docs/renders/aqua-feel-native-diff-2026-11-10.md

@@ -69,7 +69,7 @@ I read only that Starfire file's `smooth` definition; I did not render on Starfi
 
 ## The twelve sources that call `si.smooth` / `si.smoo`
 
-The earlier "eleven" was seven patches, three Ding instruments and the larynx graph; the catalog has seven patches besides the thruster **including** the larynx graph, and **four** Ding instruments (`muted-pluck` was missed). Rendered: the thruster at rpm 1200, those seven patches, and one note per instrument (261.6256 Hz, 1.5 s, not the mixed event). Three conditions, each the same Faust text on both sides: as shipped (each side's own stdlib), the old smoother pinned (`nd_smooth(s) = *(1.0 - s) : + ~ *(s)`), and imp(3) pinned. Pinning rewrites `si.smooth` and `si.smoo` in the source text; a smoother called inside another library function is not rewritten, which may account for the `dx7-algorithm-32` residual below. The tract voice rides along as a thirteenth row. Columns are native minus wasm. Raw rows: `aqua-feel-native-diff-smoother-{library,old,imp3}-2026-11-10.json`.
+The earlier "eleven" was seven patches, three Ding instruments and the larynx graph; the catalog has seven patches besides the thruster **including** the larynx graph, and **four** Ding instruments (`muted-pluck` was missed). Rendered: the thruster at rpm 1200, those seven patches, and one note per instrument (261.6256 Hz, 1.5 s, not the mixed event). Three conditions, each the same Faust text on both sides: as shipped (each side's own stdlib), the old smoother pinned (`nd_smooth(s) = *(1.0 - s) : + ~ *(s)`), and imp(3) pinned. Pinning rewrites `si.smooth` and `si.smoo` in the source text; a smoother called inside another library function is not rewritten, the `dx7-algorithm-32` residual below is not that, see `dx7-algorithm-32-native-diff-cause.md`. The tract voice rides along as a thirteenth row. Columns are native minus wasm. Raw rows: `aqua-feel-native-diff-smoother-{library,old,imp3}-2026-11-10.json`.
 
 | source | stdlib as shipped (wasm bundle vs sandbox): max err / rel RMS dB | old smoother pinned | imp3 pinned | ΔLUFS lib / old / imp3 | Δ true peak dB lib / old / imp3 | Δ centroid % lib / old / imp3 |
 |---|---|---|---|---|---|---|
@@ -91,7 +91,7 @@ Largest metric deltas over all sources and the three conditions: |LUFS| 0.017 dB
 
 What the data says:
 
-- With imp(3) pinned on both sides all 13 rows agree to -75 dB or better (Ding instruments -130 to -142 dB, thruster -99.9 dB, tract -109 dB; worst: `advanced-bird-syrinx-graph` -75.4 dB). `dx7-algorithm-32-additive-organ` sits at -79.3 (library), -79.5 (old) and -79.5 dB (imp3), so its residual is not the pinned smoother (a smoother inside a library function, or float rounding; not isolated).
+- With imp(3) pinned on both sides all 13 rows agree to -75 dB or better (Ding instruments -130 to -142 dB, thruster -99.9 dB, tract -109 dB; worst: `advanced-bird-syrinx-graph` -75.4 dB). `dx7-algorithm-32-additive-organ` sits at -79.3 (library), -79.5 (old) and -79.5 dB (imp3), so the pinned smoother does not move it. Cause found 2026-11-12 (`dx7-algorithm-32-native-diff-cause.md`): the native phasor increment rounds one 2^-30 step differently once `ba.time / ma.SR` shares the reciprocal of the sample rate (about -93 dB without it), and `si.smoo` text on the macro params is the rest (-141 dB with both removed).
 - With the old smoother pinned the thruster still differs (-21.2 dB): native's LLVM contracts the recursion to FMA and stalls at 4797.657 instead of wasm's 4796.632. The old form's stall point depends on the arithmetic; the imp(3) form does not.
 - As shipped, nine of the 13 rows differ at waveform level by -16 to -50 dB (thruster, the three syrinx patches, the larynx graph, all four Ding instruments); `advanced-alien-multivalve-graph` (-68 dB), the tract (-52.6 dB) and the two dx7 patches are closer.
 

docs/renders/dx7-algorithm-32-native-diff-cause.md

@@ -0,0 +1,57 @@
+# dx7-algorithm-32-additive-organ: why native vs wasm reads -79.5 dB
+
+**Nothing here was heard.** Offline buffers, 48 kHz, mono, 3 s, no params, same Faust text on both sides. Same toolchain as `aqua-feel-native-diff-2026-11-10.md` (wasm: faustwasm 0.20.0 / libfaust 2.90.19; native: `FaustNativeRender`, libfaust 2.90.20, `-single -ftz 2 -vec -lv 1`, faustlibraries master 3d8c6ef via `AQUASYNTH_FAUST_HOME`).
+
+## Verdict: cause found, two causes, neither a float32 floor and neither a bug in AquaSynth
+
+The -79.5 dB is a real, deterministic phase drift of the sine voices, not noise, not a quiet signal and not a sample shift.
+
+1. **Increment rounding in `os.phasor`, selected by whether `1/ma.SR` is shared** (-79.3 to -93.0 dB when removed). Lowered patches define `time = ba.time / ma.SR;`. With that division in the graph, the native compiler computes the phasor increment as `freq * fround(1/SR)`; the wasm compiler keeps `freq / SR`. For 392 Hz at 48 kHz: `fround(392/48000) = 0.008166667074`, `fround(392 * fround(1/48000)) = 0.008166666143`, one 2^-30 step (9.31e-10) apart. Measured effective increments: wasm 0.008166667074 on both, native 0.008166667074 without the division in the graph and **0.008166666143** with it. The phase then drifts linearly (about 1.3e-4 cycles over 3 s between the sides), so the waveform error grows linearly with time (0 to 1.65e-4 over 0.45 s, then follows the envelope down).
+2. **`si.smoo` library text** on `/macro/drawbar` and `/macro/shimmer` (-93.0 to -141.2 dB when also pinned). This is the known old-form vs `imp(3)` difference. The 2026-11-10 imp3 pin did not show it because cause 1 sat 14 dB above it.
+
+With both removed the patch reads **-141.2 dB (max 1.8e-7)**, the same float floor as the Ding instruments (-130 to -142 dB).
+
+## Evidence (isolating experiments; each changes one thing)
+
+Baseline reproduced through `native-diff.mjs --items smoother`: 1.65e-4 max, -79.27 dB. The signal is not quiet: RMS 0.193, peak 0.686. The last 2 s are digital silence and contribute zero error (73 % exactly-equal samples is that silence, not agreement). The error is zero at nothing but silence and rises linearly while the voices sound; lag test: lag 0 summed squared error 6.3e-5 vs 18.4 at lag +-1 (sums over 144000 samples), so no shift.
+
+| experiment | rel RMS err dB | max err |
+|---|---|---|
+| whole patch, as lowered | -79.3 | 1.65e-4 |
+| filters stripped (`fi.lowpass`/`fi.highpass`) | -79.3 | 1.66e-4 |
+| per voice, no filters: v0 / v1 / v2 / v3 / v4 / v5 | -85.8 / -79.3 / -143.3 / -73.8 / -143.1 / -109.0 | |
+| `sin(2pi*os.phasor(1,392))*0.5` alone | -163.7 | 3e-8 |
+| same sine times `age+1`, a ramp, or any slow signal | -64 to -74 | 1e-4 to 3e-3 |
+| `os.phasor(1,392)` alone | exact (0) | 0 |
+| `os.phasor(1,392) + (ba.time/ma.SR > 100)` (never true) | -33.2 | 1.0 (one wrap sample) |
+| same with `ba.time > 1e9` or `ma.SR > 1e9` instead | exact (0) | 0 |
+| same with `ba.time / 48000.0 > 100` | exact (0) | 0 |
+| effective phasor increment, native, with / without the division | 0.008166666143 / 0.008166667074 | |
+| whole patch, `time = ba.time / 48000.0;` only | -93.0 | 1.78e-5 |
+| whole patch, every `ma.SR` a literal | -93.0 | 1.78e-5 |
+| whole patch, phasor replaced by a float-accumulating one on both sides (library phasor text is fixed point in master, see below) | -84.0 | 9.4e-5 |
+| age literal + `si.smoo` pinned to old form / imp3 | -114.0 / **-141.2** | 1.3e-6 / 1.8e-7 |
+| `si.smoo` pinned to imp3 only (SR still shared) | -79.5 | 1.65e-4 |
+
+Voices 2 and 4 (588 and 980 Hz) are clean because their increments round the same either way; 196, 392 and 784 are not (1176 Hz rounds the same either way). After cause 1 is removed only voice 1 (the one scaled by `patch_param_0`, the smoothed drawbar) stays at -91.5 dB, and voice 5 (shimmer) at -109 dB: cause 2.
+
+Arithmetic check: `Math.fround(392 * Math.fround(1/48000))` = 0.008166666142642498, matching the measured native increment to 10 digits.
+
+## Scope
+
+- Other patches with the same `time = ba.time / ma.SR;` line are not affected the same way: with imp3 pinned and `time` made literal, `advanced-bird-syrinx-graph` stays -75.4, `advanced-human-larynx-graph` -94.2, `dx7-algorithm-8-bright-pair` -94.7, `advanced-alien-multivalve-graph` moves -97.9 to -110.9. Whether a patch is affected depends on whether its phasor increments round differently under division vs reciprocal multiply.
+- `native-diff.mjs --smoother imp3` therefore cannot reach the float floor for patches whose phasor increment is affected; its `-60 dB` limit still holds with the new reading (-79.5 is explained, not a worse mismatch).
+
+## Not verified
+
+- I did not read the generated C++/wasm. "Native uses `freq * fround(1/SR)`, wasm `freq/SR`" is inferred from the measured increments and the arithmetic above (exact match), not from emitted code (the wasm build has no C++ backend; no `faust` CLI here). It is also not separated whether the difference is libfaust 2.90.19 vs 2.90.20, or the `-vec -lv 1` / LLVM path, or the library's phasor text (fixed point in master, which also looks different from what the wasm bundle holds). Settling it: emit the same DSP with one libfaust version's C++ backend and compare the `instanceConstants` lines, or render the wasm side with a 2.90.20 wasm build.
+- Which side is nearer the ideal is not important: both are within one 2^-30 increment step of 392/48000 and in opposite directions (+4.1e-10, -5.2e-10).
+
+## Reproduce
+
+```
+. /var/lib/mind-aqua-self/env-review.sh
+export AQUASYNTH_FAUST_HOME=/var/lib/mind-aqua-self/faust-home-overlay-20261031
+cd tools/aqua-feel && node native-diff.mjs /var/lib/mind-aqua-self/native-diff-bin --items smoother --out "$TMPDIR/out"
+```
+then edit the lowered DSP text on both sides (as the table says) and render with `renderOnce` and `FaustNativeRender`. The experiment scripts are scratch in `$TMPDIR` and not committed.

tools/aqua-feel/native-diff.mjs

@@ -3,7 +3,7 @@
 // Both sides compile against their own stdlib (faustwasm bundles libfaust's; native reads AQUASYNTH_FAUST_HOME's share/faust), so a
 // difference here can be library text, not the host: the report records the identities. The script exits 1
 // on a render error, a length mismatch, or a LIMITS violation. LIMITS (native-compare.js; the evidence for their values is there, and native-diff.test.mjs trips each) bound the differences;
-// the sample-level bound applies only with --smoother imp3, where the largest error seen is -75.4 dB (dx7-32, unexplained, is -79.5) and a one-sample
+// the sample-level bound applies only with --smoother imp3, where the largest error seen is -75.4 dB (dx7-32 is -79.5: phasor increment rounding plus si.smoo text, docs/renders/dx7-algorithm-32-native-diff-cause.md) and a one-sample
 // shift reads about -4 dB. With the library smoother the waveforms legitimately diverge (thruster -8 to -29 dB), so no sample bound.
 // usage: node native-diff.mjs <FaustNativeRender dir|dll> [--tract PATH.dsp] [--out DIR] [--items core|smoother] [--smoother library|old|imp3]
 //   --items smoother: the 12 catalog sources that call si.smooth/si.smoo (thruster at rpm 1200, 7 patches, 4 Ding instruments, one note at 261.6256 Hz for 1.5 s)