Make public CLI and example configs truthful

[bordeauxred]

Why

Some public commands/configs imply support that is not currently wired, which can confuse first-time users exploring the examples.

What

• Decide whether clawloop run should delegate to the unified TrainConfig runner or remain a legacy benchmark-only command.
• If clawloop run remains documented, make sure documented benches work or remove them from public docs.
• Fix or move examples/configs/taubench_harness.json if it is not intended to be a valid TrainConfig.
• If TauBench is advertised as supported through examples/train_runner.py, register it in the train environment registry and add validation coverage.
• Add tests for documented CLI/config paths.

Acceptance

• No documented public command fails with Unknown benchmark.
• Public example configs either validate or are clearly marked as experimental/non-TrainConfig.
• README/examples distinguish stable demos from heavier experimental adapters.

[bordeauxred]

Codex review (2026-04-19) — addition:

If clawloop run is not refactored before launch but stays in the CLI, it must at minimum print a truthful error when an unsupported bench is requested:

Supported benches: entropic, car. For the math demo use `clawloop demo math --dry-run`.

Silent Unknown benchmark failures on benches the README mentions are a credibility-killer. Either (a) wire up the TrainConfig/ENV_BUILDERS delegation, (b) restrict the CLI's accepted bench list to ones that work, or (c) hide run from the README and --help until it's truthful.