Add measured offline Affordance collection

[yuecideng]

Description

Connect opt-in Affordance sampling to the run-env offline Task Program collection workflow. Sequential Pick/Place episodes now require measured target proximity, gripper release, and object stability before their rows can be saved. With four environments and a quota of five episodes, the final batch saves only enough accepted rows to reach five confirmed commits.

• Bind sampling to each batch/attempt and persist row-specific pose choices, episode identities, and resolved configuration fingerprints.
• Count synchronous LeRobot receipts only after files are independently readable; preserve prior episodes when resuming writes and block further writes after uncertain persistence failures. Optional trajectory output must also finish before the host confirms a row.
• Bound retries, discard rejected rows, and write a run manifest covering accepted rows, completed commits, and partial failures.
• Include a UR5 deployment, API documentation, and updated environment, Task Program, Atomic Action, and persistence context.

Scope is Affordance-only offline collection. Online collection, trajectory generation/coverage, parallel Task Program segments, and asynchronous receipt support are excluded. No dependency changes or linked issue.

Per-episode LeRobot finalize/reload adds I/O cost. Receipts provide same-process deduplication, not cross-process recovery or an atomic transaction across dataset and trajectory outputs.

Type of change

• New feature (non-breaking change which adds functionality)

Validation

• Affected production regression: 939 passed, 2 deselected (one default test exclusion and one existing DexSim compatibility failure explicitly excluded).
• Context/API tooling: 48 passed. Public API coverage: 2083/2083.
• Persistence follow-up: 60 passed, including independent dataset reloads immediately after each of three consecutive commits, before recorder shutdown, with earlier data preserved.
• Static deployment inspector: passed, 3 segments / 6 calls.
• black==26.3.1 with black ., structural checks, context map check, and git diff --check: passed.
• Code review findings on fingerprint completeness and premature LeRobot receipts were fixed and rechecked.

Draft validation limit: the real simulation smoke run failed during startup because the available DexSim build lacks WorldConfig.log_startup_info. The excluded existing config test requires DLSSConfig.frame_time_delta_ms, also absent in that build. Physical rollout success and yield remain unverified; mock tests and real LeRobot persistence tests do not establish physical task success.

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context and updated it where needed.
• Public API changes are reflected in the API docs (python docs/scripts/check_api_docs.py).
• I have added tests that prove the feature's supported contracts.
• Dependencies have been reviewed; no dependency changes are required.