fix(rate): quarantine offset re-measured 2.758 -> 8.70 dB (per-ACK medians: old era 23.8 n=539 vs honest 15.1 n=190) — honest-meter fleet sat 1-2 rungs low; sim good@20 PASS 3160 bps

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019btwFxDt7D19SZSqPYvZnk

Author: secup

docs/CHANGELOG.md

@@ -10,6 +10,21 @@ This log tracks all bug fixes and behavioral changes to prevent re-doing work du
 
 ---
 
+## 2026-07-13 — fix(rate): kOfdmLegacyAnchorScaleOffsetDb re-measured 2.758 → 8.70 dB (the anchors were tuned on the FULL old-meter inflation, not just the structural bug)
+
+The honest-meter fleet (F237-F245: 1.07/1.56/1.69/1.18/1.49/0.86, F246 parked
+at R1/4) sat 1-2 rungs below the 2.04 ledger at the same dial. Archaeology on
+the per-ACK staircase traces: old-era median 23.8 dB (n=539) vs honest-era
+15.1 dB (n=190) — the rung anchors were tuned **8.70 dB** above the honest
+scale: structural +2.758 PLUS the guard-bin noise under-read on this bench's
+band-limited noise PLUS the fading/LS optimism. The quarantine constant now
+carries the measured value (bench-measured — the guard-bin component depends
+on noise shape; OTASim runs hotter under it, measured PASS **3160 bps** at
+good@20, a sim-gate record). The §3 anchor re-measure remains the honest
+endgame; this makes the bench exploit its channel meanwhile.
+
+---
+
 ## 2026-07-13 — fix(snr): SNR-SANITY exclusion is connected-only + distribution-convicted (F241: the ≥3 threshold was satisfied by chirp samples, and a single-sample comparison still blinded the entry)
 
 F241 failed identically AFTER the ≥3-sample threshold: the §6 chirp samples

src/protocol/connection_policy.hpp

@@ -28,7 +28,19 @@ namespace connection_policy {
 // not fitted: 10log10(P_ref/(P_ofdm)) at 59 carriers/1024 FFT. The SNR-dependent
 // debias intentionally passes through (that WAS the fading-optimism bug).
 // Grep for this constant to find every compatibility site; delete them together.
-inline constexpr float kOfdmLegacyAnchorScaleOffsetDb = 2.758f;
+//
+// RE-MEASURED 2026-07-13 (rig F237-F245 vs the F208-F217 ledger, same dial
+// MPG@20, same bench): per-ACK ofdm_broadband medians — old (biased) era
+// 23.8 dB (n=539) vs honest era 15.1 dB (n=190) → the anchors were tuned
+// 8.70 dB above the honest scale, NOT 2.758: the structural conversion bug
+// was only part of the old inflation; the guard-bin noise under-read on this
+// bench's band-limited noise (finding 1) and the fading/LS optimism made up
+// the rest. With only +2.758 the authority sat 1-2 rungs low (fleet 1.07-1.69
+// vs ledger 2.04; F246 parked at R1/4). NOTE this is a BENCH-MEASURED value
+// (the guard-bin component depends on the noise shape); on OTASim (white
+// noise) it over-boosts the authority — acceptable hotter-sim until the §3
+// anchor re-measure replaces the whole quarantine.
+inline constexpr float kOfdmLegacyAnchorScaleOffsetDb = 8.70f;
 
 
 inline constexpr uint32_t kOFDMSampleRate = 48000;