fix(cli): validate before anything is sent, put status on stderr, exit 1 when nothing was measured

[TMHSDigital]

Fixes #41. Fixes #42. Fixes #15.

#41: options that misbehaved or crashed.

• --limit, --workers, --max-cases and --boot must now be positive, checked by the parser. The limit is tested with is not None rather than truthiness, so:
  • --limit 0 no longer runs every row;
  • a negative limit no longer slices from the end.
• --boot 0 used to fail with a traceback after the run and the artifact, when the calls were already paid for. It is now refused at parse time.
• Each of these is now one line, before anything is sent:
  • an option the chosen adapter doesn't take (such as --revision on typesafe_wire);
  • a missing API key, with the variable to set;
  • a --report path that is a directory.

#42: the exit code and the streams.

• Status lines and errors go to stderr, so stdout carries only the report when no --report is given, and plumbline run ... > report.md captures just the report.
• A run where no case produced a prediction writes its artifact and report, then exits 1. When every failure shares one reason, it says that reason.
• README documents which stream carries what, and the exit codes.

#15: the shared failure reason. With no figures and one shared reason (usually a missing extra or key), the report now prints that reason instead of pointing at the artifact. With several different reasons, it says how many and leaves them to the artifact.

Checked

• New tests failed on main and pass now:
  • --limit 0, --limit -1 and --boot 0 are refused, with no traceback and no results directory created;
  • --revision on typesafe_wire and a missing TYPESAFE_API_KEY are one line each;
  • a directory --report is refused before the run;
  • status goes to stderr and the report alone to stdout;
  • an all-failing run exits 1, still writes the artifact, and names the reason in stderr and in the report;
  • the report handles both the one-reason and many-reasons cases.
• Existing CLI tests that read status lines now read result.output, which holds both streams.
• ruff, mypy --strict, the full pytest suite, and the site build all pass.

🤖 Generated with Claude Code