@denaudio/den 0.0.0-stage → 0.1.0

package/docs/delay-fx.md

@@ -0,0 +1,202 @@
+# Stereo delay FX — GEN-618 / GEN-620
+
+One `delayFx` unworklet subgraph composes the existing readhead, low-pass and LFO.
+It has no loader, parameter registry, state codec, routing framework, extra effect
+engine or FFT. Call `tick(left, right, controls)` once per sample. Audio and numeric
+controls are finite f32 nodes; sync/bypass/reset are boolean nodes. Duplicate mono
+input explicitly. State is independent per engine/channel; there is no crossfeed.
+
+Configuration is immutable: `sampleRate` (integer 8000–192000, from processor ctx),
+`maxDelaySeconds` (default 8), `tone` (`lowpass`, default, or `flat`) and
+`stereoPhaseCycles` (default 0.5, wrapped modulo one). Eight seconds is a provisional
+construction default covering a whole note at 30 BPM, not a product ceiling.
+Allocation is two readhead buffers of `ceil(rate * capacity) + 2` f32 samples, plus
+scalar filter/LFO/control state. At 96 kHz the default buffer storage is 6,144,016
+bytes; callers own the memory budget. No audio-thread resize occurs.
+
+## Controls and sound
+
+| Control | Unit / policy |
+| --- | --- |
+| timeLeftSeconds / timeRightSeconds | Independent base delays when sync is false; minimum one sample, maximum construction capacity. |
+| sync, bpm, beatsLeft / beatsRight | Sync uses `60 * beats / bpm`; one beat is a quarter note. BPM domain 30–300; positive fractional beats express subdivisions, triplets and dotted values. No independent beat-count product ceiling; actual time must fit capacity. |
+| feedback | Linear multiplier, clamped to [0,0.95]. Positive repeats only. |
+| cutoffHz | Feedback tone cutoff clamped to [20,min(20000,0.24*rate)]; Q is fixed at 0.5. Flat mode bypasses tone in the feedback path. |
+| mix | Linear dry/wet crossfade, clamped to [0,1]. No equal-power boost, limiter or normalization. |
+| rateHz / depthSeconds | Shared LFO module, rate clamped to [0,20], depth [0,0.05]. Left reset phase 0, right reset phase stereoPhaseCycles. Both initialized on the first tick. |
+| bypass | Unity dry, stop feeding new input, keep the internal feedback/tone/LFO running. Unbypass may reveal remaining tail. |
+| reset | Clear logical delay history/pointers, filter history and LFO phases at this sample; new input at this sample begins new history. Dry contribution remains audible. Held reset silences wet. |
+
+A read happens before the current input write. Per channel,
+`wet[n] = delayRead(history, time[n])` and
+`write[n] = admittedInput[n] + feedback[n] * tone(wet[n])`.
+There is no additional sample in the feedback loop. The first repeat is unfiltered;
+the tone shapes each subsequent circulation. Input stop (zeros) preserves decay.
+Bypass and rejected timing admit zero new input. Reset is independent of bypass.
+
+Time, tempo and stereo-tap edits move the existing readheads immediately. There
+is no crossfade or smoothing: abrupt changes can click, skip/repeat source samples,
+and cause Doppler pitch motion. LFO modulation uses the same moving-head behavior;
+right phase 0.5 gives opposite motion, while phase 0 and equal taps preserve dual
+mono. Dry/wet or bypass steps can also click. These discontinuities are explicit
+policies, not a claim of sound approval. No arbitrary audio-rate modulation
+anti-alias guarantee is made. GEN-621's two musical settings remain separate.
+
+## Capacity rejection and clipping
+
+Both requested base times must fit the f32-representable bounds. In sync mode BPM
+must also be in range. A bad stereo request (including a too-short/nonpositive
+base time) sets `timingRejected=true`, emits unity dry on both channels and admits
+zero new input. Internally, both previous accepted base times are retained so the
+old tail continues. On reset, rejected times fall back to one sample with cleared
+history. A later valid request is accepted immediately with the same moving-head
+transition policy. There is no silently clamped rhythmic playback.
+
+LFO depth is applied after an accepted base time. Final modulation saturates at
+the buffer bounds and sets `modulationClipped=true` when clipping occurs. This
+flag is separate from invalid base timing. Consumers should display rejection
+and clipping using existing unworklet events/state if desired; no new transport
+is supplied here. Runtime sample-rate-specific compilation is unchanged: browser
+acceptance is 48 kHz only; 44.1/48/96 kHz offline coverage is separate.
+
+## Feedback bound, including cutoff modulation
+
+A fixed-frequency low-pass magnitude bound alone does not prove a time-varying
+loop stable. For the shared SVF with Q=0.5, let `g=tan(pi*cutoff/rate)`, and express
+its stored state in coordinates `p=band+low`, `q=low`. The same equations are:
+
+```
+u = (p + g*x)/(1+g)
+y = (q + g*u)/(1+g)
+pNext = r*p + (1-r)*x
+qNext = r*q + (1-r)*u
+r = (1-g)/(1+g)
+```
+
+For `0<g<1`, every expression is a convex combination. The chosen 0.24*rate cap
+keeps g below one with margin for the shared coefficient approximation. This
+property holds under sample-by-sample cutoff changes, unlike a stationary-only
+frequency response argument. Linear readhead interpolation is also convex, and
+independent stereo channels do not add energy to each other. If input magnitude
+is at most A and history starts clear, the common amplitude invariant is
+`A/(1-0.95) = 20*A` for delay storage, tone coordinates and wet output, apart from
+floating-point rounding. Thus arbitrary admitted time/cutoff modulation cannot
+create unbounded self-oscillation under these control bounds. Feedback <1, finite
+capacity and strictly positive bounded cutoff also give eventual decay after
+input stops; the tests include the slowest-cutoff/highest-gain tail. There is no
+threshold tail kill added by den; pinned unworklet flushes tiny values (<1e-30).
+Flat tone has the same loop bound without the filter stage. At DC, unity input
+can accumulate to 20; users must gain-stage the output. No hidden compensation
+conceals this behavior.
+
+## State and composition cost
+
+Readhead history remains a logical reset, not a physical memory wipe. Persistent
+slots are inherited readhead/filter/LFO state plus `first`, `acceptedLeft` and
+`acceptedRight`. Restore is supported only for the same schema, configuration and
+rate, using unworklet's existing snapshot path after parameters have rendered.
+
+Seven transient f32 scalar slots and one transient boolean materialize LFO
+outputs, final times, wet reads, cutoff and timing validity within each sample.
+The accepted-time state is read after its current-sample write, avoiding repeated
+expansion of the same selection. They are overwritten before use and add no sample latency.
+They prevent recursive expansion of the composite expression graph in pinned
+unworklet 0.4.1; no compiler/runtime replacement or upstream edit is introduced.
+
+## Evidence
+
+`tests/delay-fx.spec.ts` compares against an unbounded written-input timeline,
+independent direct-form bilinear biquad and (for time-varying cutoff) two one-pole
+cascade coordinates. It does not copy the circular storage or SVF recurrence.
+It checks repeat positions/amplitudes, fractional taps, gain bounds, filtering,
+bypass/reset/restart, stereo/mono/mix endpoints, tempo acceptance/rejection,
+modulation clipping, discontinuity policy and same-schema continuation at all
+three rates. Ordinary references allow 3e-6 absolute error, chorus/variable tone
+4e-6. Wrong waveforms fail. Extreme LFO-depth/cutoff modulation additionally checks
+the derived common amplitude invariant, separately from sample comparisons to
+avoid treating tiny sine-polynomial errors amplified by seconds-to-samples
+conversion as filter recurrence errors. Each long-tail test runs 2^20 samples
+(10.9–23.8 seconds depending on rate), with bounded peak and final tail <1e-6.
+
+`tests/delay-fx-packed.test.mjs` installs the actual tarball into an isolated locked
+consumer, strictly checks declarations, compares rendered feedback/tone against
+a separate direct-form reference, builds the standard unworklet browser path and
+exercises mix/bypass/reset/capacity rejection and snapshot restore at 48 kHz.
+
+The sustained browser gate records 2^19 frames (10.92 seconds) downstream of the
+worklet: actual input plus left/right output, with four seconds of sustained
+220-Hz input and more than six seconds of tail. `capture.ts` is a fixed test fixture
+built with existing unworklet buffers, audio ports and snapshots. It writes three
+channels on the audio thread, emits silence, and reads its snapshot after suspend.
+A guard quantum protects the requested prefix after recording fills. The frame
+counter must cover the entire requested prefix. No custom loader, transport or
+capture framework is added; the product processor is unchanged.
+
+Every captured output sample is compared to an independent reference driven by
+captured input at the unchanged 6e-6 threshold. Input continuity retains its 1e-6
+sine-recurrence threshold and exact four-second span. A ten-second native source
+control always runs, including its full sustained span; all three captured control
+channels must be identical. There are no passing-retry selection or skipped
+startup samples. Raw PCM and source/hash manifests are retained. This verifies
+graph-sample continuity, not hardware-output delivery or scheduler deadlines.
+
+The original ScriptProcessor apparatus produced false discontinuities. In one
+simultaneous recording, internal input was continuous and internal outputs matched
+the independent reference (left 7.45e-9, right 0), while external capture first
+diverged at sample 73728 and lost 640 source samples. A separate internal-only run
+had the same reference accuracy. Another saved tail failure started exactly at
+capture block 393216: shifting the reference by 128 samples restored right-channel
+bit equality and left error below 5.83e-11. The source was already silent, so a
+short native-source control could not diagnose that tail capture loss. This is
+why the baseline now sustains for ten seconds. See `delay-fx-capture-findings.json`
+for numerical results and hashes; raw evidence and diagnostic fixtures are kept
+in `artifacts/delay-fx/capture-investigation/`. The original failures are preserved,
+not converted into passing comparisons by alignment or tolerance changes.
+
+Deadline safety remains a separate unresolved question. Real AudioWorklet traces
+have included startup and non-startup CPU/wall-time budget exceedances even after
+expression materialization; replacing the recorder does not clear those findings.
+No listening approval is inferred from corrected waveform capture.
+
+`performance.mjs` separately measures the existing driver with byte-identical
+packed browser WASM. It records size/hash, controls, warm-up, mean/p99/max block
+time and total wall time for 12 seconds of audio. This host-dependent direct-WASM
+diagnostic excludes browser scheduling/copy overhead and does not replace the
+real-time gate. After the composition audit, WASM decreased from 62,653 to 45,935 bytes. Local
+direct-driver mean decreased from 0.366 to 0.071 ms per 128 frames. The latter run
+still had a 4.677 ms maximum versus a 2.667 ms budget; average cost alone does not
+establish real-time readiness. Against commit 6cf15d1, all four output/status
+channels were bit-identical over 18 cases: 44.1/48/96 kHz, flat/lowpass, and LFO
+rate/depth pairs 0/0, 4 Hz/2 ms, 12 Hz/20 ms. Each case rendered 16384 samples with
+sample-varying time, cutoff, gain, mix, sync, bypass and reset. This equivalence
+check supplements the independent numerical references; it is not a golden. That
+composition audit changed no shared module or unworklet source; no general
+real-time performance guarantee is made.
+
+The subsequent main integration includes the filter materialization described in
+`filter-materialization.md`. It preserves the persistent schema and tiny-value
+behavior while reducing this consumer's WASM to 24,707 bytes (SHA-256
+`74ab305c15c932d3d7bc32e7d4a0499439adfdfb107355f329e46d1ffea9fdb0`).
+The prior 18-case comparison against the 7ef7ce6 delay engine also found identical
+complete snapshot bytes, and bidirectional restored continuation matched continuous
+rendering with an ignored-restore negative control. The delay engine, readhead and
+LFO sources are unchanged by this main integration.
+
+The exact same generated WASM is covered by the recorded four cold candidate
+trials in `filter-materialization-results.json` (first callback wall time
+1.659–1.900 ms). Reusing matching source/WASM hashes avoids another timing sample;
+these local measurements are not universal deadline clearance or evidence of a
+hot-loop speedup. Non-first callback outliers, native-control observations and
+clock anomalies remain separate limitations. The full numerical/packed suite also
+checks the shared audition consumer introduced on main.
+
+`artifacts/delay-fx/` contains CANDIDATE WAVs/raw browser PCM, static plots and a
+manifest linking source hashes, exact settings/input, rates, dependency locks,
+package tarball and verification results. No golden is approved or updated.
+
+## Public integration
+
+`@denaudio/den/delay-fx` exports `delayFx`, `DelayFxConfig` and `DelayFxControls`.
+The isolated consumer imports that public subpath. The engine fixture remains
+separate from the Chorus/Rhythmic setting fixtures and the complete integrated
+site. Shared exports/package/lock/contracts and the UI retain one integration owner.

package/docs/delay-readhead.md

@@ -0,0 +1,121 @@
+# Delay readhead — GEN-615
+
+`src/delay-readhead.ts` supplies `delayReadhead` and `DelayReadheadConfig`.
+It is a mono unworklet subgraph, with no feedback, filter, stereo, mix, or tempo
+policy. Instantiate with a required config object and a unique stable name:
+
+```ts
+const cell = instantiate(delayReadhead, {
+  sampleRate: ctx.sampleRate,
+  maxDelaySeconds: 8,
+}, { name: 'delayLeft' });
+// Once per sample, inside forSample:
+const { output, outOfRange } = cell.tick(input, delaySeconds, reset);
+```
+
+For same-sample feedback or other processing of the delayed output, use the
+read-before-write boundary instead of `tick`:
+
+```ts
+const tap = cell.read(delaySeconds, reset);
+tap.write(input.add(tap.output.mul(0.5)));
+output.write(tap.output);
+```
+
+Each sample must use exactly one `read` followed by exactly one call to **that
+read's** `write`, or exactly one `tick`; do not combine the two forms on one instance.
+`read` captures the current delayed output and reset-adjusted pointers. Its `write`
+commits the current input and advances history once. This allows
+`writeInput[n] = input[n] + gain * delayedOutput[n]`, with no previous-output bridge
+and no extra circulation latency. The returned object and closure exist only while
+constructing the unworklet graph, not during audio processing. `tick` delegates to
+the same read/write pair. Reset masks old history before the read, then the write
+starts new history at the reset sample. The minimal feedback test uses a closed-form
+one-sample-delay impulse `[0, 1, 0.5, 0.25, ...]`, including reset and isolation at
+all three offline rates; it does not implement the dependent feedback engine.
+
+`input` / `delaySeconds` are `Node<'f32'>`; `reset` and `outOfRange` are
+`Node<'bool'>`. Input audio must be finite. The sample rate must be the enclosing
+processor's compile-time rate. Use existing AudioParams/events and snapshot APIs;
+this module introduces no control, routing, buffer, or persistence framework.
+
+## Capacity and timeline
+
+Construction allocates `ceil(sampleRate * maxDelaySeconds) + 2` f32 samples.
+The default is two seconds, configurable at construction, with no resize during
+processing. Eight seconds is covered by tests at all three offline rates. Capacity
+must include at least one sample and fit signed i32 indexing; allocation is also
+subject to the host's memory budget. These are representation constraints, not a
+musical/product duration limit. Each instance has independent history and pointers.
+
+At frame `n`, delay `d` in samples produces the linear interpolation of the input
+at timeline position `n-d`. Reads precede the write of input `n`; one sample is the
+minimum delay, and zero is not a dry feedthrough. For `d=k+f`, the weights are
+`(1-f) * x[n-k] + f * x[n-k-1]`. Both adjacent integer indices wrap independently.
+No fractional ring-coordinate subtraction is performed in f32: seconds are widened
+to f64 before rate conversion, and the fractional weight is finally cast to f32.
+This avoids losing small fractional offsets near the end of a long ring.
+
+Finite times saturate to `[1/sampleRate, maxDelaySeconds]`. `outOfRange` reports
+requests outside the **f32-representable seconds endpoints**, allowing an AudioParam
+set to either endpoint without a spurious warning caused by f32 conversion. NaN
+and negative infinity use the minimum; positive infinity uses the maximum, all
+with `outOfRange=true`. Saturation ensures safe reads but does **not** establish
+correct rhythmic playback. A later tempo consumer must use the flag to reject or
+report unavailable times; it must not silently claim that a clamped rhythm is exact.
+
+## Acoustic choice
+
+| Candidate | Time change behavior | Decision |
+| --- | --- | --- |
+| One continuously moving head, linear interpolation | Read velocity is `1 - delta(delaySamples)`; changing time bends pitch. Abrupt jumps can click or repeat/skip source material. | Selected: one explicit trajectory for modulation and steps, no hidden smoothing or extra transition state. |
+| Two stationary heads crossfaded after a change | Avoids the sustained pitch glide but overlaps two different source times; correlated content can comb or cancel. Requires a fade duration and a policy for changes during a fade. | Deferred; does not express the selected moving-head modulation behavior. |
+| Higher-order interpolation on a moving head | Retains Doppler behavior with a different high-frequency response; more taps and possible overshoot. | Not needed for this initial bounded two-tap readhead; a future quality change requires new evidence. |
+
+Linear interpolation is not spectrally transparent: for fractional weight `f`,
+`H(w)=(1-f)+f*exp(-jw)` apart from integer delay. At half-sample offset its magnitude
+is `abs(cos(w/2))`, reaching zero at Nyquist. Nonnegative weights sum to one, so it
+cannot overshoot the extrema of its two finite inputs. Arbitrary fast modulation
+can alias; no anti-aliasing or click-free transition guarantee is made. Abrupt
+controls take effect at their sample with no smoothing. A delay slope of +1 holds
+the source position, and -1 doubles its traversal rate; analytic tests verify both.
+Human sound evaluation remains a later batched instrument/FX step.
+
+## Reset and state
+
+Reset applies before the read at the dispatched sample. It resets the logical
+write pointer and valid-history count, producing zero immediately. The current
+input is then written as the first sample of the new history. Held reset keeps
+output silent. There is no pre-reset contribution to either fractional tap.
+
+This is a constant-time **logical history clear**, not a physical memory wipe.
+Pinned unworklet 0.4.1 has no buffer-clear DSL operation. Old storage bytes can
+remain in a snapshot until overwritten but are unreadable by this readhead; this
+is not a data-erasure API. Reset may click. `history` (persistent f32 buffer),
+`cursor` (i32), and `valid` (i32) use unworklet state with stable instance prefixes.
+Snapshot continuation is supported only with the same schema, capacity, and rate.
+The tests restore history, compare continuation, and prove reset cannot resurrect it.
+
+## Evidence and integration
+
+`tests/delay-readhead.spec.ts` uses an unbounded input-timeline reference, independent
+of ring storage, and analytic impulse/pitch expectations. Tests cover minimum and
+maximum, fractional wrap, steps, modulation, invalid controls, state isolation,
+reset across block/wrap boundaries, and same-schema continuation. Sample comparisons
+use absolute tolerance 2e-6 (4e-6 for analytic pitch tests); scrubbing must remain zero.
+The fixture rejects an intentionally wrong waveform. No audio golden is used.
+
+`tests/delay-packed.test.mjs` packs the package, installs it into an isolated locked
+consumer, checks declarations strictly, renders offline at 44100/48000/96000, and
+builds and renders Chromium at **48000 only**. It imports the public
+`@denaudio/den/delay-readhead` subpath. It neither
+imports repository source nor edits an installed dependency. The existing entry
+test retains the other-rate browser rejection probes.
+
+Artifacts in `artifacts/delay/` are CANDIDATE: raw WAVs, static waveforms, the packed
+tarball/consumer lock, browser samples, and a manifest linking source/settings/input
+and file hashes. CI uploads these through the existing artifact step.
+
+The integration owner exposes `./delay-readhead` with its declarations and ESM
+implementation. The packed fixture exercises that public import; package/lock,
+shared consumers and shared contracts stay under one integration owner.

package/docs/delay-settings-candidate.json

@@ -0,0 +1,333 @@
+{
+  "status": "CANDIDATE",
+  "approval": null,
+  "previousApprovedAudio": null,
+  "description": "Initial reference candidates, not approved golden. No normalization or limiter.",
+  "sourceCommit": "ace4deafaf41f8c4336df7fdb532bbac10a6be42",
+  "sourceDirty": false,
+  "sourceFiles": {
+    "src/delay-settings.ts": "fb7a8ec25889514d9fd75a883c55afe58d2ffad412eb2c35918dd9ad8aa1d626",
+    "src/delay-fx.ts": "1df28860ffee67b594088ece185a81827fec90abb2acd81199ea90df047c420a",
+    "src/delay-readhead.ts": "b859b816bdccb9e638a79f07c8084761fb4f951a943ca2d48fb962162477ce38",
+    "src/filter.ts": "61d49d6b1e52b17791237e00dec5c5089e594ee0030ecaa37f2c632ecb441278",
+    "src/lfo.ts": "a2d73e0c389b91b62e3e833635dc7406d354dcce2d76e83d12e96abce7ca3daf",
+    "tests/fixtures/delay-settings.mjs": "ebf999b689f5f7608547a9d41b2cd72924712946216833992573786393a66318",
+    "scripts/render-delay-settings.mjs": "5da957ade4a82f7863d27d237250cbfc6c4b5cfb4a6260b9fa1ae00c81001815",
+    "scripts/plot-delay-settings.py": "e027a9958cadf9ea1321de1d64ccf689e3db2bba0e44728673f5e250ea21fe19",
+    "tests/delay-settings.spec.ts": "0e5eb94a49c484a871810d65f334f4da65e7db852213f132fac1495eadc18eb9",
+    "tests/delay-settings-packed.test.mjs": "947fdbfa4ef5d1e8cf82365bf4e4ece7640e75868dbd9528809ca83af96bd52a",
+    "package-lock.json": "9aebb16d62097a630ef0022e375f0570dfcc7c082746e74e4b9427637c6fa53a"
+  },
+  "dependencies": {
+    "den": "0.0.0",
+    "unworklet": "0.4.1",
+    "lockSHA256": "9aebb16d62097a630ef0022e375f0570dfcc7c082746e74e4b9427637c6fa53a"
+  },
+  "sampleRate": 48000,
+  "channels": 2,
+  "frames": 262144,
+  "durationSeconds": 5.461333333333333,
+  "initialState": "Fresh engine: zero history/filter states and initial LFO phase; no restored snapshot",
+  "seed": null,
+  "midi": null,
+  "audioEncoding": "IEEE float32 WAV",
+  "browserScope": "48kHz engine packed gate is separate; these candidate files are offline renders",
+  "cases": [
+    {
+      "id": "chorus-base",
+      "setting": {
+        "config": {
+          "maxDelaySeconds": 2,
+          "tone": "flat",
+          "stereoPhaseCycles": 0.5
+        },
+        "parameters": {
+          "timeLeftSeconds": 0.018,
+          "timeRightSeconds": 0.018,
+          "sync": false,
+          "bpm": 120,
+          "beatsLeft": 1,
+          "beatsRight": 1,
+          "feedback": 0,
+          "cutoffHz": 3200,
+          "mix": 0.45,
+          "rateHz": 0.65,
+          "depthSeconds": 0.003,
+          "bypass": false,
+          "reset": false
+        }
+      },
+      "events": [],
+      "source": "Identical mono L/R; 220/330/440Hz sine sum gains 0.12/0.06/0.03; 20ms linear attack, 50ms release; source stops at 2s.",
+      "input": {
+        "file": "chorus-base-input.wav",
+        "sha256": "7323fd06940d219a2c0e96af964394d19fa99d879d3b5a97390943e87d134473"
+      },
+      "raw": {
+        "file": "chorus-base-raw.wav",
+        "sha256": "3eeeffd71e1a7d5e23729f1b29f630ab9befd15994a58b36e3c529f1d7dc1461"
+      },
+      "peaks": [
+        0.19121864438056946,
+        0.19117864966392517
+      ],
+      "verification": {
+        "independentTimelineBiquadMaxError": [
+          2.0489096641540527e-07,
+          1.9371509552001953e-07
+        ],
+        "tolerance": 5e-06,
+        "scrubbedSamples": 0
+      },
+      "listeningFocus": "Slow widening and pitch motion; mono fold coloration; rate/depth edit can click.",
+      "waveform": {
+        "file": "chorus-base-waveform.png",
+        "sha256": "44840f45455ee91e93a918495e23f6a99070dd85de6960a72e618cf92c3eb450",
+        "bucketSamples": 128,
+        "amplitudeRange": [
+          -0.5,
+          0.5
+        ]
+      }
+    },
+    {
+      "id": "chorus-edits",
+      "setting": {
+        "config": {
+          "maxDelaySeconds": 2,
+          "tone": "flat",
+          "stereoPhaseCycles": 0.5
+        },
+        "parameters": {
+          "timeLeftSeconds": 0.018,
+          "timeRightSeconds": 0.018,
+          "sync": false,
+          "bpm": 120,
+          "beatsLeft": 1,
+          "beatsRight": 1,
+          "feedback": 0,
+          "cutoffHz": 3200,
+          "mix": 0.45,
+          "rateHz": 0.65,
+          "depthSeconds": 0.003,
+          "bypass": false,
+          "reset": false
+        }
+      },
+      "events": [
+        {
+          "sample": 72000,
+          "values": {
+            "rateHz": 1.3,
+            "depthSeconds": 0.006
+          }
+        },
+        {
+          "sample": 144000,
+          "values": {
+            "bypass": true
+          }
+        },
+        {
+          "sample": 168000,
+          "values": {
+            "bypass": false
+          }
+        },
+        {
+          "sample": 216000,
+          "values": {
+            "reset": true
+          }
+        },
+        {
+          "sample": 216001,
+          "values": {
+            "reset": false
+          }
+        }
+      ],
+      "source": "Identical mono L/R; 220/330/440Hz sine sum gains 0.12/0.06/0.03; 20ms linear attack, 50ms release; source stops at 2s.",
+      "input": {
+        "file": "chorus-edits-input.wav",
+        "sha256": "7323fd06940d219a2c0e96af964394d19fa99d879d3b5a97390943e87d134473"
+      },
+      "raw": {
+        "file": "chorus-edits-raw.wav",
+        "sha256": "eaafa1a74a47b189a3bf6c26cf287c9c6c12e4813801f3fc6bc29c82e566d8a3"
+      },
+      "peaks": [
+        0.19121313095092773,
+        0.19117864966392517
+      ],
+      "verification": {
+        "independentTimelineBiquadMaxError": [
+          3.0919909477233887e-07,
+          1.9371509552001953e-07
+        ],
+        "tolerance": 5e-06,
+        "scrubbedSamples": 0
+      },
+      "listeningFocus": "Slow widening and pitch motion; mono fold coloration; rate/depth edit can click.",
+      "waveform": {
+        "file": "chorus-edits-waveform.png",
+        "sha256": "9c2ab4f7b0ffc1d195b24c78ff02d992bac62e65935301317bd15d5224dcea24",
+        "bucketSamples": 128,
+        "amplitudeRange": [
+          -0.5,
+          0.5
+        ]
+      }
+    },
+    {
+      "id": "rhythmic-delay-base",
+      "setting": {
+        "config": {
+          "maxDelaySeconds": 2,
+          "tone": "lowpass",
+          "stereoPhaseCycles": 0.5
+        },
+        "parameters": {
+          "timeLeftSeconds": 0.375,
+          "timeRightSeconds": 0.5,
+          "sync": true,
+          "bpm": 120,
+          "beatsLeft": 0.75,
+          "beatsRight": 1,
+          "feedback": 0.48,
+          "cutoffHz": 3200,
+          "mix": 0.35,
+          "rateHz": 0,
+          "depthSeconds": 0,
+          "bypass": false,
+          "reset": false
+        }
+      },
+      "events": [],
+      "source": "Identical mono L/R; four 0.5s plucks at t=0/0.5/1/1.5; 0.3*exp(-local/0.065)*(sin(2pi*110*local)+0.25*sin(2pi*330*local)); stops at 2s.",
+      "input": {
+        "file": "rhythmic-delay-base-input.wav",
+        "sha256": "b58394aba9656fd2949937c71881d6e9a45add8c41cb8247a726389aa3e2982c"
+      },
+      "raw": {
+        "file": "rhythmic-delay-base-raw.wav",
+        "sha256": "3421ba3682892b307e7f426cdbf512ae9f47fd578ff954f6117465a2dc0c9fc9"
+      },
+      "peaks": [
+        0.18629704415798187,
+        0.32658955454826355
+      ],
+      "verification": {
+        "independentTimelineBiquadMaxError": [
+          1.4901161193847656e-08,
+          2.9802322387695312e-08
+        ],
+        "tolerance": 5e-06,
+        "scrubbedSamples": 0
+      },
+      "listeningFocus": "Dotted-eighth/quarter spacing and darkening repeats; abrupt tempo edit can jump/click; bypass hides but preserves tail.",
+      "waveform": {
+        "file": "rhythmic-delay-base-waveform.png",
+        "sha256": "5e420ffe22efd3dfc57955286ae9d66bf838672f3b7a2f6a33d123c30fe2c467",
+        "bucketSamples": 128,
+        "amplitudeRange": [
+          -0.5,
+          0.5
+        ]
+      }
+    },
+    {
+      "id": "rhythmic-delay-edits",
+      "setting": {
+        "config": {
+          "maxDelaySeconds": 2,
+          "tone": "lowpass",
+          "stereoPhaseCycles": 0.5
+        },
+        "parameters": {
+          "timeLeftSeconds": 0.375,
+          "timeRightSeconds": 0.5,
+          "sync": true,
+          "bpm": 120,
+          "beatsLeft": 0.75,
+          "beatsRight": 1,
+          "feedback": 0.48,
+          "cutoffHz": 3200,
+          "mix": 0.35,
+          "rateHz": 0,
+          "depthSeconds": 0,
+          "bypass": false,
+          "reset": false
+        }
+      },
+      "events": [
+        {
+          "sample": 72000,
+          "values": {
+            "bpm": 180,
+            "feedback": 0.65
+          }
+        },
+        {
+          "sample": 144000,
+          "values": {
+            "bypass": true
+          }
+        },
+        {
+          "sample": 168000,
+          "values": {
+            "bypass": false
+          }
+        },
+        {
+          "sample": 216000,
+          "values": {
+            "reset": true
+          }
+        },
+        {
+          "sample": 216001,
+          "values": {
+            "reset": false
+          }
+        }
+      ],
+      "source": "Identical mono L/R; four 0.5s plucks at t=0/0.5/1/1.5; 0.3*exp(-local/0.065)*(sin(2pi*110*local)+0.25*sin(2pi*330*local)); stops at 2s.",
+      "input": {
+        "file": "rhythmic-delay-edits-input.wav",
+        "sha256": "b58394aba9656fd2949937c71881d6e9a45add8c41cb8247a726389aa3e2982c"
+      },
+      "raw": {
+        "file": "rhythmic-delay-edits-raw.wav",
+        "sha256": "a8e395239a030f920735f61c770b63e6cd40eaba2541bd604235763c74e7ae78"
+      },
+      "peaks": [
+        0.1877381056547165,
+        0.30593806505203247
+      ],
+      "verification": {
+        "independentTimelineBiquadMaxError": [
+          2.9802322387695312e-08,
+          2.9802322387695312e-08
+        ],
+        "tolerance": 5e-06,
+        "scrubbedSamples": 0
+      },
+      "listeningFocus": "Dotted-eighth/quarter spacing and darkening repeats; abrupt tempo edit can jump/click; bypass hides but preserves tail.",
+      "waveform": {
+        "file": "rhythmic-delay-edits-waveform.png",
+        "sha256": "72bbeef4e0453e180968e6a520a197fd0d26454768e50d739f5ff7c4da311a33",
+        "bucketSamples": 128,
+        "amplitudeRange": [
+          -0.5,
+          0.5
+        ]
+      }
+    }
+  ],
+  "plotter": {
+    "matplotlib": "3.10.8",
+    "numpy": "2.3.5"
+  }
+}