[Lane 6 · review] Precision/Recall/F1 — DARC review and @DarcValidated #1235

Sub-issue of #1222 (Precision, Recall and F1Score metrics).

Lane: 6 · Review (DARC sign-off)
Skill needed: a second pair of eyes with ML-metrics and Kotlin literacy. Must not be the implementer of #1226–#1229 or the author of #1234.
Size: s (~2 h)
Blocked by: #1229 and #1234 merged.

What to do

Walk all four DARC phases against the shipped code, per Contributing → DARC: advanced
contribution workflow → Criteria for setting @DarcValidated:

1. D — Document. The three partials have non-empty math / intuition / examples /
   references tags; none are placeholders.
2. A — Assess. Run every documented example against develop; shapes, dtypes and
   numbers must match the prose. Spot-check the zero-division and macro-vs-micro claims
   against the tests. Cross-check the reference fixture (pred [2,0,1,2], target [2,1,1,0]) against a one-off sklearn.metrics.precision_recall_fscore_support run.
3. R — Research. Click every citation in the partials; each must still resolve and
   support the specific claim it backs.
4. C — Code. The implementation matches the documented contract (averaging modes,
   zero-division, dim / threshold semantics, single accumulator per metric).
5. If all four pass, open a PR adding
   @DarcValidated(by = "First Last <you@example.com>", on = "YYYY-MM-DD") to the
   compute() override of Precision, Recall and F1Score. If any fails, send the
   specific item back to its sub-issue — don't annotate.

Acceptance

• Review findings posted on [Feature]: Precision, Recall and F1Score metrics #1222 (pass, or the list sent back)
• @DarcValidated PR merged, or a clear statement of what's blocking it
• [Feature]: Precision, Recall and F1Score metrics #1222 can be closed (all lanes done, ground-truth explicitly deferred to SKEEP-005 / [SKEEP-005]: Ground-truth validation for stateful metrics #1223)

Notes

Precedent to flag, not to solve: @DarcValidated targets FUNCTION and today the KSP
OperatorDocProcessor only reads it off TensorOps functions — so the badge will be
recorded in source but will not render on any generated page for a Metric. Mention
that in the PR so whoever owns the doc pipeline can decide whether Metric
implementations become a second processed category. That is a follow-up, not a blocker.