Adds `docker_image = "docker.io/savatar101/env-data-agent-train:base"` under `[environment]` in all 250 task.toml files: the image every task's Dockerfile already builds FROM, which is all those Dockerfiles do besides `WORKDIR /workdir`. Harbor then uses the prebuilt image directly instead of building one per task, and the test set runs on HF sandboxes, which need a prebuilt image. Same change as the train set (https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-harbor-train/discussions/2), except for the image reference: this uses the `:base` tag rather than the digest. HF Jobs uses the image as a label, capped at 100 characters, and `docker.io/savatar101/env-data-agent-train@sha256:...` is 113, so a digest pin makes every HF sandbox fail with `Too big: expected string to have <=100 characters at labels.name`.
`_tool_calls()` read /workdir/.n_tool_calls and $N_TOOL_CALLS. Nothing writes that file and no task.toml sets that variable, so the count was always None and `_tool_efficiency` always returned null. Harbor's reward model is dict[str, float|int], so the null rejected the WHOLE dict and took `correctness` with it.
The only surviving path was the empty-submission early return, which hardcodes floats -- so the suite graded failures correctly and crashed on every successful answer. Measured on Qwen3.5-4B via opencode: of 250 tasks, the 86 invoking `grader.py --json` failed with '2 validation errors for VerifierResult' whenever the agent actually produced an answer.
The tool-call count is not the sandbox's to know -- it lives in the capture proxy, which sees every model call. A reward belongs in the verifier only if the sandbox is what makes it computable: correctness needs the data, the gold and the tolerances; tool efficiency needs a trace. Efficiency moves to the trainer, gated on corre…