fix(behavior): stop scoring frames whose window straddles a bout edge

[gbradham]

What

The rolling window is causal: row i's features summarise frames
i-window+1 .. i. For the first window-1 frames of every bout, that span
still covers the previous behaviour. Cross-validation was asking the model
to name a behaviour from features that mostly describe a different one, then
counting it wrong when it obliged.

Those frames cannot be classified correctly however good the model is. Scoring
them measures where the annotator drew the boundary, not the classifier.

Only the leading edge is affected — the end of a bout is fine, its window
lies wholly inside it.

Impact

On a 34-session cohort with ~25-frame bouts at window 8, these are 28% of all
scored frames. Applying the same rule to a held-out evaluation moved macro F1
+0.022 (0.685 → 0.707) and accuracy +0.022 — performance that was always
there and was being hidden by frames the model had no way to get right.

This shifts reported CV numbers upward. Runs measured before and after are
not directly comparable. The change is toward the honest figure, but it is a
behaviour change and worth knowing when comparing against previously recorded
results.

Training is untouched

Deliberately. A partly-filled window is still a real input the deployed model
will meet at a genuine behaviour transition, and dropping those rows would
shrink the training set for no gain. This is a scoring change only.

n_rows_scored

n_rows_kept counts rows assembled for training, which now differs from what
was measured. The result dict gains n_rows_scored so nobody divides by the
wrong denominator.

One existing test asserted sum(support) == n_rows_kept — that invariant is
precisely what this changes, so it now asserts == n_rows_scored plus
n_rows_scored <= n_rows_kept. Worth a look, since it changes an existing
assertion rather than adding one.

Tests

Ten cases covering the rule directly: leading frames excluded, trailing frames
kept, bouts shorter than the window never scorable, session boundaries not
bridged, frame-number gaps restarting the window, window=1 a no-op,
out-of-order rows handled, and the excluded count equalling window-1
across several window sizes.

2813 passed, 1 skipped; ruff clean.

An intermittent single failure appeared in one full-suite run and did not
recur with a fixed seed. Same pre-existing flake noted on #102, unrelated to
this change.