[D08] lvef_upper fallback injects LVEF=100 label artifacts into the cohort

[duckyquang]

Task metadata

Field	Value
Category	Data
Priority	High
Complexity	Low
Est. days	1
Feeds	#62 (D07), retrain after fix
Found by	#67 (E10) diagnosis

Description

The E10 diagnosis (scripts/diagnose_baseline_gap.py, PR #74) found that the lvef_upper fallback in build_cohort.py's LVEF source chain is feeding label artifacts into the cohort:

• 24 cohort rows take their label from lvef_upper; 18 of them carry LVEF = 100.0, which is not a physiologically real ejection fraction — these look like range-upper-bound parses (e.g. ">55%" or "hyperdynamic" coded as an upper limit), not measurements.
• The 6 such rows in the test split score 43.0 MAE against the canonical fused checkpoint, versus 10.42 overall and 5.17 on lvef_3d-labeled rows.
• Excluding just those 6 rows moves test MAE from ~10.42 to ~9.6 — roughly a full MAE point from 2.4% of the split.

Steps:

1. Look at what lvef_upper actually holds in the source table — if it is the upper bound of a reported range, it should either be excluded from the fallback chain or converted to a range midpoint, not used as-is.
2. Update the LVEF coalesce order in build_cohort.py accordingly, and record the affected row count in the funnel output.
3. Rebuild the manifest and note that M10 probes should be retrained on it (cheap — the full cycle is ~90s on CPU per TECHNICAL 9).

Acceptance criteria:

• A stated decision on lvef_upper (exclude vs midpoint), grounded in what the field actually contains.
• Rebuilt manifest with the change reflected in the funnel counts.
• A note on [D07] Paired cohort recovery — 4,103 shared subjects to 1,208 rows #62, since label sourcing interacts with the cohort-size funnel.