fix(training): repair W2 residual-momentum leak-free read + leakfree starvation observability (L4469 CF3)

[cipher813]

What

Closes the two W2/W1 observe-read defects called out in ROADMAP L4469 CF3 ("fix the two broken observe reads"). The reads are OBSERVE-mode (gate nothing) but block the W2 promotion verdict and muddy the leak-free battery the champion/challenger arc reads.

Bug B — the real crash (meta_trainer.py W2 read)

The residual-momentum standalone leak-free read re-ran the scorer on X_resid_mom_raw[canonical_finite_mask]. X_resid_mom_raw is full-training length N; canonical_finite_mask is OOS length → boolean-index shape mismatch (IndexError) → the read fell into status="error" every run.

Fix: each OOS row already stores residual_momentum_score — the same scorer output, computed leak-free on the test slice at fold time (~L1516, predict_array(X_resid_mom_raw[te])[local_idx]). Pull the stored per-row value (mirrors the W2.3 factor_momentum_ratio idiom just below it): identical in value, correctly OOS-aligned, matched row-for-row to meta_y / _rm_dates.

Bug A — not a defect, made diagnosable (leakfree_meta_ic.py)

The single-path leak-free meta IC returns n_folds=0 while canonical-label coverage is thin — an expanding purged+embargoed WF needs ~min_test + forward_days + test_size unique dates to clear even one fold, and CPCV is the data-efficient working read meanwhile. The graceful insufficient_folds sentinel already prevents a crash. This adds n_unique_dates to both returns so the manifest shows starvation vs a real defect instead of leaving it to be guessed. No fold-forcing (no-shortcuts default) — self-heals as coverage grows (~Aug).

Why this way

• Reuses the already-computed leak-free predictions rather than recomputing — SOTA/no-bandaid; the stored value is provably identical.
• Refuses to manufacture folds from insufficient data; surfaces the cause instead. The next manifest will confirm whether n_folds=0 is pure date-starvation.

Tests

• test_starvation_reports_unique_date_count — n_unique_dates present on the starved path.
• TestW2LeakfreeReadShape — documents the full-length-array-masked-by-OOS-mask IndexError and asserts the stored-row idiom is OOS-aligned.
• test_leakfree_meta_ic.py + test_meta_trainer_residual_momentum_gate.py: 26 passed.

OBSERVE-only; gates nothing; live ensemble + inference path unchanged.

🤖 Generated with Claude Code