@openmixer/omx-contract 2.7.0 → 2.8.0

package/README.md

@@ -49,7 +49,7 @@ pkg-config --cflags omx-contract
 ```
 
 Pin one version and build against its renders instead of keeping a copy: the installed package at exactly
-that version (`pkg-config --exact-version=2.7.0 omx-contract`), else the npm tarball, unpacked, and
+that version (`pkg-config --exact-version=2.8.0 omx-contract`), else the npm tarball, unpacked, and
 `omx-contract render --target c --out build/omx-contract/include`. A kernel's controls and readouts, in their
 order, are in `share/omx-contract/omx-contract.json`; how a declaration is written is in
 [docs/kernel-files.md](docs/kernel-files.md), and what a readout is in [docs/readouts.md](docs/readouts.md).

package/data/kernels/comp.json

@@ -127,7 +127,7 @@
       "name": "detectorTopology",
       "kind": "choice",
       "global": "DETECTOR_TOPOLOGY",
-      "doc": "Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, openmixer's ssl-bus-compressor spec §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.",
+      "doc": "Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, the SSL bus compressor design §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.",
       "ids": [
         "feedforward",
         "feedback"
@@ -158,7 +158,7 @@
   "constants": {
     "PROGRAM_RELEASE": {
       "kind": "sheet",
-      "doc": "The program-dependent release's model constants — the LA-2A's opto cell as DATA (`docs/design/specs/2026-09-26-compressor-models.md` §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.",
+      "doc": "The program-dependent release's model constants — the LA-2A's opto cell as DATA (compressor models design §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.",
       "value": {
         "chargeMs": 1000,
         "dischargeMs": 5000,

package/data/kernels/deesser.json

@@ -128,7 +128,7 @@
   "constants": {
     "DEESS_KNEE_DB": {
       "kind": "scalar",
-      "doc": "The knee, in dB — a CONSTANT of the stage and not a row field (`2026-09-22-native-deesser-stage.md` §4b): a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.",
+      "doc": "The knee, in dB — a CONSTANT of the stage and not a row field: a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.",
       "unit": "dB",
       "value": 6
     }

package/data/kernels/delay.json

@@ -5,7 +5,7 @@
       "name": "fxDelayTime",
       "kind": "travel",
       "global": "FX_DELAY_TIME_RANGE",
-      "doc": "The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by `output-delay-ceiling.test.ts`. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word \"delay\" is why they keep being confused for each other.",
+      "doc": "The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by the engine's conformance test. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word \"delay\" is why they keep being confused for each other.",
       "travel": {
         "min": 0,
         "max": 2000,
@@ -38,7 +38,7 @@
     {
       "name": "tone",
       "kind": "travel",
-      "doc": "The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching `DEFAULT_DELAY_STATE.tone`. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.",
+      "doc": "The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching the console's delay default tone. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.",
       "travel": {
         "min": 0,
         "max": 1,
@@ -84,7 +84,7 @@
   "constants": {
     "FX_DELAY_PINGPONG_DEFAULT": {
       "kind": "scalar",
-      "doc": "The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the server's `DEFAULT_DELAY_STATE` (which now reads it) so the DPF plugin generator (`packages/omx-plugins/tools/params-gen.mjs`, `2026-09-25-omx-plugins-dpf.md` §3a) reads the same fact the desk seeds a fresh strip with.",
+      "doc": "The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the console's delay default state (which now reads it) so the DPF plugin generator (omx-plugins, DPF plugin design §3a) reads the same fact the desk seeds a fresh strip with.",
       "unit": "",
       "value": false
     }

package/data/kernels/drive.json

@@ -161,6 +161,22 @@
         "12 kHz",
         "16 kHz"
       ]
+    },
+    {
+      "name": "oversample",
+      "kind": "choice",
+      "doc": "The drive's OVERSAMPLING factor: the shaper runs inside the console's one oversampler at 1x, 2x or 4x, and the id is the factor (`omx_drive.h`'s `omx_drive_factor_of` member set, `struct omx_drive.os_factor`). A field of the console's drive row (`DriveState.oversample`, the console's DRIVE_OVERSAMPLE_FACTORS; the set is named for its control), 4 at come-up there and in omx-dsp's drive face (OMX_DRIVE_INSTANCE_FACTOR, the oversampler's largest).",
+      "ids": [
+        1,
+        2,
+        4
+      ],
+      "default": 4,
+      "labels": [
+        "1x",
+        "2x",
+        "4x"
+      ]
     }
   ],
   "constants": {

package/data/kernels/eq.json

@@ -107,7 +107,7 @@
     {
       "name": "gain",
       "kind": "travel",
-      "doc": "The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (`core/src/eq.ts` derives `EQ_GAIN_LIMIT_DB` from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.",
+      "doc": "The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (the engine derives its gain limit from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.",
       "travel": {
         "min": -15,
         "max": 15,
@@ -162,7 +162,7 @@
   "constants": {
     "EQ_MAX_BANDS": {
       "kind": "scalar",
-      "doc": "The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (`output-delay-ceiling.test.ts` reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.",
+      "doc": "The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (the engine's conformance test reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.",
       "unit": "",
       "value": 24
     },

package/data/kernels/fbs.json

@@ -130,7 +130,7 @@
     },
     "FBS_MAX_NOTCH_DEPTH_DB": {
       "kind": "scalar",
-      "doc": "THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (2026-09-08-room-analysis.md §4 — \"never more than the FBS notch depth law\"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.",
+      "doc": "THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (the room-analysis design §4 — \"never more than the FBS notch depth law\"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.",
       "unit": "dB",
       "value": 18
     },

package/data/kernels/hrp.json

@@ -24,7 +24,7 @@
     },
     "HRP_MAX_VOICES": {
       "kind": "scalar",
-      "doc": "How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. `core/src/output-delay-ceiling.test.ts` reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.",
+      "doc": "How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. the engine's conformance test reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.",
       "unit": "",
       "value": 8
     },
@@ -60,7 +60,7 @@
     },
     "HRP_DEFAULT_MAX_AUTO_BANDS": {
       "kind": "scalar",
-      "doc": "How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer \"will this fit?\" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. `eq-budget-fits.test.ts` asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.",
+      "doc": "How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer \"will this fit?\" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. the engine's conformance test asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.",
       "unit": "",
       "value": 8
     }

package/data/kernels/limiter.json

@@ -59,7 +59,7 @@
   ],
   "tables": {
     "LIMITER_LIMITS": {
-      "doc": "The precision limiter's travels (`2026-09-27-precision-limiter.md` §2), rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by `harness/contract-limits-gen.mjs`, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth."
+      "doc": "The precision limiter's travels, rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by the engine's limits generator, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth."
     }
   }
 }

package/data/kernels/mixmatrix.json

@@ -3,7 +3,7 @@
   "constants": {
     "BUS_INPUT_CAP": {
       "kind": "scalar",
-      "doc": "The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. `native-input-ceiling.test.ts` holds this to the C header; spec `docs/design/specs/2026-09-14-console-allocation.md` §9 carries the ruling and the measured cost of width.",
+      "doc": "The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. the engine's conformance test holds this to the C header; the console allocation design §9 carries the ruling and the measured cost of width.",
       "unit": "",
       "value": 128
     }

package/data/kernels/reverb.json

@@ -4,7 +4,7 @@
     {
       "name": "plateModDepth",
       "kind": "travel",
-      "doc": "PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it (`2026-09-26-native-fx-catalogue.md` §2, Reverb).",
+      "doc": "PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it.",
       "travel": {
         "min": 0,
         "max": 400,
@@ -111,7 +111,7 @@
     {
       "name": "reverse",
       "kind": "travel",
-      "doc": "REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds (`2026-07-15-native-delay-reverb.md`, Amendment 2026-09-17 §B). A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.",
+      "doc": "REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds. A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.",
       "travel": {
         "min": 50,
         "max": 500,

package/data/kernels/rta.json

@@ -3,7 +3,7 @@
   "constants": {
     "RTA_TARGET_BIN_HZ": {
       "kind": "scalar",
-      "doc": "The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; `rta-window-derivation.test.ts` reads it back from the header.",
+      "doc": "The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; the engine's conformance test reads it back from the header.",
       "unit": "Hz",
       "value": 5.9
     },

package/data/kernels/transient.json

@@ -69,7 +69,7 @@
   ],
   "tables": {
     "TRANSIENT_LIMITS": {
-      "doc": "The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. `harness/contract-limits-gen.mjs` emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921)."
+      "doc": "The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. the engine's limits generator emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921)."
     }
   },
   "constants": {

package/data/kernels/trim.json

@@ -5,7 +5,7 @@
       "name": "trimDb",
       "kind": "travel",
       "global": "TRIM_RANGE",
-      "doc": "The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (openmixer's TRIM_RANGE, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.",
+      "doc": "The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (the console's trim range, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.",
       "travel": {
         "min": -24,
         "max": 24,

package/data/rates.json

@@ -2,7 +2,7 @@
   "$schema": "../schema/omx-contract.schema.json",
   "STANDARD_SAMPLE_RATES": {
     "kind": "list",
-    "doc": "The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` (`2026-07-15-clock-rate-honest-force-rate.md`) — this is the ceiling a write is refused by, never a menu.",
+    "doc": "The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` — this is the ceiling a write is refused by, never a menu.",
     "unit": "Hz",
     "values": [
       44100,

package/dist/ts/omx-contract.d.ts

@@ -255,13 +255,13 @@ export declare const COMP_KINDS: readonly ["comp", "limiter"];
 export declare const DETECTOR_OVERSAMPLINGS: readonly ["auto", "off", "x4"];
 
 /**
- * Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, openmixer's ssl-bus-compressor spec §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.
+ * Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, the SSL bus compressor design §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.
  * @see data/kernels/comp.json
  */
 export declare const DETECTOR_TOPOLOGY: readonly ["feedforward", "feedback"];
 
 /**
- * The program-dependent release's model constants — the LA-2A's opto cell as DATA (`docs/design/specs/2026-09-26-compressor-models.md` §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.
+ * The program-dependent release's model constants — the LA-2A's opto cell as DATA (compressor models design §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.
  * @see data/kernels/comp.json
  */
 export declare const PROGRAM_RELEASE: {
@@ -381,13 +381,13 @@ export declare const DEESS_RELEASE_RANGE: {
 export declare const DEESS_MODES: readonly ["split", "wideband"];
 
 /**
- * The knee, in dB — a CONSTANT of the stage and not a row field (`2026-09-22-native-deesser-stage.md` §4b): a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.
+ * The knee, in dB — a CONSTANT of the stage and not a row field: a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.
  * @see data/kernels/deesser.json
  */
 export declare const DEESS_KNEE_DB: 6;
 
 /**
- * The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by `output-delay-ceiling.test.ts`. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word "delay" is why they keep being confused for each other.
+ * The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by the engine's conformance test. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word "delay" is why they keep being confused for each other.
  * @see data/kernels/delay.json
  */
 export declare const FX_DELAY_TIME_RANGE: {
@@ -413,7 +413,7 @@ export declare const FX_DELAY_FEEDBACK_RANGE: {
 };
 
 /**
- * The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching `DEFAULT_DELAY_STATE.tone`. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.
+ * The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching the console's delay default tone. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.
  * @see data/kernels/delay.json
  */
 export declare const DELAY_TONE_RANGE: {
@@ -448,7 +448,7 @@ export declare const DELAY_MIX_RANGE: {
 export declare const DELAY_PINGPONGS: readonly ["off", "on"];
 
 /**
- * The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the server's `DEFAULT_DELAY_STATE` (which now reads it) so the DPF plugin generator (`packages/omx-plugins/tools/params-gen.mjs`, `2026-09-25-omx-plugins-dpf.md` §3a) reads the same fact the desk seeds a fresh strip with.
+ * The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the console's delay default state (which now reads it) so the DPF plugin generator (omx-plugins, DPF plugin design §3a) reads the same fact the desk seeds a fresh strip with.
  * @see data/kernels/delay.json
  */
 export declare const FX_DELAY_PINGPONG_DEFAULT: false;
@@ -548,6 +548,12 @@ export declare const DRIVE_STEREO_LINKS: readonly ["off", "on"];
  */
 export declare const DRIVE_HF_ROLLOFFS: readonly [0, 12000, 16000];
 
+/**
+ * The drive's OVERSAMPLING factor: the shaper runs inside the console's one oversampler at 1x, 2x or 4x, and the id is the factor (`omx_drive.h`'s `omx_drive_factor_of` member set, `struct omx_drive.os_factor`). A field of the console's drive row (`DriveState.oversample`, the console's DRIVE_OVERSAMPLE_FACTORS; the set is named for its control), 4 at come-up there and in omx-dsp's drive face (OMX_DRIVE_INSTANCE_FACTOR, the oversampler's largest).
+ * @see data/kernels/drive.json
+ */
+export declare const DRIVE_OVERSAMPLES: readonly [1, 2, 4];
+
 /**
  * The bias at `character = +1`, in the SHAPER's own input units — `OMX_DRIVE_BIAS_MAX`. DERIVED, not chosen (§4b). In the saturating limit every knee tends to `sgn`, the output is a rectangular wave of duty `d = ½ + asin(b)/π`, and its `n`th harmonic goes as `sin(nπd)/n` — so `|H2|` is maximal at `d = ¾`, which is `b = sin(π/4) = 1/√2` exactly. Measured too: the H2-maximising bias for a full-scale sine at unity drive is 0.712 (soft), 0.658 (tape), 0.645 (tube), and this one value costs 0.000 / 0.035 / 0.088 dB of H2 against each curve's own optimum. ONE normalised bias for all three knees, not three.
  * @see data/kernels/drive.json
@@ -612,7 +618,7 @@ export declare const EQ_FREQ_RANGE: {
 };
 
 /**
- * The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (`core/src/eq.ts` derives `EQ_GAIN_LIMIT_DB` from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.
+ * The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (the engine derives its gain limit from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.
  * @see data/kernels/eq.json
  */
 export declare const EQ_GAIN_RANGE: {
@@ -651,7 +657,7 @@ export declare const EQ_NOTCH_Q_RANGE: {
 export declare const EQ_BAND_ONS: readonly ["off", "on"];
 
 /**
- * The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (`output-delay-ceiling.test.ts` reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.
+ * The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (the engine's conformance test reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.
  * @see data/kernels/eq.json
  */
 export declare const EQ_MAX_BANDS: 24;
@@ -846,7 +852,7 @@ export declare const DEFAULT_THRESHOLDS: {
 };
 
 /**
- * THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (2026-09-08-room-analysis.md §4 — "never more than the FBS notch depth law"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.
+ * THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (the room-analysis design §4 — "never more than the FBS notch depth law"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.
  * @see data/kernels/fbs.json
  */
 export declare const FBS_MAX_NOTCH_DEPTH_DB: 18;
@@ -1094,7 +1100,7 @@ export declare const HRP_AMOUNT_RANGE: {
 export declare const HRP_MAX_CUT_DB: 6;
 
 /**
- * How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. `core/src/output-delay-ceiling.test.ts` reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.
+ * How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. the engine's conformance test reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.
  * @see data/kernels/hrp.json
  */
 export declare const HRP_MAX_VOICES: 8;
@@ -1130,13 +1136,13 @@ export declare const HRP_COLLISION_CENTS: 100;
 export declare const HRP_ATTRIBUTION_CENTS: 60;
 
 /**
- * How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer "will this fit?" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. `eq-budget-fits.test.ts` asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.
+ * How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer "will this fit?" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. the engine's conformance test asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.
  * @see data/kernels/hrp.json
  */
 export declare const HRP_DEFAULT_MAX_AUTO_BANDS: 8;
 
 /**
- * The precision limiter's travels (`2026-09-27-precision-limiter.md` §2), rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by `harness/contract-limits-gen.mjs`, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth.
+ * The precision limiter's travels, rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by the engine's limits generator, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth.
  * @see data/kernels/limiter.json
  */
 export declare const LIMITER_LIMITS: {
@@ -1167,7 +1173,7 @@ export declare const LIMITER_LIMITS: {
 };
 
 /**
- * The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. `native-input-ceiling.test.ts` holds this to the C header; spec `docs/design/specs/2026-09-14-console-allocation.md` §9 carries the ruling and the measured cost of width.
+ * The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. the engine's conformance test holds this to the C header; the console allocation design §9 carries the ruling and the measured cost of width.
  * @see data/kernels/mixmatrix.json
  */
 export declare const BUS_INPUT_CAP: 128;
@@ -1342,7 +1348,7 @@ export declare const PITCH_KERNEL: {
 };
 
 /**
- * PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it (`2026-09-26-native-fx-catalogue.md` §2, Reverb).
+ * PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it.
  * @see data/kernels/reverb.json
  */
 export declare const REVERB_PLATE_MOD_DEPTH_RANGE: {
@@ -1449,7 +1455,7 @@ export declare const REVERB_HIGHCUT_RANGE: {
 };
 
 /**
- * REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds (`2026-07-15-native-delay-reverb.md`, Amendment 2026-09-17 §B). A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.
+ * REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds. A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.
  * @see data/kernels/reverb.json
  */
 export declare const REVERB_REVERSE_RANGE: {
@@ -1745,7 +1751,7 @@ export declare const ROTARY_STOP_EPS: 0.0001;
 export declare const ROTARY_RING_FLOATS: 256;
 
 /**
- * The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; `rta-window-derivation.test.ts` reads it back from the header.
+ * The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; the engine's conformance test reads it back from the header.
  * @see data/kernels/rta.json
  */
 export declare const RTA_TARGET_BIN_HZ: 5.9;
@@ -1781,7 +1787,7 @@ export declare const SRC_ROLLOFF: 0.92;
 export declare const SRC_MAX_L: 1024;
 
 /**
- * The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. `harness/contract-limits-gen.mjs` emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921).
+ * The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. the engine's limits generator emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921).
  * @see data/kernels/transient.json
  */
 export declare const TRANSIENT_LIMITS: {
@@ -1928,7 +1934,7 @@ export declare const TREMOLO_TRAVELS: {
 };
 
 /**
- * The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (openmixer's TRIM_RANGE, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.
+ * The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (the console's trim range, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.
  * @see data/kernels/trim.json
  */
 export declare const TRIM_RANGE: {
@@ -2094,7 +2100,7 @@ export declare const EQ_CONTINUOUS_TRAVELS: {
 export declare const READOUT_RATE_HZ: 50;
 
 /**
- * The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` (`2026-07-15-clock-rate-honest-force-rate.md`) — this is the ceiling a write is refused by, never a menu.
+ * The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` — this is the ceiling a write is refused by, never a menu.
  * @see data/rates.json
  */
 export declare const STANDARD_SAMPLE_RATES: readonly [44100, 48000, 88200, 96000, 176400, 192000];

package/dist/ts/omx-contract.js

@@ -265,7 +265,7 @@ export const DETECTOR_OVERSAMPLINGS = [
 ];
 
 /**
- * Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, openmixer's ssl-bus-compressor spec §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.
+ * Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, the SSL bus compressor design §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.
  * @see data/kernels/comp.json
  */
 export const DETECTOR_TOPOLOGY = [
@@ -274,7 +274,7 @@ export const DETECTOR_TOPOLOGY = [
 ];
 
 /**
- * The program-dependent release's model constants — the LA-2A's opto cell as DATA (`docs/design/specs/2026-09-26-compressor-models.md` §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.
+ * The program-dependent release's model constants — the LA-2A's opto cell as DATA (compressor models design §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.
  * @see data/kernels/comp.json
  */
 export const PROGRAM_RELEASE = {
@@ -397,13 +397,13 @@ export const DEESS_MODES = [
 ];
 
 /**
- * The knee, in dB — a CONSTANT of the stage and not a row field (`2026-09-22-native-deesser-stage.md` §4b): a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.
+ * The knee, in dB — a CONSTANT of the stage and not a row field: a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.
  * @see data/kernels/deesser.json
  */
 export const DEESS_KNEE_DB = 6;
 
 /**
- * The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by `output-delay-ceiling.test.ts`. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word "delay" is why they keep being confused for each other.
+ * The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by the engine's conformance test. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word "delay" is why they keep being confused for each other.
  * @see data/kernels/delay.json
  */
 export const FX_DELAY_TIME_RANGE = {
@@ -429,7 +429,7 @@ export const FX_DELAY_FEEDBACK_RANGE = {
 };
 
 /**
- * The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching `DEFAULT_DELAY_STATE.tone`. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.
+ * The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching the console's delay default tone. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.
  * @see data/kernels/delay.json
  */
 export const DELAY_TONE_RANGE = {
@@ -467,7 +467,7 @@ export const DELAY_PINGPONGS = [
 ];
 
 /**
- * The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the server's `DEFAULT_DELAY_STATE` (which now reads it) so the DPF plugin generator (`packages/omx-plugins/tools/params-gen.mjs`, `2026-09-25-omx-plugins-dpf.md` §3a) reads the same fact the desk seeds a fresh strip with.
+ * The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the console's delay default state (which now reads it) so the DPF plugin generator (omx-plugins, DPF plugin design §3a) reads the same fact the desk seeds a fresh strip with.
  * @see data/kernels/delay.json
  */
 export const FX_DELAY_PINGPONG_DEFAULT = false;
@@ -587,6 +587,16 @@ export const DRIVE_HF_ROLLOFFS = [
   16000
 ];
 
+/**
+ * The drive's OVERSAMPLING factor: the shaper runs inside the console's one oversampler at 1x, 2x or 4x, and the id is the factor (`omx_drive.h`'s `omx_drive_factor_of` member set, `struct omx_drive.os_factor`). A field of the console's drive row (`DriveState.oversample`, the console's DRIVE_OVERSAMPLE_FACTORS; the set is named for its control), 4 at come-up there and in omx-dsp's drive face (OMX_DRIVE_INSTANCE_FACTOR, the oversampler's largest).
+ * @see data/kernels/drive.json
+ */
+export const DRIVE_OVERSAMPLES = [
+  1,
+  2,
+  4
+];
+
 /**
  * The bias at `character = +1`, in the SHAPER's own input units — `OMX_DRIVE_BIAS_MAX`. DERIVED, not chosen (§4b). In the saturating limit every knee tends to `sgn`, the output is a rectangular wave of duty `d = ½ + asin(b)/π`, and its `n`th harmonic goes as `sin(nπd)/n` — so `|H2|` is maximal at `d = ¾`, which is `b = sin(π/4) = 1/√2` exactly. Measured too: the H2-maximising bias for a full-scale sine at unity drive is 0.712 (soft), 0.658 (tape), 0.645 (tube), and this one value costs 0.000 / 0.035 / 0.088 dB of H2 against each curve's own optimum. ONE normalised bias for all three knees, not three.
  * @see data/kernels/drive.json
@@ -667,7 +677,7 @@ export const EQ_FREQ_RANGE = {
 };
 
 /**
- * The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (`core/src/eq.ts` derives `EQ_GAIN_LIMIT_DB` from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.
+ * The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (the engine derives its gain limit from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.
  * @see data/kernels/eq.json
  */
 export const EQ_GAIN_RANGE = {
@@ -709,7 +719,7 @@ export const EQ_BAND_ONS = [
 ];
 
 /**
- * The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (`output-delay-ceiling.test.ts` reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.
+ * The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (the engine's conformance test reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.
  * @see data/kernels/eq.json
  */
 export const EQ_MAX_BANDS = 24;
@@ -1003,7 +1013,7 @@ export const DEFAULT_THRESHOLDS = {
 };
 
 /**
- * THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (2026-09-08-room-analysis.md §4 — "never more than the FBS notch depth law"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.
+ * THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (the room-analysis design §4 — "never more than the FBS notch depth law"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.
  * @see data/kernels/fbs.json
  */
 export const FBS_MAX_NOTCH_DEPTH_DB = 18;
@@ -1263,7 +1273,7 @@ export const HRP_AMOUNT_RANGE = {
 export const HRP_MAX_CUT_DB = 6;
 
 /**
- * How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. `core/src/output-delay-ceiling.test.ts` reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.
+ * How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. the engine's conformance test reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.
  * @see data/kernels/hrp.json
  */
 export const HRP_MAX_VOICES = 8;
@@ -1299,13 +1309,13 @@ export const HRP_COLLISION_CENTS = 100;
 export const HRP_ATTRIBUTION_CENTS = 60;
 
 /**
- * How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer "will this fit?" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. `eq-budget-fits.test.ts` asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.
+ * How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer "will this fit?" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. the engine's conformance test asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.
  * @see data/kernels/hrp.json
  */
 export const HRP_DEFAULT_MAX_AUTO_BANDS = 8;
 
 /**
- * The precision limiter's travels (`2026-09-27-precision-limiter.md` §2), rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by `harness/contract-limits-gen.mjs`, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth.
+ * The precision limiter's travels, rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by the engine's limits generator, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth.
  * @see data/kernels/limiter.json
  */
 export const LIMITER_LIMITS = {
@@ -1336,7 +1346,7 @@ export const LIMITER_LIMITS = {
 };
 
 /**
- * The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. `native-input-ceiling.test.ts` holds this to the C header; spec `docs/design/specs/2026-09-14-console-allocation.md` §9 carries the ruling and the measured cost of width.
+ * The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. the engine's conformance test holds this to the C header; the console allocation design §9 carries the ruling and the measured cost of width.
  * @see data/kernels/mixmatrix.json
  */
 export const BUS_INPUT_CAP = 128;
@@ -1518,7 +1528,7 @@ export const PITCH_KERNEL = {
 };
 
 /**
- * PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it (`2026-09-26-native-fx-catalogue.md` §2, Reverb).
+ * PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it.
  * @see data/kernels/reverb.json
  */
 export const REVERB_PLATE_MOD_DEPTH_RANGE = {
@@ -1625,7 +1635,7 @@ export const REVERB_HIGHCUT_RANGE = {
 };
 
 /**
- * REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds (`2026-07-15-native-delay-reverb.md`, Amendment 2026-09-17 §B). A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.
+ * REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds. A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.
  * @see data/kernels/reverb.json
  */
 export const REVERB_REVERSE_RANGE = {
@@ -1931,7 +1941,7 @@ export const ROTARY_STOP_EPS = 0.0001;
 export const ROTARY_RING_FLOATS = 256;
 
 /**
- * The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; `rta-window-derivation.test.ts` reads it back from the header.
+ * The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; the engine's conformance test reads it back from the header.
  * @see data/kernels/rta.json
  */
 export const RTA_TARGET_BIN_HZ = 5.9;
@@ -1967,7 +1977,7 @@ export const SRC_ROLLOFF = 0.92;
 export const SRC_MAX_L = 1024;
 
 /**
- * The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. `harness/contract-limits-gen.mjs` emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921).
+ * The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. the engine's limits generator emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921).
  * @see data/kernels/transient.json
  */
 export const TRANSIENT_LIMITS = {
@@ -2117,7 +2127,7 @@ export const TREMOLO_TRAVELS = {
 };
 
 /**
- * The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (openmixer's TRIM_RANGE, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.
+ * The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (the console's trim range, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.
  * @see data/kernels/trim.json
  */
 export const TRIM_RANGE = {
@@ -2287,7 +2297,7 @@ export const EQ_CONTINUOUS_TRAVELS = {
 export const READOUT_RATE_HZ = 50;
 
 /**
- * The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` (`2026-07-15-clock-rate-honest-force-rate.md`) — this is the ceiling a write is refused by, never a menu.
+ * The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` — this is the ceiling a write is refused by, never a menu.
  * @see data/rates.json
  */
 export const STANDARD_SAMPLE_RATES = [

package/include/omxcontract/omx_contract_limits.h

@@ -696,6 +696,13 @@ enum omx_drive_hf_rolloffs {
 };
 #define OMX_DRIVE_HF_ROLLOFFS_COUNT 3u
 #define OMX_DRIVE_HF_ROLLOFFS_DEFAULT OMX_DRIVE_HF_ROLLOFFS_0
+enum omx_drive_oversamples {
+  OMX_DRIVE_OVERSAMPLES_1 = 1,
+  OMX_DRIVE_OVERSAMPLES_2 = 2,
+  OMX_DRIVE_OVERSAMPLES_4 = 4,
+};
+#define OMX_DRIVE_OVERSAMPLES_COUNT 3u
+#define OMX_DRIVE_OVERSAMPLES_DEFAULT OMX_DRIVE_OVERSAMPLES_4
 enum omx_drive_stereo_links {
   OMX_DRIVE_STEREO_LINKS_OFF = 0,
   OMX_DRIVE_STEREO_LINKS_ON = 1,

package/lib/semver.mjs

@@ -20,7 +20,7 @@
 const RANK = { none: 0, patch: 1, minor: 2, major: 3 };
 
 /**
- * The development-phase rule (openmixer spec 2026-10-05-omx-contract §3.6): while every consumer of
+ * The development-phase rule (omx-contract spec §3.6): while every consumer of
  * omx-contract is a FreeMixer repository named here, removing an item is MINOR. The rule ends, and
  * this becomes null so a removal is MAJOR again, in the release that names the first consumer
  * outside the list.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@openmixer/omx-contract",
-  "version": "2.7.0",
+  "version": "2.8.0",
   "description": "The FreeMixer declarations as data: a JSON Schema, the data, and one CLI that renders them into C, TypeScript and JSON",
   "type": "module",
   "license": "GPL-3.0-or-later",

package/render/c/contract-limits-render.mjs

@@ -1,11 +1,10 @@
 // SPDX-License-Identifier: GPL-3.0-or-later
 // Copyright (C) 2026 Pau Aliagas <linuxnow@gmail.com>
-// Moved unchanged from openmixer harness/contract-limits-render.mjs at 6c6e709f4429 (omx-contract spec
-// §3.4): a frozen export of openmixer's whole sheet renders through it byte for byte (test/round-trip.test.mjs).
+// Moved unchanged from the console engine's contract-limits renderer at 6c6e709f4429 (omx-contract spec
+// §3.4): a frozen export of the engine's whole sheet renders through it byte for byte (test/round-trip.test.mjs).
 /**
  * The RENDER half of `contract-limits-gen.mjs`: the header text as a pure function of the
- * declaration sheet it is handed. Split out so the declaration-perturbation ratchet (R-094,
- * `packages/server/src/declaration-perturbation-conformance.test.ts`) can render the header over
+ * declaration sheet it is handed. Split out so the engine's declaration-perturbation test can render the header over
  * a PERTURBED sheet and see whether every define follows it — the generator itself imports the
  * built modules and calls this, so there is one renderer and no second copy.
  */
@@ -220,7 +219,7 @@ function genDeclaredListsBlock(DECLARED_LISTS, definedElsewhere) {
  * over `<PREFIX>_LIMITS`, `declaredScalars`, `declaredSets`, `declaredLists`), then `extra` — the
  * names that are not in the core barrel (`OMX_CLAP_HOST_EXTENSIONS`) or a test's override of one
  * door. The generator and every render test build the sheet here, so a new door is added once,
- * never appended to a hand list per caller (2026-09-15-no-stored-derivations.md §9).
+ * never appended to a hand list per caller.
  */
 export function contractLimitsSheet(coreNs, declarations, extra = {}) {
   return {

package/render/doors.mjs

@@ -1,8 +1,8 @@
 // SPDX-License-Identifier: GPL-3.0-or-later
 // Copyright (C) 2026 Pau Aliagas <linuxnow@gmail.com>
 /**
- * The resolved data cut into the C renderer's doors, as openmixer's discovery cut its barrels
- * (packages/core/src/native-limit-tables.ts, declared-scalars.ts, declared-lists.ts): a `travels`
+ * The resolved data cut into the C renderer's doors, as the engine's discovery cut its barrels
+ * (its native limit tables, declared scalars and declared lists): a `travels`
  * table named <PREFIX>_LIMITS is a native table; every scalar, sheet and single travel is scalars
  * (its finite-number leaves, a travel's `unit`/`defaultFrom` labels skipped); a list is a declared
  * list unless its item says `"c": { "render": false }`; a set is an enum; a `c.alias` is the core-limit spelling.

package/share/omx-contract/omx-contract.json

@@ -1,6 +1,6 @@
 {
   "name": "@openmixer/omx-contract",
-  "version": "2.7.0",
+  "version": "2.8.0",
   "items": {
     "PAN_RANGE": {
       "rel": "data/kernels/balance.json",
@@ -302,7 +302,7 @@
     "DETECTOR_TOPOLOGY": {
       "rel": "data/kernels/comp.json",
       "kind": "set",
-      "doc": "Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, openmixer's ssl-bus-compressor spec §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.",
+      "doc": "Where the compressor's detector listens: `feedforward` hears the slot's input (or its key), `feedback` hears the slot's own pre-make-up output DYN_FB_LOOP_DELAY_SAMPLES late, through one envelope stage, and a keyed slot stays feed-forward (omx-dsp's OMX_DYN_TOPOLOGY_FEEDFORWARD 0 and _FEEDBACK 1, in that order; the set is named for its control, so C spells them OMX_DETECTOR_TOPOLOGY_*). The console's comp row carries it (`detectorTopology`, the SSL bus compressor design §3a) and comes up `feedforward`, so a zeroed atom is the compressor every slot runs.",
       "default": "feedforward",
       "value": [
         "feedforward",
@@ -312,7 +312,7 @@
     "PROGRAM_RELEASE": {
       "rel": "data/kernels/comp.json",
       "kind": "sheet",
-      "doc": "The program-dependent release's model constants — the LA-2A's opto cell as DATA (`docs/design/specs/2026-09-26-compressor-models.md` §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.",
+      "doc": "The program-dependent release's model constants — the LA-2A's opto cell as DATA (compressor models design §3b). `chargeMs`/`dischargeMs` are the word's memory constants, shared by every profile; each profile's `fastMs`, `slowMinMs`, `slowMaxMs` and `share` fit the unit's published two-stage recovery. Not operator controls: generated into `omx_contract_limits.h` as `OMX_PROGRAM_RELEASE_*`, which libomxdsp's `omx_env_program_release` reads.",
       "value": {
         "chargeMs": 1000,
         "dischargeMs": 5000,
@@ -443,14 +443,14 @@
     "DEESS_KNEE_DB": {
       "rel": "data/kernels/deesser.json",
       "kind": "scalar",
-      "doc": "The knee, in dB — a CONSTANT of the stage and not a row field (`2026-09-22-native-deesser-stage.md` §4b): a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.",
+      "doc": "The knee, in dB — a CONSTANT of the stage and not a row field: a de-esser without a soft knee latches audibly on the first sibilant of a phrase, and no other value of this number is a de-esser anyone would dial. The controller pushes it to the node and this is the same number; a second spelling would be a picture that disagrees with the audio about where the reduction starts.",
       "unit": "dB",
       "value": 6
     },
     "FX_DELAY_TIME_RANGE": {
       "rel": "data/kernels/delay.json",
       "kind": "travels",
-      "doc": "The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by `output-delay-ceiling.test.ts`. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word \"delay\" is why they keep being confused for each other.",
+      "doc": "The FX delay's time travel. `max` is the NATIVE ring's ceiling — the same number as `OMX_FXDELAY_MAX_MS` in `omx_delay.h`, pinned to it by the engine's conformance test. Distinct from the per-route ALIGNMENT delay (`OUTPUT_DELAY_RANGE`, 1 s, `OMX_DELAY_MAX_MS`): one is a musical echo with feedback and mix, the other is a pure latency. Two facts, two rings, two ceilings — sharing the word \"delay\" is why they keep being confused for each other.",
       "c": {
         "alias": {
           "fact": "delay",
@@ -484,7 +484,7 @@
     "DELAY_TONE_RANGE": {
       "rel": "data/kernels/delay.json",
       "kind": "travels",
-      "doc": "The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching `DEFAULT_DELAY_STATE.tone`. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.",
+      "doc": "The FX delay's TONE — a NORMALIZED_RANGE instance with its own come-up value, matching the console's delay default tone. See REVERB_SIZE_RANGE for why this is a named constant rather than a default on the shared shape.",
       "shape": "travel",
       "value": {
         "min": 0,
@@ -529,7 +529,7 @@
     "FX_DELAY_PINGPONG_DEFAULT": {
       "rel": "data/kernels/delay.json",
       "kind": "scalar",
-      "doc": "The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the server's `DEFAULT_DELAY_STATE` (which now reads it) so the DPF plugin generator (`packages/omx-plugins/tools/params-gen.mjs`, `2026-09-25-omx-plugins-dpf.md` §3a) reads the same fact the desk seeds a fresh strip with.",
+      "doc": "The FX delay's PING-PONG come-up value — a toggle, so it has no travel, only this. Declared here rather than as a literal in the console's delay default state (which now reads it) so the DPF plugin generator (omx-plugins, DPF plugin design §3a) reads the same fact the desk seeds a fresh strip with.",
       "unit": "",
       "value": false
     },
@@ -699,6 +699,22 @@
         16000
       ]
     },
+    "DRIVE_OVERSAMPLES": {
+      "rel": "data/kernels/drive.json",
+      "kind": "set",
+      "doc": "The drive's OVERSAMPLING factor: the shaper runs inside the console's one oversampler at 1x, 2x or 4x, and the id is the factor (`omx_drive.h`'s `omx_drive_factor_of` member set, `struct omx_drive.os_factor`). A field of the console's drive row (`DriveState.oversample`, the console's DRIVE_OVERSAMPLE_FACTORS; the set is named for its control), 4 at come-up there and in omx-dsp's drive face (OMX_DRIVE_INSTANCE_FACTOR, the oversampler's largest).",
+      "default": 4,
+      "labels": [
+        "1x",
+        "2x",
+        "4x"
+      ],
+      "value": [
+        1,
+        2,
+        4
+      ]
+    },
     "DRIVE_BIAS_MAX": {
       "rel": "data/kernels/drive.json",
       "kind": "scalar",
@@ -805,7 +821,7 @@
     "EQ_GAIN_RANGE": {
       "rel": "data/kernels/eq.json",
       "kind": "travels",
-      "doc": "The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (`core/src/eq.ts` derives `EQ_GAIN_LIMIT_DB` from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.",
+      "doc": "The parametric-band gain travel, in dB. The ± scalar `clampEqGain` enforces is READ off this (the engine derives its gain limit from `max`), so the bound is stated once and a surface cannot draw a wider knob than the clamp will honour.",
       "shape": "travel",
       "value": {
         "min": -15,
@@ -854,7 +870,7 @@
     "EQ_MAX_BANDS": {
       "rel": "data/kernels/eq.json",
       "kind": "scalar",
-      "doc": "The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (`output-delay-ceiling.test.ts` reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.",
+      "doc": "The maximum number of parametric bands a channel EQ may hold — the operator can add up to this many. Generated into `omx_contract_limits.h` as `OMX_EQ_MAX_BANDS`, which `omx_biquad.h` reads (the engine's conformance test reads the header): the native lane allocates exactly this many biquad slots per channel and silently clamps above it, so the model, the wire, and the DSP agree by test, not by memory.",
       "unit": "",
       "value": 24
     },
@@ -1170,7 +1186,7 @@
     "FBS_MAX_NOTCH_DEPTH_DB": {
       "rel": "data/kernels/fbs.json",
       "kind": "scalar",
-      "doc": "THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (2026-09-08-room-analysis.md §4 — \"never more than the FBS notch depth law\"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.",
+      "doc": "THE NOTCH DEPTH LAW: the deepest cut any machine planter may propose from one reading, in dB. A number, not a literal at the call site, because two planters now lean on it. FBS derives a candidate's depth from the peak-to-average ratio around its bin and stops here; Room Analysis derives a mode's depth from its height above the median curve and is held to the SAME ceiling (the room-analysis design §4 — \"never more than the FBS notch depth law\"). Written twice as `18` those two bounds would be equal by coincidence and could drift apart in silence. It is a CEILING ON THE PROPOSAL, not the final gain: a band that lands as a rampable BELL is bounded again by `EQ_GAIN_RANGE`, which is tighter. A planter must honour both and REPORT when either bit, never quietly deliver a shallower cut than it was asked for.",
       "unit": "dB",
       "value": 18
     },
@@ -1447,7 +1463,7 @@
     "HRP_MAX_VOICES": {
       "rel": "data/kernels/hrp.json",
       "kind": "scalar",
-      "doc": "How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. `core/src/output-delay-ceiling.test.ts` reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.",
+      "doc": "How many simultaneous musical voices the detector may report on one channel. PINNED to `OMX_HRP_MAX_VOICES` in `pipewire-native/src/hrp_pitch.h` — the C voice set is a FIXED-SIZE array, so a TS number above the macro promises voices the analyser structurally cannot return, and one below it silently drops notes the C already found. the engine's conformance test reads the header and holds the pair equal. EIGHT (operator sizing, 2026-09-01): a guitar sounds six strings, a piano or string-section voicing sits at eight. It is a SEARCH ceiling, not a cost — the detector abandons a round as soon as no credible candidate remains, so monophonic material costs what it always did.",
       "unit": "",
       "value": 8
     },
@@ -1489,14 +1505,14 @@
     "HRP_DEFAULT_MAX_AUTO_BANDS": {
       "rel": "data/kernels/hrp.json",
       "kind": "scalar",
-      "doc": "How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer \"will this fit?\" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. `eq-budget-fits.test.ts` asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.",
+      "doc": "How many bands HRP may plant on ONE channel — its half of the machine budget, and the number that makes the budget FORECASTABLE rather than emergent. Without it the count is whatever the material happens to provoke: the placement law is sparse, so a clean note plants two bands and a resonant room plants nine, and nobody can answer \"will this fit?\" until it does not. With FBS senior (§9.3(b)) that discovery would happen at showtime, to the junior planter, silently. Declared instead, the whole sum is arithmetic an operator can read before a note is played: operator bands + FBS maxAutoBands + HRP maxAutoBands <= EQ_MAX_BANDS EIGHT — ONE CORRECTION PER VOICE at full polyphony (operator sizing, 2026-09-01). The detector's ceiling is HRP_MAX_VOICES voices, a guitar's six strings and a piano/string voicing's eight, and a budget below the voice count would decide by arithmetic which of the notes actually sounding may be corrected — a limit nobody could read off the desk. Scarcity is handled where it belongs instead: the candidates from every voice meet in ONE globally ranked list (spec §29) and the head of it is what gets planted. It is NOT a claim to seniority, and it is larger than FBS's six. Seniority is the ALLOCATION ORDER — `HrpPlanter` takes only `EQ_MAX_BANDS − others.length`, the remainder after the operator and every other machine — so a bigger junior default cannot reach a senior slot; it only means HRP asks for more of what is left. the engine's conformance test asserts both halves. The per-channel DEFAULT; the row's travel is HRP_AUTO_BANDS_RANGE, so a channel that needs more can have more, as an explicit decision with a visible cost.",
       "unit": "",
       "value": 8
     },
     "LIMITER_LIMITS": {
       "rel": "data/kernels/limiter.json",
       "kind": "travels",
-      "doc": "The precision limiter's travels (`2026-09-27-precision-limiter.md` §2), rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by `harness/contract-limits-gen.mjs`, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth.",
+      "doc": "The precision limiter's travels, rendered into `omx_contract_limits.h` as `OMX_LIMITER_*` by the engine's limits generator, so `mix_limiter.h` reads them and never restates one (R-094). The ceiling is a TRUE peak in dBFS (the ×4 detector); the look-ahead sets the declared latency; the release is the envelope's time constant on the gain depth.",
       "shape": "table",
       "value": {
         "ceilingDb": {
@@ -1528,7 +1544,7 @@
     "BUS_INPUT_CAP": {
       "rel": "data/kernels/mixmatrix.json",
       "kind": "scalar",
-      "doc": "The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. `native-input-ceiling.test.ts` holds this to the C header; spec `docs/design/specs/2026-09-14-console-allocation.md` §9 carries the ruling and the measured cost of width.",
+      "doc": "The input cap a summing bus can be sized to — **the engine's declared input ceiling**, read from `@freemixer/declarations` rather than re-typed here. A gig with more than this many inputs needs more than one engine node; below it a bus is sized to `min(N, BUS_INPUT_CAP)`. It was a literal 64 until 2026-09-15, said to mirror the engine's `MAX_BUS_INPUTS` — and it did, but that 64 was the PipeWire builtin mixer's single-tier fan-in (8 × 8), never the engine's bound: the native node has allocated `OMX_MAX_STRIPS` (128) strips all along. Two literals agreeing by hand is how the console came to refuse a 96-channel desk it could carry. the engine's conformance test holds this to the C header; the console allocation design §9 carries the ruling and the measured cost of width.",
       "unit": "",
       "value": 128
     },
@@ -1721,7 +1737,7 @@
     "REVERB_PLATE_MOD_DEPTH_RANGE": {
       "rel": "data/kernels/reverb.json",
       "kind": "travels",
-      "doc": "PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it (`2026-09-26-native-fx-catalogue.md` §2, Reverb).",
+      "doc": "PLATE's tank MODULATION DEPTH, in percent of Dattorro's own excursion (±8 samples at his 29 761 Hz tuning rate, scaled to the live rate the way the tank lengths are). 100 is the depth the plate has always run at, so the come-up is bit-identical to the plate before the field; 0 brings back the unmodulated tank's ring; 400 is the wide, chorused EMT-style swing. The rate stays 1 Hz. Only PLATE reads it.",
       "shape": "travel",
       "value": {
         "min": 0,
@@ -1836,7 +1852,7 @@
     "REVERB_REVERSE_RANGE": {
       "rel": "data/kernels/reverb.json",
       "kind": "travels",
-      "doc": "REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds (`2026-07-15-native-delay-reverb.md`, Amendment 2026-09-17 §B). A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.",
+      "doc": "REVERSE's MIRROR WINDOW — how long a slice of the wet is played backward, in milliseconds. A DECLARED travel and never derived from `size`'s RT60: an RT60 of seconds cannot be buffered at 192 kHz, and a window silently clamped to fit is the silent-clamp defect. The ceiling is what `mix_reverb.h`'s pool budgets a ring for (`OMX_REVERB_REVERSE_MAX_MS`), so the console offers exactly the travel the kernel accepts — nothing wider, nothing narrower.",
       "shape": "travel",
       "value": {
         "min": 50,
@@ -2167,7 +2183,7 @@
     "RTA_TARGET_BIN_HZ": {
       "rel": "data/kernels/rta.json",
       "kind": "scalar",
-      "doc": "The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; `rta-window-derivation.test.ts` reads it back from the header.",
+      "doc": "The RTA's target bin resolution — generated into `omx_contract_limits.h` as `OMX_RTA_TARGET_BIN_HZ`, which `mix_dsp.h` reads; the engine's conformance test reads it back from the header.",
       "unit": "Hz",
       "value": 5.9
     },
@@ -2209,7 +2225,7 @@
     "TRANSIENT_LIMITS": {
       "rel": "data/kernels/transient.json",
       "kind": "travels",
-      "doc": "The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. `harness/contract-limits-gen.mjs` emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921).",
+      "doc": "The legal range of each transient-designer field (engineering units), transient spec §3. Every kind comes up flat: no per-kind default differs. the engine's limits generator emits each as `OMX_TRANSIENT_<FIELD>_{MIN,MAX,DEFAULT}` for `mix_transient.h`'s resolve; the row publishes them on OPTIONS when it lands (issue #921).",
       "shape": "table",
       "value": {
         "attackDb": {
@@ -2369,7 +2385,7 @@
     "TRIM_RANGE": {
       "rel": "data/kernels/trim.json",
       "kind": "travels",
-      "doc": "The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (openmixer's TRIM_RANGE, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.",
+      "doc": "The digital input trim, in dB: the console's fine head gain, applied first in a channel's lane as a click-free ramp from the last block's gain to this one (the console's trim range, -24..+24 dB in 0.1 dB steps, spelled minDb/maxDb/stepDb there). It comes up at 0 dB, the trim a channel has when none is set. omx-dsp's trim face takes it as `trim_db`.",
       "shape": "travel",
       "value": {
         "min": -24,
@@ -2549,7 +2565,7 @@
     "STANDARD_SAMPLE_RATES": {
       "rel": "data/rates.json",
       "kind": "list",
-      "doc": "The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` (`2026-07-15-clock-rate-honest-force-rate.md`) — this is the ceiling a write is refused by, never a menu.",
+      "doc": "The standard sample rates (Hz) the desk recognises — the pro-audio superset. One source of truth for the clock controller's `clock.allowed-rates` seed, the force row's validation ceiling and the device-rate probe fallback. What the operator is OFFERED is always the DEVICE's own `supportedRates` — this is the ceiling a write is refused by, never a menu.",
       "unit": "Hz",
       "value": [
         44100,
@@ -2886,6 +2902,11 @@
           "name": "hfRolloff",
           "kind": "choice",
           "global": "DRIVE_HF_ROLLOFFS"
+        },
+        {
+          "name": "oversample",
+          "kind": "choice",
+          "global": "DRIVE_OVERSAMPLES"
         }
       ]
     },