Vla Evaluation Harness 0.7.0

Original release notes

v0.7.0

More reliable recording and concurrent container runs, with CPU rendering for additional benchmarks.

vla-eval merge is replaced by vla-eval export [-o OUTPUT_DIR]. Update scripts that call merge; its flags are not aliases for the new command.
Shards must share an eval ID and output directory. Export the resulting recording-.sqlite after the shards finish. scripts/run_sharded.sh handles this; single-shard runs export automatically.
Export preserves the last committed attempt per work item and marks incomplete runs as partial. Retry with the same eval ID, config, and task ordering. Legacy databases remain exportable, but missing historical completion/identity metadata cannot be reconstructed.

Benchmark entry fields can be overridden with --benchmark-field, e.g. --benchmark-field episodes_per_task=5. --param remains for benchmark constructor parameters. #155
ManiSkill2 and SimplerEnv support --render cpu using isolated Mesa lavapipe. Rebuild or update their images to use this path. GPU rendering remains the default. #158
Charliecloud prefers SquashFS when the required tools are available, with directory-image fallback. #159

Serialize shared Charliecloud image storage operations and bind temporary evaluation configs outside its shared temporary directory. #156, #157
Serialize SQLite writes and export from a consistent snapshot; preserve retry/task identity, track shard completion, and surface required recording/export failures. #160, #161
Accept SQLite 3.24+ with the required JSON functions instead of requiring 3.38+. Capability checks happen when a recording store opens. #162
Benchmark smoke tests share production Docker configuration handling, including user mapping and timeout cleanup. Model-server documentation clarifies per-episode policy state reset. #161, #163

Full changelog: v0.6.0...v0.7.0.