[feat] Count a scorer's labels by value in the dashboard #1472

Parent: #1137

Goal

The scorer view of #1182 counts a measurement's labels as passed and failed. A scorer with more than two labels loses what told them apart. partial and bad both fail. For a routing question (#1469), missed and extra both fail, and fit and idle both pass, so its precision and recall cannot be read.

Count the labels of a measurement by value too.

Example

teamai:routing:ops:release · rubric 5b2d0a · session · claude / <recorded-model> · attended · 30 days
label    score   labels
fit      1       31
idle     1       112
extra    0       6
missed   0       9
158 labels · 143 passed · 15 failed · 91%

precision(ops:release) ≈ 31 / (31 + 6) = 84%
recall(ops:release)    ≈ 31 / (31 + 9) = 78%

The view shows the counts. Precision and recall are read from them; the view does not compute them.

Scope

• Under the pass rate of a measurement, one row per label value it holds, with the value, its score and its count. Rows go by score, highest first, then by value.
• The same labels as the pass rate: after the newest-scoredAt dedup and the window, so the counts add up to the measurement's label count.
• Values as the labels stored them. The scorer file is not read to rename, add or reorder them.
• The read API returns the counts on the selected measurement (selected.cohort). Docs: the Scorers section of docs/guide/dashboard.md and docs/guide/zh-CN/dashboard.md.

Acceptance criteria

1. A measurement with fixture labels fit ×3, missed ×1 and extra ×2 shows those three rows, and they add up to its label count of 6.
2. A label rescored in a later month with another value counts once, under its newest value.
3. A value with no label in the window has no row.

Out of scope

• Counts by value inside groups and in the split at a team commit.
• Computing precision, recall or any other metric.
• Labels from benchmark runs, which the view already leaves out.

Dependencies

• Blocked by: [feat] Show scorer results in the dashboard #1182, the scorer view this extends.
• Related: [feat] Judge whether team skills load when they should, without writing a scorer #1469, whose routing questions need these counts.