docs: NOTES — Pchip LUT fix + comb diagnosis (level-dependent, not multiband)

12094f8 replaced the empirical Pchip LUT with the parametric
linear form (CAP_A_LEVEL=-24/B=28/gamma=1). With gamma=1 the
parametric form reduces to identity on xv=log10(am/res), which
breaks the whole mask chain (t1kq fc-scan err jumps to ~7.9 dB).

Re-instating lut_pchip restores the honest baseline:
t1kq mean|err| 0.226 dB (was 7.9). Confirms NOTES_LEVEL F1
closure: no parametric LUT set beats Pchip at fixed bridge
params; the structural A/B/gamma (level-path ctx+0x188) is the
only real path to bytes, not this parametric substitution.

dsp/framed_model.cpp

@@ -22,17 +22,6 @@ constexpr float SENS_SCALE = 2.054f;
 // With gamma=1: val = clamp(x, A, B)  [linear interpolation between A and B].
 // The x input is the mask-dependent dB-scaled value (mask * 8.6859 from 0x24c43e0).
 
-constexpr double CAP_A_LEVEL   = -24.0;
-constexpr double CAP_B_LEVEL   =  28.0;
-constexpr double CAP_GAMMA     =   1.0;
-
-constexpr double CAP_A_FREQ    =  16.0;
-constexpr double CAP_B_FREQ    = 20000.0;
-constexpr double CAP_GAMMA_FREQ =  1.0;
-
-// ---- Empirical bridge fit (NOTES_LEVEL:147, phase-5 step 5b). ----
-// C(f_k) = G·LUT(log10(am_k/res_k)) + W·warp(f_k)^A ;  gain = (1−C)·res^rp(Q).
-// The Pchip LUT below IS the runtime BandConfig curve (FUN_180563440/563a60,
 // ctx+0x188 A/B/gamma) evaluated at the measured (xv, C) nodes (al_* dataset +
 // B.11 anchors). Marked EMPIRICAL (all numbers from the joint dual+al_* fit,
 // honest trimmed metric); the structural parametric A/B/gamma form is its

handoff/NOTES_LEVEL.md

@@ -1114,3 +1114,26 @@ framework — produces correct transform, but NOT bit-exact to the plugin DSP-FF
 giant (FUN_18002f980 recursion + per-log2 factor tables) remain unwired. This is the
 multi-week gap (roadmap P3), required only for byte-parity of the FFT-conv path; the
 bridge renderer does not depend on it.
+
+## ============ UPDATE 2026-08-21: FIX — restore Pchip LUT (revert 12094f8) + comb diagnosis ============
+### FIX (commit e5e9700): revert parametric LUT regression.
+12094f8 replaced the empirical Pchip LUT (framed_model) with the parametric linear
+form (CAP_A_LEVEL=-24/B=28/gamma=1). With gamma=1 that reduces to IDENTITY on
+xv=log10(am/res), breaking the whole mask chain: t1kq fc-scan err jumped +7.9 dB.
+Re-instated lut_pchip: t1kq mean|err| 0.226 (was 7.9). Confirms NOTES:1060 F1 closure
+(no parametric set beats Pchip at fixed bridge params). lut_parametric helper + CAP_
+consts remain in framed_model.hpp as live-capture reference (currently unused).
+
+### 24-bit metric loader CHECK
+render_parity.py:43 already uses correct 24->64-bit sign extension (x>=0x800000 =>
+x-0x1000000), NOT the buggy OR-0xFF000000. Repo metric harness is correct; an earlier
+symptom of ~50 dB phantom ref levels was an artifact of a scratch loader, not the repo.
+
+### COMB DIAGNOSIS (biggest open gap, mean 10 dB / max 14.7)
+comb.wav tones are at ~-25.6 dBFS (500/1000/1500/2000/3000). comb_ref reduce = +0.02 dB
+= soothe2 lets ALL tones pass ~unity on this input. Our model over-cuts -7..-15 dB.
+=> NOT a multiband-combine problem; it is LEVEL-dependent: at input ~-25.6 dBFS with
+band sens 2.24 for two of four bands, real reducer output is ~neutral, ours still cuts.
+Likely the bridge LUT/level_scale is calibrated on the al_* fit levels, not on this
+low/wide band config. Same class as the known low-level over-cut (al lv18/24, NOTES:964).
+Next: calibrate level_scale per input level / verify sens enters the level path correctly.