Use when a batch job is slow and you need the exact call site, query
    count, and fix — wherewent counts/groups/attributes deterministically in
    Python (zero token cost) and hands back machine fields to act on.

    Args:
        command: The job as an argv LIST, e.g. ["python", "job.py"] or
            ["python", "-m", "mypkg"]. TRUST BOUNDARY: this list is executed
            directly with NO shell (no shell=True) — no interpolation, no
            injection. The agent platform's sandbox is the outer boundary.
        unit_function: Optional SPEC (e.g. "myapp.jobs:process_receivable")
            to enable per-unit trend analysis (rising per-unit cost).
        timeout_s: Max seconds to let the job run (default 600). Real jobs can
            run for hours — on timeout the child is stopped and PARTIAL
            results are returned with timed_out=True. Partial data is a
            first-class use: read the per-unit growth trend.

    Returns an envelope: {"result": <enriched wherewent JSON or null>,
    "exit_code", "timed_out", "stderr_tail", "error"}. Each finding in
    result.findings carries fix/call_site/calls/evidence. Act on `fix`; cite
    `call_site` + `evidence`.
    

    For the "job already ran under `wherewent run --save out.json`" case —
    no re-run. Returns the same enriched schema analyze_job produces, with
    each finding carrying fix/call_site/calls/evidence.

    Args:
        path: Path to a JSON file written by `wherewent run --save`.