Drop range upper-bound LVEF readings from the label chain

[duckyquang]

Summary

Closes #75. lvef_upper / rest_lvef_upper are the high end of a reported range, not a point EF — on the canonical run 18 of the 24 cohort rows lvef_upper labelled carry LVEF = 100.0, and its 6 test rows sit at 43.1 MAE against the fused checkpoint vs 10.41 overall (docs/results/baseline_gap.json, fused_mae_by_lvef_source). Those are range-parse artifacts feeding the regression label, so the chain now excludes them. I went with exclude over midpoint: with the upper end pinned at 100 a midpoint is still a bound artifact, not a measurement.

The exclusion is visible, not silent:

• one lvef_candidates CTE does the parse / range-gate / dedupe once over the chain names plus the excluded bounds, with an accepted flag marking chain membership — the cohort path just filters WHERE accepted
• the funnel gains an explicit label-selection stage, so its diff is exactly what the exclusion costs and it shows up in cohort_funnel.* and the CONSORT flowchart instead of being lumped into "no LVEF"
• the LVEF source breakdown reports the dropped studies with excluded=TRUE (their pct / is_fallback stay null so the labelled shares still sum to 100)
• the rule rides in the cohort summary's drop_rules

Result

Nothing rebuilt — this machine has no BigQuery access or embedding cache, so there are no new cohort numbers here. 117 tests pass locally (pytest), ruff check and ruff format --check clean. From E10's measured numbers the rebuild should lose ~24 of the 1,208 manifest rows and move test MAE from ~10.4 toward ~9.6, but that's a prediction from the diagnosis, not a run.

Notes

• Stacked on Real fairness, calibration, and baseline-gap artifacts from the canonical run #74 (base is feat/eval-artifacts), so the diff here is just the one commit; retarget to main once Real fairness, calibration, and baseline-gap artifacts from the canonical run #74 merges. Merge order: Real fairness, calibration, and baseline-gap artifacts from the canonical run #74 → this.
• @kevzho you have the working BigQuery + embedding setup — can you run build_cohort.py → build_echo_hubert_manifest.py → train_probes.py → the evals once this lands? The funnel and breakdown will print the exclusion counts, and the docs restate happens from those artifacts the same way we just did it on Real fairness, calibration, and baseline-gap artifacts from the canonical run #74.
• Left a note on [D07] Paired cohort recovery — 4,103 shared subjects to 1,208 rows #62 about the interaction with the cohort funnel.