[Lane 2 · kotlin-core] Precision metric #1227

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

Lane: 2 · Kotlin core
Skill needed: Kotlin. Read Accuracy.kt and ConfusionMatrixAccumulator first — this task is mostly wiring.
Size: s (~2–3 h)
Blocked by: #1226 (accumulator + Averaging)

What to do

1. Add public class Precision(averaging: Averaging = Averaging.MACRO, dim: Int = -1, threshold: Float? = null) : Metric
   in skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/nn/metrics/Precision.kt:
   • override val name = "precision";
   • update(...) delegates to a private ConfusionMatrixAccumulator(dim, threshold);
   • compute() = reduce(acc.counts(), averaging) { tp, fp, fn -> if (tp + fp == 0L) 0.0 else tp.toDouble() / (tp + fp) };
   • reset() delegates.
     Mirror Accuracy.kt's KDoc style (what the metric means, the dim / threshold semantics, what averaging does).
2. Add factory functions at the bottom of the file mirroring accuracy() / binaryAccuracy():
   public fun precision(averaging: Averaging = Averaging.MACRO, dim: Int = -1): Metric and
   public fun binaryPrecision(threshold: Float = 0.5f): Metric.
3. Tests in skainet-backends/skainet-backend-cpu/src/commonTest/kotlin/sk/ainet/exec/metrics/PrecisionTest.kt,
   copying AccuracyTest.kt's tensor helpers. Cover, with hand-computed expected values
   (see the parent issue and [Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224 for the reference fixture pred [2,0,1,2], target [2,1,1,0]):
   • macro multi-class with Int32 index targets → 0.5;
   • micro multi-class → 0.5 (and assert it equals Accuracy on the same batch);
   • the same batch with soft (FP32 one-hot) targets gives the same result;
   • binary with threshold, 2-class logits;
   • accumulation across two update() calls equals one call on the concatenated batch;
   • zero-division: a class with TP+FP+FN = 0 contributes 0.0 (per [Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224), and
     compute() on a fresh metric returns 0.0.
4. ./gradlew :skainet-lang:skainet-lang-core:apiDump and
   ./gradlew :skainet-backends:skainet-backend-cpu:jvmTest --tests "sk.ainet.exec.metrics.*".

Acceptance

• Precision + precision() + binaryPrecision() public and in the API dump
• All six test cases above pass on JVM
• No duplicated argmax / iteration logic — everything goes through the accumulator
• PR links this issue and [Feature]: Precision, Recall and F1Score metrics #1222; PR body states "unit tests only; ground-truth wiring tracked in SKEEP-005 ([SKEEP-005]: Ground-truth validation for stateful metrics #1223)"

Notes

Precision answers "of what I predicted positive, how much was right". The twin task #1228 (Recall) is the same shape with tp / (tp + fn) — if you take both, fine, but they are filed separately so two people can each ship one.