feat(leakfree): auto-resolve CPCV/horizon embargo to the label horizon (L4488a)

[cipher813]

L4488a — first item of the model-rotation scaffolding arc (L4488); scoring-honesty prerequisite.

The bug

WF_EMBARGO_DAYS defaulted to 0, leaving the CPCV after-test boundary open. The excluded zone around a test group [a,b] is [a − forward_days, b + embargo_days] (leakfree_meta_ic.py:53), so with embargo=0 train rows at b+1 … b+forward_days leak via overlapping labels with the test's late observations. That inflated every leak-free read — and more at longer horizons (60d needs a ~60-row embargo vs ~21 for 21d), making it a real confound for the 60d-vs-21d horizon call. The planned raise was bundled into W1.4, which challenger-first superseded (CF4), so it got orphaned.

The fix

WF_EMBARGO_DAYS default → None = auto = the label horizon (the correct overlapping-label embargo; makes the CPCV exclusion symmetric). Resolved inside leakfree_meta_oos_ic + cpcv_meta_oos_ic + expanding_wf_folds (None → forward_days); the horizon curve inherits per-horizon h automatically since it passes forward_days=h through. An explicit int (incl. 0) still overrides.

• Prod predictor.yaml doesn't set embargo_days → auto applies in production.
• predictor.sample.yaml updated to document the auto behavior (was a misleading explicit 0).
• Note: the proper overlapping-label embargo is ≈forward_days, not the plan's old ~0.01·T (that's only the residual serial-correlation buffer, undersized here).

Effect

The leaderboard's leak-free CPCV/horizon reads become honest — the prerequisite for L4488b (net-of-cost scoring) and L4488d (the horizon call). The current ~0.18 CPCV is expected to drop somewhat (the honest number), and the 60d>21d gap will shrink once both are properly embargoed.

Tests

+1 (None auto-resolves to horizon; explicit 0 honored; cpcv too). Suite 1362→1363.