or model code, so none of the runtime invariants apply.

_{Advisory, not a merge gate. Rules: AGENTS.md. Reply to discuss.}

[dghosef]

lgtm