[Lane 2 · kotlin-core] F1Score metric #1229

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

Lane: 2 · Kotlin core
Skill needed: Kotlin. Read Precision.kt / Recall.kt from #1227 / #1228 first — F1Score composes their formulas over one accumulator.
Size: s (~2–3 h)
Blocked by: #1227 and #1228 (or at least #1226 — F1 only needs the accumulator, but its tests cross-check against Precision/Recall).

What to do

1. Add public class F1Score(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/F1Score.kt:
   • override val name = "f1";
   • one private ConfusionMatrixAccumulator — do not hold a Precision and a
     Recall and call update() on both; that walks every batch twice;
   • compute() = reduce(acc.counts(), averaging) { tp, fp, fn -> per-class F1 } where
     per-class F1 = 2·tp / (2·tp + fp + fn) (algebraically 2PR/(P+R), but this form has
     a single zero-division check: 0.0 when 2·tp + fp + fn == 0).
   • Note that for MACRO, F1 is the mean of per-class F1s — not the F1 of macro-P and
     macro-R. Say so in the KDoc; it is the most common F1 implementation mistake.
2. Factories: public fun f1Score(averaging: Averaging = Averaging.MACRO, dim: Int = -1): Metric
   and public fun binaryF1Score(threshold: Float = 0.5f): Metric.
3. Tests in skainet-backends/skainet-backend-cpu/src/commonTest/kotlin/sk/ainet/exec/metrics/F1ScoreTest.kt,
   reference fixture pred [2,0,1,2], target [2,1,1,0]:
   • macro F1 → 0.4444 (per-class [0, 0.667, 0.667]), and assert it is not
     2·0.5·0.5/(0.5+0.5) = 0.5 — the mean-of-F1s vs. F1-of-means trap;
   • micro F1 → 0.5, equal to micro precision, micro recall and accuracy;
   • binary with threshold on 2-class logits, hand-computed;
   • a class with TP+FP+FN = 0 contributes 0.0; fresh metric compute() is 0.0;
   • accumulation across two update() calls equals one concatenated call;
   • computeForBatch (from Metric.kt) round-trips through reset().
4. ./gradlew :skainet-lang:skainet-lang-core:apiDump and
   ./gradlew :skainet-backends:skainet-backend-cpu:jvmTest --tests "sk.ainet.exec.metrics.*".

Acceptance

• F1Score + f1Score() + binaryF1Score() public and in the API dump
• All test cases above pass on JVM
• Batches are walked once per update()
• PR body states "unit tests only; ground-truth wiring tracked in SKEEP-005 ([SKEEP-005]: Ground-truth validation for stateful metrics #1223)"

Notes

Reference: sklearn f1_score docs — https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html ;
torchmetrics F1Score — https://lightning.ai/docs/torchmetrics/stable/classification/f1_score.html .
Closing this issue is what unblocks the platform-parity lane (#1230–#1233) and review (#1235).