docs: record per-epoch Apple re-run of fallback-excluded FL signal

[wpfleger96]

Appends a dated entry to the decision log in docs/fl-rera-validation.md and rewrites the open question. Documentation only; no analysis or reporting code changes.

The previous entry left a pointer that the fallback-excluded FL signal was more inversely related to Apple breathing disturbance over the full date range (ρ −0.128). Re-running per constant-settings epoch (12 Rx periods ≥10 days, via validate-apple and sweep-thresholds --target apple --include-fallback 1,0 with production knobs pinned) shows:

• The inverse relation belongs to fl_class_ge4_pct specifically, which is already rule-matched-only by construction, so it was never a fallback artefact (9/12 epochs negative; P11 −0.44, P14 −0.72).
• The fallback-excluded RERA proxy has no within-epoch signal (5 neg / 5 pos / 2 undefined, most epochs < 50 events). The full-range −0.128 was an epoch confound and is retracted.
• The two SNORE instruments diverge: the run-based proxy and rera_index lean positive in the two largest APAP epochs while FL prevalence is flat.
• fl_class_ge4_pct vs Apple sleep efficiency leans negative (8/12), so high-FL nights have worse sleep efficiency yet fewer Apple BD — Apple's own signals disagree, making Apple-BD construct mismatch the leading explanation.
• rera_index vs Apple awake seconds is positive in 8/12 epochs (P20 +0.47 p=0.025; P21 +0.35 p=0.043).

The open question narrows to why rule-matched FL prevalence inversely tracks Apple BD while the RERA proxy does not, with the denominator test and Apple sleep efficiency / awake seconds as the next anchors.