[Lane 5 · docs] Precision/Recall/F1 — doc partials (math / intuition / examples / references) #1234

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

Lane: 5 · Docs / DARC
Skill needed: AsciiDoc and technical writing; enough ML background to explain precision vs. recall. No Kotlin beyond reading a 30-line class.
Size: s (~2–3 h for all three)
Blocked by: nothing for the math / intuition / references tags (start now); the examples tag needs #1229 so the example actually runs.

What to do

1. Write three partials, one per metric, following the tag structure of
   docs/modules/ROOT/partials/ops/tensorops/matmul.adoc:
   • docs/modules/ROOT/partials/ops/metrics/precision.adoc
   • docs/modules/ROOT/partials/ops/metrics/recall.adoc
   • docs/modules/ROOT/partials/ops/metrics/f1score.adoc
     Each with non-placeholder // tag::math[], // tag::intuition[], // tag::examples[],
     // tag::references[] blocks. Use stem:[…] / [stem] (LaTeX; the playbook sets
     stem: latexmath) for the formulas. The math block must state the per-class
     definition, what macro vs. micro averaging does, and the zero-division convention
     adopted in [Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224.
2. Metrics are not TensorOps functions, so the operator-doc generator will not pick
   these partials up automatically. Surface them by adding a "Precision, recall and F1"
   subsection to docs/modules/ROOT/pages/how-to/metrics-and-perf-testing.adoc that
   include::partial$ops/metrics/f1score.adoc[tag=intuition] (and the others by tag).
3. The examples block must be real code that compiles against develop after [Lane 2 · kotlin-core] F1Score metric #1229 —
   ideally a tagged snippet in skainet-docs-samples included via include::example$…,
   like the existing TrainingDemo.kt accuracy example on that page; a plain
   [source,kotlin] block is acceptable if you say why.
4. Build the site locally (see Contributing → Build from source / the docs.yml
   workflow) and check the math renders.

Acceptance

• Three partials, all four tags non-empty, math renders
• Metrics how-to page shows the new section with the intuition text and a runnable example
• Every reference link resolves and supports the claim next to it (the reviewer in
  [Lane 6 · review] Precision/Recall/F1 — DARC review and @DarcValidated #1235 will click each one)

Notes

Intuition to convey: precision answers "of what I predicted positive, how much was right";
recall answers "of what was actually positive, how much did I catch"; F1 is their harmonic
mean, so it punishes a classifier that is high on one only because the other is near zero —
which is exactly where raw accuracy stays silent (predict the majority class everywhere:
high accuracy, near-zero minority recall). Also spell out that macro F1 is the mean of
per-class F1s, not the F1 of the macro means.

References to start from: sklearn precision_recall_fscore_support, torchmetrics
F1Score, and https://en.wikipedia.org/wiki/F-score .