[rybosome]

Executive summary

The human summary from mic run is now easier to read. Score means are rounded to three decimals in an aligned Score table, and a new Elapsed line shows wall-clock time. Nothing else in the summary changes, and run.json and --json stay exact.

Before:

Status       completed
Trials       30 completed / 30 admitted
Failures     0 task, 0 scoring
necessity.exact: mean=0.976207729468599 count=30
necessity.false_unneeded: mean=0.0 count=30
necessity.unneeded_recall: mean=0.6666666666666666 count=12

After:

Status       completed
Elapsed      3:02
Trials       30 completed / 30 admitted
Failures     0 task, 0 scoring
Score        necessity.exact            mean 0.976  count 30
Score        necessity.false_unneeded   mean 0.000  count 30
Score        necessity.unneeded_recall  mean 0.667  count 12

Why

These summaries were hard to scan. Full float precision was shown, the metric lines were the only ones that didn't follow the labelled-column layout, and there was no indication of how long the run took. This is the third PR in the streaming-output work, agreed during the design discussion.

Design and implementation

• Layout. Every line keeps a 12-column label, so the summary stays easy to grep (Score, Requirement, Source). Metric names and means are padded to their widest value so the columns line up.
• Rounding (console._mean):
  • three decimal places by default;
  • exponent form for values that would round to 0.000 or are 1e6 or larger, e.g. 4.000e-04, so small and large values aren't flattened;
  • n/a when a metric has no recorded scores.
• Requirement lines keep the exact actual value. Rounding there could print mean 0.900 next to FAIL …>=0.9. The test uses that exact case.
• Elapsed time. The CLI measures wall-clock time around the run, including sink finishing, and passes it to format_summary(manifest, elapsed=...). It reuses the progress display's duration format (4.2s, 3:02). The manifest gets no new field, so there's no schema change. Python callers of format_summary without elapsed get no Elapsed line.
• Docs. docs/cli.md now describes the human summary, its rounding and the exact requirement values. The README run example is updated.

Verification and confidence

• New tests/reporting/test_console.py covers:
  • exact aligned output for the means from the original report (0.976, 0.667, 0.000);
  • exponent form for tiny and huge means, sign kept on negatives, n/a for empty metrics;
  • Elapsed placement and format, and its absence when no time is passed;
  • an exact requirement actual next to a rounded mean.
• Updated assertions: the summary checks in test_html and the CLI acceptance test, including an end-to-end check of the Score line and the Elapsed line.
• Manual previews: a suite run with a failing --require, a metric with no scores, and two sources all render as expected.
• scripts/verify.py: 782 tests, 0 failures. Typing, lint, format and the coverage gates all pass. The Node report checks can't run locally (no Node); they passed in CI's full verification job, along with the Linux, macOS, Windows and distribution jobs.

Compatibility and risks

• Human stdout output changes. Metric lines go from task.metric: mean=X count=N to Score task.metric mean X count N, and an Elapsed line is added after Status. Scripts that parse the human summary instead of --json would need updating; --json is the documented machine interface and is unchanged.
• No changes to the Python API, artifact formats, exit codes or dependencies. format_summary gains an optional keyword-only argument.

Relationships and tracking

Builds on #30 (engine observer hooks) and #31 (live progress display), both merged; this PR targets main. It completes the planned stack.

Unverified or follow-up work

None.

🤖 Generated with Claude Code