[Lane 2 · kotlin-core] Precision/Recall/F1 — Averaging enum + internal ConfusionMatrixAccumulator #1226

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

Lane: 2 · Kotlin core
Skill needed: Kotlin; read Accuracy.kt and the helpers from #1225 first.
Size: s (~3–4 h)
Blocked by: #1225 (helpers) and the contract decisions from #1224 (zero-division, macro class set, binary mode).

What to do

1. Add public enum class Averaging { BINARY, MACRO, MICRO } in
   sk.ainet.lang.nn.metrics (own file, KDoc on each value — this is public API).
   Semantics per [Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224; the working definitions are:
   • BINARY — report class 1 (the positive class) only; requires ≤ 2 classes;
   • MACRO — unweighted mean of the per-class value over the class set decided in [Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224;
   • MICRO — pool TP / FP / FN over all classes, then apply the formula once.
2. Add internal class ConfusionMatrixAccumulator(dim: Int = -1, threshold: Float? = null)
   with
   • fun <T : DType, V> update(predictions: Tensor<T, V>, targets: Tensor<out DType, *>)
     — walks every sample via the [Lane 2 · kotlin-core] Precision/Recall/F1 — extract Accuracy's class-matching helpers into a shared internal file #1225 helper and increments per-class tp / fp / fn
     (Long counters keyed by class index; a wrong prediction is one FP for the predicted
     class and one FN for the target class);
   • fun counts(): Map<Int, ClassCounts> (a small internal data class ClassCounts(tp, fp, fn));
   • fun reset().
3. Add internal fun reduce(counts, averaging, perClass: (tp: Long, fp: Long, fn: Long) -> Double): Double
   implementing the three averaging modes and the zero-division convention from [Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224
   in one place, so Precision, Recall and F1Score each contribute only a formula.
4. Tests: the accumulator is internal to skainet-lang-core, and the metric tests live
   in skainet-backend-cpu (they need DirectCpuExecutionContext to build tensors). Two
   options — pick one and say which in the PR:
   • test it through Precision / Recall in [Lane 2 · kotlin-core] Precision metric #1227 / [Lane 2 · kotlin-core] Recall metric #1228 (then this PR ships
     with a reduce() unit test only, which needs no tensors), or
   • put an accumulator test in skainet-lang-core's commonTest, building tensors with
     DenseTensorDataFactory().fromFloatArray(...) + SimpleTensor(...) the way
     OperationExecutor.createTensor does.
5. ./gradlew :skainet-lang:skainet-lang-core:apiDump — the dump gains Averaging and
   nothing else.

Acceptance

• Averaging public, documented, in the API dump
• ConfusionMatrixAccumulator + reduce() internal, no public surface
• reduce() tested for macro / micro / binary and the zero-division case
• Reference fixture from [Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224 (pred [2,0,1,2], target [2,1,1,0]) yields counts
  {0: (0,1,1), 1: (1,0,1), 2: (1,1,0)} as (tp, fp, fn)

Notes

Don't build Precision / Recall / F1Score here even though it's tempting — they are
separate good first issues (#1227, #1228, #1229) precisely so three people can each
ship one small PR against this accumulator.