fix: correct EXP-2793 R² values and Trio TDD in Apr-20 reports

- timescale-architecture: CSF R² 0.302→0.305, ISF 0.185→0.195, Basal ~0.05→0.108
- settings-extraction: Trio median TDD 52→60 (per EXP-2807 verified data)

All other claims in the 5 new Apr-20 reports verified against
EXP-2790-2807 JSONs (causal-architecture, aid-author-recommendations,
report-cards reports passed verification).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: bewest

docs/60-research/settings-extraction-report-2026-04-20.md

@@ -82,7 +82,7 @@ ISF×TDD = 4551 rather than 1800.
 
 | Metric | Loop | Trio | OpenAPS |
 |--------|------|------|---------|
-| Median TDD (U/day) | 46 | 52 | 35 |
+| Median TDD (U/day) | 46 | 60 | 35 |
 | Actual Basal % | 15% | 8% | 30% |
 | ISF Extracted (mg/dL/U) | 55 | 98 | 82 |
 | ISF Profile (mg/dL/U) | 49 | 62 | 50 |

docs/60-research/timescale-architecture-report-2026-04-20.md

@@ -56,10 +56,10 @@ BG dynamics behave fundamentally differently by metabolic context:
 
 | Category | AR(2) R² | Hourly R² | Interpretation |
 |----------|----------|-----------|----------------|
-| CSF (meals) | 0.302 | Highest | Carb absorption is structured |
-| ISF (corrections) | 0.185 | Moderate | Insulin action more variable |
+| CSF (meals) | 0.305 | Highest | Carb absorption is structured |
+| ISF (corrections) | 0.195 | Moderate | Insulin action more variable |
 | UAM (unexplained) | 0.099 | Low | Truly unexplained events |
-| Basal (quiet) | ~0.05 | Low | Homeostatic noise |
+| Basal (quiet) | 0.108 | Low | Homeostatic noise |
 
 Using separate models per category instead of a global model **doubles prediction accuracy** (R²: 0.228 → 0.418 at 5-min, EXP-2796).