From 33a36bfe68f1ad0063176b41073c9965862588b7 Mon Sep 17 00:00:00 2001
From: Matiq <kokagishi@proton.me>
Date: Tue, 18 Aug 2026 10:06:25 +0300
Subject: [PATCH] =?UTF-8?q?B.7:=20XML=20mode=3D1=20=D0=B8=D1=81=D0=BF?=
 =?UTF-8?q?=D0=BE=D0=BB=D0=BD=D1=8F=D0=B5=D1=82=20case8=20(Q=3D+0x80c,=20f?=
 =?UTF-8?q?var=3Dsens/20),=20=D0=BD=D0=B5=20case1;=20Q=5Feff-=D0=BC=D0=B0?=
 =?UTF-8?q?=D0=BF=D0=BF=D0=B8=D0=BD=D0=B3?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 bandshape.py     |  84 +++++++++++++++++++++++++++++++++
 fit_bandshape.py | 119 +++++++++++++++++++++++++++++++++++++++++++++++
 roadmap.md       |  17 +++++--
 3 files changed, 217 insertions(+), 3 deletions(-)
 create mode 100644 bandshape.py
 create mode 100644 fit_bandshape.py

diff --git a/bandshape.py b/bandshape.py
new file mode 100644
index 0000000..482b66b
--- /dev/null
+++ b/bandshape.py
@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+"""bandshape.py — эмпирическая форма полосы детектора из рендеров soothe2.
+
+Тон 1 кГц (tone1kq.wav / tone1k.wav), одна полоса b1 с on=1, sens=12,
+freq=f_c меняется по серии t1kq_only1_<fc> / t1k_b1f_<fc>.
+Глубина подавления тона при стабильном состоянии = -20log10(|H(f=1k)|) полосы.
+Кривая D(f_c) = форма |H(f_c)| (Q-ширина), нормируется на D(1000).
+
+Результаты: stdout "fc  D_dB  D_norm_dB", а также печать подбора к моделям
+case1 (Q=0.707) и case8 (Q=0.9999978542327881) с хелпером cfa.
+"""
+import numpy as np, wave, glob, os, sys, re
+
+SR = 44100
+
+def read_wav(path):
+    w = wave.open(path, 'rb')
+    sw, nc, n = w.getsampwidth(), w.getnchannels(), w.getnframes()
+    d = np.frombuffer(w.readframes(n), dtype=np.uint8).reshape(n, nc, sw)
+    w.close()
+    ch = d[:, 0, :]
+    if sw == 3:
+        v = (ch[:, 0].astype(np.int64) | (ch[:, 1].astype(np.int64) << 8) |
+             (ch[:, 2].astype(np.int64) << 16))
+        v = (v ^ (1 << 23)) - (1 << 23)
+        return v.astype(np.float64) / (1 << 23)
+    v = (ch[:, 0].astype(np.int64) | (ch[:, 1].astype(np.int64) << 8))
+    v = (v ^ (1 << 15)) - (1 << 15)
+    return v.astype(np.float64) / (1 << 15)
+
+def rms_db(x, t0, t1):
+    s0, s1 = int(t0 * SR), int(t1 * SR)
+    seg = x[s0:s1]
+    if seg.size == 0: return -999.0
+    return 20.0 * np.log10(np.sqrt(np.mean(seg ** 2)) + 1e-12)
+
+def band1_freq(rpp):
+    data = open(rpp, 'rb').read()
+    s = data.decode('utf-8', 'replace')
+    i = s.find('VST3: soothe2')
+    line = s[i:].split('\n', 1)[1]
+    j = line.find('\n      >')
+    import base64
+    b = re.sub(r'[^A-Za-z0-9+/=]', '', line[:j])
+    d = base64.b64decode(b)
+    for m in re.finditer(r'<PARAM id="([^"]+)" value="([^"]+)"', d.decode('utf-8', 'replace')):
+        if m.group(1) == 'band1 freq':
+            return float(m.group(2))
+    return None
+
+def measure_series(prefix, dry_path, freqs, t0=1.6, t1=5.5, tag=''):
+    dry = read_wav(dry_path)
+    rows = []
+    for fc in freqs:
+        rpp = f'/home/m/soothe-bt/{prefix}_{fc}.rpp'
+        wav = f'/home/m/soothe-bt/{prefix}_{fc}.wav'
+        if not (os.path.exists(rpp) and os.path.exists(wav)):
+            continue
+        fcr = band1_freq(rpp)
+        fx = read_wav(wav)
+        n = min(dry.size, fx.size)
+        red = rms_db(dry[:n], t0, t1) - rms_db(fx[:n], t0, t1)
+        rows.append((fcr if fcr else fc, red))
+    rows.sort()
+    if not rows:
+        print(f'[{tag}] no data')
+        return []
+    center = min(rows, key=lambda r: abs(r[0] - 1000.0))
+    for fc, red in rows:
+        print(f'[{tag}] fc={fc:10.3f}  D={red:7.3f} dB  norm={red-center[1]:7.3f} dB')
+    return rows
+
+if __name__ == '__main__':
+    measure_series('t1kq_only1',
+                   '/home/m/soothe-bt/tone1kq.wav',
+                   [800, 900, 950, 980, 1000, 1020, 1050, 1100, 1200], tag='t1kq_only1')
+    print()
+    measure_series('t1k_b1f',
+                   '/home/m/soothe-bt/tone1k.wav',
+                   [500, 800, 900, 950, 1000, 1050, 1100, 1200, 1500, 2000], tag='t1k_b1f')
+    print()
+    measure_series('t1kq_b1f',
+                   '/home/m/soothe-bt/tone1kq.wav',
+                   [500, 800, 900, 950, 1000, 1050, 1100, 1200, 1500, 2000], tag='t1kq_b1f')
diff --git a/fit_bandshape.py b/fit_bandshape.py
new file mode 100644
index 0000000..d99a09d
--- /dev/null
+++ b/fit_bandshape.py
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""fit_bandshape.py — подбор модели формы полосы к измеренной кривой.
+
+Модели (из декомпиляции, /tmp/det/detector_decoded.md):
+  case1 (mode=1 bell): FUN_1805343e0(fs, fc, Q=0.707)
+      w = 1/FUN_181a14cfa(fc*pi/fs);  k = (1/Q)*w;  w2=w*w;  a=1/(k+1+w2)
+      A=[a,2a,a];  B=[1,(1-w2)*2a,(1-k+w2)*a]
+  case8 (RBJ bell+gain): FUN_180533ec0(fs, fc, Q=0.9999978542327881, gain=10^(sens/20))
+      w0=fc*2pi/fs; c=cos(w0); s=sin(w0); p=(c*0.5)/Q
+      alpha=p*gain; alpha2=p/gain
+      A=[alpha+1, s*(-0.5), 1-alpha];  B=[alpha2+1, s*(-0.5), 1-alpha2]
+Передаточная функция H(f) = |(A0+A1 z^-1+A2 z^-2)/(B0+B1 z^-1+B2 z^-2)|, z=e^{j2pi f/fs}.
+Кривая подавления тона f_tone при полосе на fc: D(fc) = -20log10(|H(f_tone)|), норм.
+"""
+import numpy as np
+from bandshape import measure_series
+
+FS = 44100.0
+
+def cfa_candidates(x):
+    return {
+        'sin': np.sin(x),
+        'sqrt': np.sqrt(x),
+        'tan': np.tan(x),
+        'asin': np.arcsin(np.clip(x, -1, 1)),
+        'csc': 1.0/np.sin(x),
+        'cot': 1.0/np.tan(x),
+        'none': x,
+    }
+
+def cfa_fn(name):
+    return cfa_candidates(1.0)[name].__class__ if False else {
+        'sin': np.sin, 'sqrt': np.sqrt, 'tan': np.tan,
+        'asin': lambda x: np.arcsin(np.clip(x, -1, 1)),
+        'csc': lambda x: 1.0/np.sin(x),
+        'cot': lambda x: 1.0/np.tan(x),
+        'none': lambda x: x,
+    }[name]
+
+def h_mag(A, B, f, fs=FS, zsign=-1.0):
+    z = np.exp(zsign*2j*np.pi*f/fs)
+    num = B[0] + B[1]*z + B[2]*z**2
+    den = A[0] + A[1]*z + A[2]*z**2
+    return np.abs(2.0*num/den)
+
+def model_case1(fc, f_tone, Q, cfa, fs=FS):
+    x = fc*np.pi/fs
+    w = 1.0/cfa(x)
+    k = (1.0/Q)*w
+    w2 = w*w
+    a = 1.0/(k+1.0+w2)
+    A = [a, 2*a, a]
+    B = [1.0, (1.0-w2)*2*a, (1.0-k+w2)*a]
+    return h_mag(A, B, f_tone, fs)
+
+def model_case8(fc, f_tone, Q, gain, fs=FS):
+    w0 = fc*2*np.pi/fs
+    c, s = np.cos(w0), np.sin(w0)
+    p = (c*0.5)/Q
+    alpha = p*gain
+    alpha2 = p/gain
+    A = [alpha+1.0, s*(-0.5), 1.0-alpha]
+    B = [alpha2+1.0, s*(-0.5), 1.0-alpha2]
+    return h_mag(A, B, f_tone, fs)
+
+def norm_db(d):
+    d = np.asarray(d, float)
+    return d - d[np.argmax(d)]
+
+def fit(rows, pred_fn, f_tone=1000.0):
+    fc = np.array([r[0] for r in rows])
+    meas = np.array([r[1] for r in rows])
+    pred = np.array([pred_fn(f, f_tone) for f in fc])
+    pred_db = norm_db(-20*np.log10(np.maximum(pred, 1e-12)))
+    meas_db = norm_db(meas)
+    se = np.mean((pred_db - meas_db)**2)
+    rmse = np.sqrt(se)
+    return rmse, pred_db, meas_db, fc
+
+def run(rows, tag):
+    print(f'=== {tag} ===')
+    best = []
+    cfa_fns = cfa_fn
+    # case1
+    for qname, q in [('Q0.707', 0.707), ('Q1.0', 1.0), ('Q0.99999', 0.9999978542327881)]:
+        for cfa_name in ['sin', 'sqrt', 'tan', 'asin', 'csc', 'cot', 'none']:
+            def make(cfa_name, q):
+                def f_(fc, ft):
+                    return model_case1(fc, ft, q, cfa_fns(cfa_name))
+                return f_
+            rmse, p, m, fc = fit(rows, make(cfa_name, q))
+            best.append((rmse, f'case1 {qname} cfa={cfa_name}', p))
+    # case8
+    for qname, q in [('Q0.99999', 0.9999978542327881), ('Q1.0', 1.0), ('Q0.707', 0.707)]:
+        for gain in [10.0**(12.0/20.0), 1.0]:
+            def make(q, gain):
+                def f_(fc, ft):
+                    return model_case8(fc, ft, q, gain)
+                return f_
+            rmse, p, m, fc = fit(rows, make(q, gain))
+            best.append((rmse, f'case8 {qname} gain={gain:.2f}', p))
+    best.sort()
+    print(f'{"model":45s} rmse')
+    for rmse, name, _ in best[:8]:
+        print(f'{name:45s} {rmse:.4f}')
+    print()
+    # best curve vs meas
+    rmse, name, p = best[0]
+    m = np.array([r[1] for r in rows])
+    print(f'best={name} rmse={rmse:.4f}')
+    for i, r in enumerate(rows):
+        print(f'  fc={r[0]:8.1f} meas={m[i]:7.3f} model={p[i]:7.3f} dB')
+    return best
+
+if __name__ == '__main__':
+    rows = measure_series('t1kq_only1',
+                   '/home/m/soothe-bt/tone1kq.wav',
+                   [800, 900, 950, 980, 1000, 1020, 1050, 1100, 1200], tag='t1kq_only1')
+    run(rows, 't1kq_only1')
diff --git a/roadmap.md b/roadmap.md
index f518474..368aeac 100644
--- a/roadmap.md
+++ b/roadmap.md
@@ -334,6 +334,17 @@
 - **Матем. хелперы 0x181a14xxx — IAT-стабы** (`JMP [0x181bab3xx]`), цели 0x6fffff... вне дампа
   (грузит плагин). Семантика `FUN_181a14cfa` не подтверждена: кандидаты sin/sqrt/tan —
   проверяется на харнессе против baseline (corr 0.99475/SNR 19.80/diff −0.065).
-- **ОТКРЫТЫЙ ВОПРОС**: какой case фактически исполняется для band1 при XML mode=1
-  (case1: Q=0.707 фикс vs case8: Q из +0x80c + gain) — нужен замер. Вопрос (c): factory default
-  band mode=8 (0x18052f306), а XML mode=1 — где маппинг.
+- **ОТВЕТ НА ВОПРОС (c): XML mode=1 реально исполняет case8 (mode=8/default), НЕ case1**.
+  Доказательство — серия `dual_b1q_*` (тон 500+2000, band1 на 500, sens=12, mode=1):
+  подавление @2000 (red@500≈10.2 насыщена и q-независима) монотонно убывает с q:
+  q=0.1→+5.00 dB, 0.3→+2.73, 0.5→+1.5, 0.7→+0.9, 1.0→+0.47, 1.5→+0.09, 2.0→−0.04,
+  3.0→−0.06. case1 (фикс Q=0.707) q НЕ читает → исключён. case8 (Q=+0x80c, fvar=0.6=sens/20):
+  delta(2000)=[4.94,1.9,0.85,0.46,0.24,0.11,0.06,0.03] rmse≈0.42, верный знак/направление
+  (завышен при q≥0.5 — модель шире факта). Свободный Q_eff(q): ratio≈0.7–0.95 при q≤1,
+  → рост при q≥1.5 (Δ<0 — вне bell). Структура `FUN_180533ec0` подтверждена.
+  mode4-путь (Q=0.5 фикс, fvar=(q−0.707)/2+1) даёт ОТРИЦАТЕЛЬНЫЙ тренд (delta 0.57→−5.45) —
+  несовместим. Вывод: case8-структура (A=[1+α,−2s,1−α], B=[1+α2,−2s,1−α2], α=p·f, α2=p/f,
+  p=0.5·cos(w0)/Q, fvar≈sens/20 или sin(sens/20)=0.565) — модель формы полосы; fvar-различение
+  sin vs x — на харнессе (Q_eff ratio 0.96/0.70 vs 1.25/0.75 при q=0.1/1.0).
+  Несоответствие формы t1kq_only (D(800)=−0.92 при Q=0.99999; case8 даёт ~0) — полоса на тон
+  шире биквада → учёт WOLA/region-smoothing маски (п. B.4) + Q_eff-маппинг — следующий шаг.