[Lane 1 · numerics] Precision/Recall/F1 — averaging-mode and zero-division conventions #1224

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

Lane: 1 · Numerics / Research
Skill needed: comfortable reading scikit-learn / torchmetrics source and docs. No Kotlin required, no SKaiNET knowledge required.
Size: s (~2–4 h)
Blocked by: nothing — this is the first task in the feature and unblocks #1226.

What to do

1. Read sklearn.metrics.precision_recall_fscore_support (docs + source) and
   torchmetrics.classification.{Precision, Recall, F1Score} (docs + source).
2. Produce a short table — one row per edge case, one column per library — for:
   • empty batch (no samples at all);
   • a class that appears in neither predictions nor targets (TP+FP+FN = 0);
   • a class that appears only in targets (never predicted) and vice versa;
   • all predictions wrong;
   • ties in the argmax (two logits exactly equal).
     For each: what does the library return for precision / recall / F1 under macro and
     micro, and do the two libraries agree?
3. Answer the three decisions the parent issue leaves open, each as a one-paragraph
   recommendation with a link a reviewer can click:
   • Zero-division: 0.0 (sklearn default) vs. NaN vs. a parameter. The parent
     recommends 0.0 to match Accuracy.compute() on an empty accumulator.
   • Class set for macro averaging: union of classes seen in targets ∪ predictions
     (sklearn) vs. the full class dimension predictions.shape[dim] (which counts
     never-seen classes as F1 = 0). These differ on small batches.
   • Binary mode: does binary mean "report the positive class (index 1) only", and
     should F1Score take a threshold: Float? like Accuracy does, or is binary purely
     averaging = BINARY?
4. Post the table and the three recommendations as a comment on [Feature]: Precision, Recall and F1Score metrics #1222 under
   its Research section. That comment is the deliverable — not code.

Acceptance

• Edge-case table posted on [Feature]: Precision, Recall and F1Score metrics #1222 with links to the exact sklearn / torchmetrics
  doc sections or source lines
• Three explicit recommendations (zero-division, macro class set, binary/threshold)
• A maintainer has replied "adopted" (or amended) so [Lane 2 · kotlin-core] Precision/Recall/F1 — Averaging enum + internal ConfusionMatrixAccumulator #1226 can start from a fixed contract

Notes

The hand-computed reference fixture the Kotlin sub-issues will test against is:
predicted classes [2, 0, 1, 2], targets [2, 1, 1, 0] (3 classes). Expected under
sklearn: per-class precision [0, 1, 0.5], recall [0, 0.5, 1], F1 [0, 0.667, 0.667];
macro P = R = 0.5, macro F1 = 0.444; micro P = R = F1 = 0.5 (= accuracy). Please confirm
those numbers by actually running precision_recall_fscore_support once and paste the
output — that turns a hand calculation into a citable reference.