M6 hot convergence

docs/DECISIONS.md

@@ -1028,3 +1028,18 @@ Consequences: no rate-change delay (latency = the oversampler's: 49 samples FIR
 `harness/reports/M6/20261017_plugin_cpu.md`: default knobs 26.4 % (Half 18.9 %), every knob 10 50.2 % (33.2 %), knob automation
 32–35 % (≈ 24 %); at the owner's NAM-matched settings 47.3 % (Half 31.5 %). Null test 26 / 26; automation (96 kHz, Full, exact J4,
 1e-4 V) 0 non-finite, 20 rescued first-solve failures. The null test and the CPU tool default to Full.
+
+## D-061 EL34: no plate current below the cathode (5 V softplus of the plate voltage; reference model change)
+Date: 2026-10-17 · Status: accepted (owner: "approve the tubes.lib change")
+Context: with the speaker load (D-059) the power amp at HOT had 37–56 first-solve Newton failures per DI clip at 384 kHz
+(more at 768 kHz, 43 unrescued) and the b56 / b6 SPICE references stopped with "timestep too small". Diagnosis
+(PROGRESS 2026-10-17): every failure at pair B's plate port during a flyback (pair A +1100 … +3400 V, pair B −700 …
+−2600 V), where Koren's unclamped atan(Vpk/kvb) gives a reverse plate current and the screen knee switches at Vpk = 0;
+not the iteration budget, tolerance, matrix precision, halvings or the half-step code (every-sample half-steps at fs
+reproduce 2·fs to −67.6 dB).
+Decision: in `tubes.lib` (EL34 Bp, Bs) and `devices::Pentode`, Vpk enters atan through P = max(Vpk, 0) + 5·ln(1 +
+exp(−|Vpk|/5)): the plate current vanishes below the cathode, the cut-off current goes to the screen, and both are smooth
+at 0 V. Above +20 V, P = Vpk within 0.1 V (DC operating point unchanged).
+Consequences: engine diagnostic: HOT DI 0 power-amp failures at 8x and 16x; CRUNCH / DRIVE unchanged (−218 … −224 dB vs
+the previous engine), HOT differs by −35 … −37 dB. tests/data EL34 sweeps regenerated; the pentode test asserts the new
+below-cathode behaviour. SPICE references with the power amp are re-rendered.

docs/PROGRESS.md

@@ -826,6 +826,22 @@ Notes:
 - Owner listening 2026-10-17 (NAM-matched settings, same IR): Power Amp Rate Full vs Half, "can't hear any real audible
   difference". The earlier NAM comparison ("pretty accurate") was made at Half (the restored instance setting).
 - Owner 2026-10-17: keep Full as the default (D-060 confirmed after the listening test); Half stays selectable per instance.
+- Owner 2026-10-17: fix the HOT convergence with the speaker load. Plan: (1) check the half-step handling of the speaker
+  inductors (every-sample half-steps at fs must reproduce 2·fs without half-steps, D-043); (2) characterise the failing
+  power-amp solves (ports, residuals, iterations); (3) fix (Newton step control / fallback / discretisation) and re-verify
+  HOT, the plugin model check and the acceptance.
+- Diagnosis: every-sample half-steps at 8x reproduce 16x to −67.6 dB (half-step code correct); failures grow with the step
+  rate (384 kHz: 1, 768 kHz: 34–37) and are unchanged by maxIter 64, tolerances 1e-6 … 1e-8, double-double matrices or 4 / 8
+  halvings (8: stall). All at pair B's plate in flybacks below the cathode (−700 … −2600 V): Koren's unclamped
+  atan(Vpk/kvb) gives reverse plate current there. Hard clamp: mixed; 5 V softplus with a matching screen knee: 0 failures
+  at 8x and 16x. Owner approved the tubes.lib change (D-061).
+- D-061 verification (new references): M4 full acceptance (768 kHz) 107 PASS; 384 kHz exact J4 107 PASS; HOT 20 PASS / 1 FAIL:
+  DI ESR −40.3 … −41.3 dB (D-059 alone −33.5 … −35.7), Newton failures 0 / 0 / 0 / 2 (rescued; the FAIL); b56 106 PASS; b6
+  96 PASS / 1 FAIL (the known 5 kHz 20 V Newton median, D-043); b5open 88 PASS. The b56 / b6 SWP-lo-0.1V references still stop
+  with "timestep too small" at RELTOL 1e-8 (node y / the recorded nScr source, same times as before D-061: not the tube
+  model) and are now rendered once more at RELTOL 1e-7 by `render_spice` (as the M3 cascade), noted in the reports. Plugin
+  model check: Full 2x HOT −24.2 dB (PASS), unseen −20.8 dB, 94 first-solve failures (2 unrescued) over 28 runs, nearly all in
+  the extreme validation settings (presets: 0 / 0 / 2); Half 2x HOT −18.9 dB. Presence knob jump −62.9 dB; null test 26 / 26.
 
 ## Questions for the owner
 _(Claude Code: add questions here and also ask them in the session.)_

docs/WDF_DESIGN.md

@@ -80,17 +80,21 @@ Vs'  = max(Vsk, 1e−3)
 u    = kp · (1/mu + Vgk / Vs')
 E1   = Vs' · softplus(u) / kp
 F    = E1 > 0 ? 2 · E1^ex : 0
-Ip   = (F / kg1) · atan(Vpk / kvb)
-Is   = F · ( 1/kg2 + (1.5707963 − atan(max(Vpk, 0) / kvb)) / kg1 )    (tubes.lib's literal for π/2)
+P    = max(Vpk, 0) + 5 · ln(1 + exp(−|Vpk| / 5))                      (5 V softplus of Vpk, D-061)
+Ip   = (F / kg1) · atan(P / kvb)
+Is   = F · ( 1/kg2 + (1.5707963 − atan(P / kvb)) / kg1 )              (tubes.lib's literal for π/2)
 Ig   = diode as in 3.1 with Rs = 1000 Ω + 1 Ω
 
 ∂E1/∂Vgk = sigmoid(u)
 ∂E1/∂Vsk = softplus(u)/kp − sigmoid(u) · Vgk / Vs'          (0 when Vsk < 1e−3)
 ∂F/∂x    = 2 · ex · E1^(ex−1) · ∂E1/∂x
-∂atan(Vpk/kvb)/∂Vpk = (1/kvb) / (1 + (Vpk/kvb)²)
+∂P/∂Vpk = logistic(Vpk / 5)
+∂atan(P/kvb)/∂Vpk = (∂P/∂Vpk / kvb) / (1 + (P/kvb)²)
 ```
-Note: as in SPICE, `atan(Vpk/kvb)` is not clamped in the plate equation, so Ip would go negative if the
-plate fell below the cathode. Replicate SPICE exactly; log it if it ever happens in the harness.
+Note (D-061): Koren's form used atan(Vpk/kvb) unclamped, so Ip went negative below the cathode, and the screen knee
+switched at Vpk = 0. With the speaker load (D-059) the plates fly far below the cathode at high Master and the power-amp
+Newton failed there; the 5 V softplus gives no plate current below the cathode, smooth through 0 V, and is equal to
+Vpk within 0.1 V above +20 V. SPICE (tubes.lib) uses the same expression.
 
 Matched-pair equivalent (B6): multiply Ip, Is, Ig and all their derivatives by 2.
 

engine/src/devices/Pentode.h

@@ -19,12 +19,14 @@ struct PentodeCurrents {
 /// Koren pentode with this repository's screen-current modification, as in reference/spice/tubes.lib
 /// (subckt EL34, WDF_DESIGN §3.2):
 ///   Vs' = max(Vsk, 1 mV), u = kp·(1/mu + Vgk/Vs'), E1 = Vs'·softplus(u)/kp, F = 2·E1^ex,
-///   Ip = (F/kg1)·atan(Vpk/kvb),  Is = F·(1/kg2 + (1.5707963 − atan(max(Vpk, 0)/kvb))/kg1),
+///   P = s(Vpk) = max(Vpk, 0) + 5·ln(1 + exp(−|Vpk|/5)) (5 V softplus, D-061),
+///   Ip = (F/kg1)·atan(P/kvb),  Is = F·(1/kg2 + (1.5707963 − atan(P/kvb))/kg1),
 ///   Ig = grid diode (Rs = rgi + 1 Ω).
-/// As in SPICE, atan(Vpk/kvb) in Ip is not clamped (Ip < 0 below the cathode). The 1 GΩ Re1 / Rcp
+/// No plate current below the cathode (it goes to the screen), smooth at Vpk = 0 (D-061). The 1 GΩ Re1 / Rcp
 /// helpers, the diode's CJO / TT and the inter-electrode capacitances are not part of the device.
 /// `count` scales every current and derivative: 2 = a matched pair with identical drive (SPEC §4.2 B6).
 struct Pentode {
+    static constexpr double kPlateSoft = 5.0;  // V, plate softplus scale (D-061)
     double mu = 14.77, ex = 1.046, kg1 = 194.3, kg2 = 2243.0, kp = 53.82, kvb = 40.43;  // D-058 (was Koren 11 / 1.35 / 650 / 4200 / 60 / 24)
     GridDiode grid{1e-9, 1.0, 1000.0 + 1.0};
     double count = 1.0;
@@ -45,20 +47,23 @@ struct Pentode {
             const double df_dvgk = df_de1 * sg;
             const double df_dvsk = clamped ? 0.0 : df_de1 * (sp / kp - sg * vgk / vs);
 
-            const double x = vpk / kvb;
+            // P = 5 V softplus of Vpk (overflow-free form), dP/dVpk = logistic(Vpk/5) (D-061)
+            const double e = std::exp(-std::fabs(vpk) / kPlateSoft);
+            const double pv = std::fmax(vpk, 0.0) + kPlateSoft * std::log1p(e);
+            const double dpv = vpk >= 0.0 ? 1.0 / (1.0 + e) : e / (1.0 + e);
+            const double x = pv / kvb;
             const double a = std::atan(x);
-            const double da = 1.0 / (kvb * (1.0 + x * x));
+            const double da = dpv / (kvb * (1.0 + x * x));
             c.ip = f / kg1 * a;
             c.dip_dvgk = df_dvgk / kg1 * a;
             c.dip_dvsk = df_dvsk / kg1 * a;
             c.dip_dvpk = f / kg1 * da;
 
-            const double knee = vpk > 0.0 ? a : 0.0;
-            const double sfac = 1.0 / kg2 + (1.5707963 - knee) / kg1;
+            const double sfac = 1.0 / kg2 + (1.5707963 - a) / kg1;
             c.is = f * sfac;
             c.dis_dvgk = df_dvgk * sfac;
             c.dis_dvsk = df_dvsk * sfac;
-            c.dis_dvpk = vpk > 0.0 ? -f * da / kg1 : 0.0;
+            c.dis_dvpk = -f * da / kg1;
         }
         if (count != 1.0) {
             for (double* p : {&c.ig, &c.is, &c.ip, &c.dig_dvgk, &c.dis_dvgk, &c.dis_dvsk, &c.dis_dvpk, &c.dip_dvgk,

harness/render_spice.py

@@ -29,6 +29,9 @@ def select_signals(block, preset, pattern=None):
     return [s for s in sigs if pattern is None or pattern in s.id]
 
 
+RETRY_RELTOL = 1e-7  # one retry after "timestep too small" (the M3 cascade's P56_RETRY_RELTOL)
+
+
 def render(block, preset, signals, jobs=None, log=print):
     """Returns {'op': {...}, 'ac': (f, h) or None, 'tran': {signal id: {node: array}}, 'recorded': {signal id:
     {name: array}}, 'notes': [...]}. Blocks driven by recordings (b5open, b6) get them from their source runs
@@ -46,7 +49,13 @@ def one(sig):
         t0 = time.time()
         rec, notes = recordings.recordings(block, knobs, sig)
         x = block_input(block, sig) if block.input_node else None
-        res = spice.tran(text, knobs, x, nodes, recorded=rec or None)
+        try:
+            res = spice.tran(text, knobs, x, nodes, recorded=rec or None)
+        except spice.SpiceError as e:  # as the M3 cascade (render_p56): one retry at RELTOL 1e-7, noted in the report
+            if 'Timestep too small' not in str(e):
+                raise
+            res = spice.tran(text, knobs, x, nodes, recorded=rec or None, reltol=RETRY_RELTOL)
+            notes = list(notes) + [f'rendered at RELTOL {RETRY_RELTOL:g} (ngspice: timestep too small at {config.SPICE_RELTOL:g})']
         if block.engine_supplies:  # the engine block runs on this run's own supply waveforms (b56)
             rec = {n: res[n] for n in block.engine_supplies}
         log(f'  spice {block.id}/{preset.id}/{sig.id}: {len(sig.x) / config.FS:.3f} s signal, {time.time() - t0:.1f} s')