svelte-kit-vice/docs/architecture/sema.md

---
title: Sema — the perceptual engine and channels
type: reference
audience: human + agent
authority: E1 architecture — the semantic domain, the emit contract, the resolution cascade and the channels
status: current
source: migrated from src/uix/sema/README.md (2026-07-02, docs-book F7.2)
---

# Sema

`Sema` defines UIX's canonical semantic domain and orchestrates the emission
of perceptual signals.

## What it is

- canonical families (8): `contact`, `commit`, `signal`, `handle`, `emerge`,
  `shift`, `sustain`, `delegate`
- canonical intents: `neutral`, `affirm`, `fulfill`, `risk`, `threat`, `loss`
- a per-family intent policy (`SEMA_FAMILY_POLICY`) with **two axes**:
  `intentRequirement` (`'required'` | `'optional'` | `'forbidden'`) — the
  compile-time gate that shapes the discriminated union — and
  `intentGuidance` (`'expected'` | `'contextual'` | `'discouraged'`) — a
  doctrinal hint for lint/tooling
- normalization between the structured shape and the canonical label
- minimal domain validation
- `EngineSemantic` as the channel registry + occurrence dispatch

`Sema` does not decide which event happened. The provider decides.
`EngineSemantic` receives the occurrence and dispatches it to the registered
perceptual channels. Each channel materializes the signal in its modality
(DOM, audio, vibration).

## What it no longer is

`Sema` is no longer a monolithic multimodal runtime.

`EngineSemantic` does not contain:

- a global accessibility policy
- a cross-channel perceptual map
- decisions about which modal effect to apply

Those live in each channel separately:

- `EngineSemantic` keeps a channel registry and dispatches each signal
- the DOM projector materializes the signal as `data-event*` in the DOM
- `SoundChannel`, `HapticChannel` and future modal engines register as
  independent channels that receive the signal and decide how to materialize
  it on their plane

### Channel registry — open

The channel set is **NOT closed**. The framework ships canonical signatures
for `sound` and `haptic`. The `visual` channel exists as the projection/hold
meta-channel but has no slice of its own in `EffectiveSignature`. Apps can add
channels with a signature via declaration merging:

```ts
// app bootstrap
declare module '$uix/sema' {
	interface SemaChannelSignatures {
		a11y: A11ySignature; // narrator / live-region
		voice: VoiceSignature; // text-to-speech
	}
}
```

Only the visual channel is mandatory (with the `visual: false` escape).
Sound and haptic are opt-in. Any new channel registers its `Channel` and
receives the dispatch.

## The `emit` contract

```ts
semantic.emit(signal: SemanticSignal): Promise<string>
```

The resolved value is the **occurrence id**. Transient signals can ignore it;
a non-transient one (`untilFix` / `untilAction` / `stateBound`, see
[persistence](#persistence)) must be held so the caller can end it with
`clear(id)` or `clearTarget(el)` — there is no other handle on it.

One signature. It covers the three scenarios when composed with `dom.apply`:

```ts
// Structural change without a signal
dom.apply(change);

// Structural change with a signal
await semantic.emit(signal);
dom.apply(change);

// Signal without structural change
void semantic.emit(signal);
```

### Promise semantics — sequential strict

`emit(signal)` resolves when the `VisualChannel` has completed its whole
materialization:

1. `VisualChannel.prepare()` projected `data-event*` onto the DOM
2. the attributes lived in the DOM for the configured `hold`
3. the attributes were already removed (cleanup complete)
4. the Promise resolves

This is **sequential strict**: the caller applies the structural commit
AFTER the perceptual signal has finished. There is no parallelism between the
event and the state change.

Non-visual channels (sound, haptic) are fire-and-forget: they start in
parallel with the visual one but do not affect the Promise's timing.

### `emit`'s internal lifecycle

```
1. The engine generates the occurrence's id/session
2. The engine runs `channel.prepare(...)` on the registered channels
   - `VisualChannel.prepare()` projects `data-event*` for the cascade
3. The engine resolves the cascade and dispatches the signal to ALL registered channels
   - non-visual channels (sound, haptic) → fire-and-forget (not awaited)
   - the visual channel → awaited
4. The engine resolves the Promise when the visual channel has finished
   (strict sequential semantics: cleanup BEFORE the resolve)
```

### Channels as modules

Sema is organized in symmetric perceptual channels:

```
src/uix/sema/
├── engine.ts             registry + channel prepare/dispatch + cascade composition
├── resolver.ts           resolveSignature(signal, opts): EffectiveSignature
├── stamp.ts              stampEventAttrs / unstampEventAttrs (data-event-*)
├── channels.ts           channel ids, signatures and override types
├── sounds.ts             nominal sound repository + dynamic recipes
├── sema-map.ts           SEMA_MAP data + per-component Sema packs
├── components/           per-component perceptual packs (CSEM)
│   ├── dialog.ts         dialogSema — cascade rules + preloadSamples
│   ├── toast.ts          (future)
│   └── …
└── chans/
    ├── types.ts          Channel interface — handle(signal, effective)
    ├── visual.ts         VisualChannel (data-event projection + hold)
    ├── sound.ts          SoundChannel (gate + reduction policy; the Web Audio
    │                     runtime lives in the `$sound` art, injected)
    └── haptic.ts         HapticChannel (Vibration API + categorical kinds)
```

The engine does not mutate DOM attributes directly: it runs the generic
`channel.prepare(...)` hook. In the visual channel that hook delegates the
projection to a `SignalProjector`. When `ActiveUix` builds it, that projector
writes through UIX's `ActiveDom`. The engine resolves each signal into an
`EffectiveSignature` with `hold`, `sound`, `haptic` and future typed
channels, and dispatches `(signal, effective)` to each channel. Each channel
reads its slice (`effective.sound` for audio, `effective.haptic` for
vibration, etc.) or ignores the signature if it doesn't use it. Only the
visual channel blocks the caller with the perceptual hold; the rest are
fire-and-forget.

### Resolver and sema-map — the resolution cascade

On every `emit`, the engine:

1. Runs the channels' `prepare` hooks. The `VisualChannel` projects the
   semantic tokens `data-event`, `data-event-family`, `data-event-intent`,
   `data-event-direction`, `data-event-phase`, `data-event-id` onto
   `signal.target`. These are the tokens the cascade and eidos's CSS read.
2. Calls `resolveSignature(signal, opts)`, which applies the cascade
   (canonical numbering **1 · 2 · 3 · 4 · 5a · 5b** — the same in
   `engine.ts`, `resolver.ts` and CLAUDE.md; each layer overrides the
   previous):

```
THE SOUND — two lookups, and nothing modulates:

  nombre = per-emit ?? cascada ?? pack ?? morfo ?? familia[verbo] ?? familia.default
  sonido = pack[`${nombre}.${intent}`] ?? pack[nombre] ?? nada

EVERYTHING ELSE — a CSS-like cascade, each layer over the previous:

  1.  FAMILY base       — SEMA_MAP.families[f].base (haptic + activeChannels + hold)
  2.  INTENT deltas     — SEMA_MAP.intents[i].deltas; HAPTIC ONLY since 2026-08-06
  3.  MORFO overrides   — signal.overrides + signal.channels (minus `sound`)
  4.  RUNTIME overrides — engineOpts.overrides.runtime, baked into the map
  5a. PACK cascade      — engineOpts.components (minus their `sound`)
  5b. APP cascade       — engineOpts.overrides.cascade
```

**The sound does not participate in that cascade, and that is the design.**
An intent SELECTS a whole sound; it never bends one. A pack NAMES a sound; it
never authors one. There is nothing to override because there are no
parameters to override — a rule carries a name, and a name replaces a name.

What that removed, and why it is worth knowing before you go looking for the
machinery that used to be here: sound used to be resolved through all six
layers, with intent deltas adding to a family base signature and pack rules
replacing primitives afterwards. Measured across the 71 packs, that produced
214 rules authoring 33 distinct signatures, of which five were the same 700 Hz
note at five volumes and the «direction» variants differed by 20 Hz. The theory
was expressive; the output was mush. It also generated a whole class of defect
(D.7, the S-07 finding, three guards) whose only job was to police arithmetic
that should never have been in a component.

**Precedence among the things that NAME.** Most specific wins: a name passed at
trigger-time beats the app cascade, which beats a pack rule, which beats the
morfo event, which beats the family's verb table. This INVERTED the old order,
in which the cascade beat the per-emit override — whoever fires the occurrence
knows it best.

**The verb tier is what empties the packs.** `SEMA_MAP.families[f].sounds` is
keyed by the event's canonical VERB, with `default` as the fallback cell, so
`emerge` + `close` finds `close` without one line in any overlay's pack. The
verb travels in the signal for exactly this reason (it used to be computed and
discarded — audit S-40). A pack only speaks when it DIFFERS from the default:
the framework's 71 packs carry 30 sound rules between them.

3. Dispatches to each channel with the resolved signature.
4. Awaits the VisualChannel (which contributes the hold).
5. Runs the cleanup of the handles returned by `prepare`.

**Overrides vs deltas convention — numbers.** Applies to the HAPTIC channel
and to `hold`; the sound channel has no numbers to converge on any more.

- **Layer 2 (intent.deltas)**: `intensity: 0.2` adds to the base. Modifiers.
- **Layers 3, 4, 5a/5b (overrides)**: `intensity: 0.7` sets. Like CSS.
- To add explicitly from an override layer: `{ op: 'add', value: 0.1 }`.
- To multiply: `{ op: 'multiply', factor: 1.2 }`.
- To replace non-numeric primitives: `{ op: 'replace', value: ... }`.

If `signal.family` is missing or not in the map, it returns an empty
`EffectiveSignature` — the channels no-op.

### Semantic tokens — `data-event-*`

`VisualChannel.prepare()` projects the following attrs onto `signal.target`
BEFORE the cascade resolves. Rules with selectors over these attrs match
natively via `target.matches()` / `target.closest()`:

| Attr                   | Value                | Origin                            |
| ---------------------- | -------------------- | --------------------------------- |
| `data-event`           | `'signal-alert-close-fail'` | `signal.name`              |
| `data-event-family`    | `'signal'`           | `signal.family`                   |
| `data-event-intent`    | `'threat'`           | `signal.intent` (when present)    |
| `data-event-direction` | `'forward'`          | `signal.direction` (when present) |
| `data-event-phase`     | `'active'`           | while the hold lasts              |
| `data-event-id`        | `'sig-42'`           | occurrence id                     |

`data-event-direction` (`forward` / `backward`, `SemaDirection`) is the SENSE of
a traversal, decided per emit: the event name already separates
`shift-enter-mode` from `shift-exit-mode`, but one `shift-navigate` goes to the
previous month and the next one to the following month under the same name. It
is a sense, never an axis — eidos maps it onto the inline axis so `:dir(rtl)`
flips it. A route with no clear sense (a month picked from a select) stamps
nothing.

Those tokens are the **cross-channel contact surface**: sema's cascade
(`sound`, `haptic` and future channels) reads them with CSS selectors, the
same way eidos's CSS reads them to tint borders / animate states during the
hold. One perceptual surface, separate owners.

### The surface is ONE SLOT — ownership and `regime`

Those six attrs are a **single slot per element**. Two occurrences on one node
cannot both express, and the framework spent a year not saying so: three
independent audits found the same defect and none closed it (fable **S1**
2026-07-01 · sema **S-17** 2026-08-05 · blocks **A-36 / A-65**, reproduced in
the browser). Two rules now hold it shut.

**1 · The unstamp checks ownership.** `unstampEventAttrs` retires the
projection only when `data-event-id` still matches the signal that wrote it.
Before 2026-08-10 it took the signal and discarded it, so the occurrence
finishing FIRST wiped whichever one currently held the slot — measured on the
knob, a `commit-set` projection died 24 ms into a 240 ms hold.

**2 · `regime` says what an ARRIVING occurrence does to a busy surface**
(`SemaRegime`, declared per morfo event, default `replace`):

| Value | Behaviour |
| --- | --- |
| `replace` | Takes the surface at once, displacing the live projection. **Default** — and correct only since rule 1. |
| `queue` | Waits for the live occurrence's HOLD to elapse, then takes it. |

`queue` waits for the **hold**, never for the whole expression.
`awaitExpression` keeps the projection alive until the target's animations
finish (capped at `MAX_EXPRESSION_WAIT_MS`), which is right for unstamping and
wrong for queueing: gating on it delayed a Toggle's queued `commit-toggle` by
**1.6 s**, because unrelated transitions kept the node busy. The engine and the
channel resolve that number through one path (`VisualChannel.holdMsFor`).

**Reach for `regime` last.** It is for pairs the doctrine REQUIRES on one node
and no re-targeting can separate — a toggle's `contact-press` + `commit-toggle`
(its provider IS the button), the knob's `handle-drop` + `commit-set`. When the
collision comes from a **redirection** instead, the fix is to stop redirecting:
that is how A-36 closed, with the overlays' `emerge-open` giving up a
`targetOverride` left over from when they were `sequence: 'pre'`. Measured
after: the Button's `contact-activate` stamps the trigger and the Drawer's
`emerge-open` stamps its content — two surfaces, both expressing.

The redirection those overlays DID need — landing `emerge-close` on the trigger
once the content has unmounted — is not a `targetOverride` either: it is
**mount-state resolution, declared in the morfo as `targetFallback`** and
resolved by the runtime through the same single path that moves a11y focus
(`resolveEmitTarget`). A rule may select a fallback part — the census counts it
stampable — but it only fires in the degraded mount; the routine surface is
still the declared target. The field, its invariants and its split from
`allowedTargets`: [`architecture/morfo.md` §Step 5.5](./morfo.md).

`collapse` and `lock` were declared here from the founding commit and never
meant anything; both were retired on 2026-08-10 rather than left as a contract
that lies. The reasons are in `SemaRegime`'s own docblock.

### Cascade rules — flat CSS-like shape

```ts
interface SemaCascadeRule {
  selector: string                    // CSS selector — matches state attrs + event tokens
  priority?: number                   // CSS-specificity override (optional)
  channels?: readonly SemaChannelId[] // restricts / silences active channels
  sound?: …
  haptic?: …
}
```

One rule = one selector + one block of deltas. No intermediate
`overrides: { eventLabel: ... }` layer — the event's identity is read from
the selector via the `[data-event*]` tokens.

### Cascade selectors — typed builder, no hand-written strings

Cascade rules in `sema/components/*.ts` MUST build their `selector` via
`semaSelector(morfo, partKebab, matchers?)` from `$uix/morfo`. The helper
closes the loop between morfo's part/event contract and the selectors the
cascade evaluates.

```ts
import { semaSelector } from '$uix/morfo';
import { dialogMorfo } from '$uix/morfo/components/dialog';

const onContent = (matchers?: Parameters<typeof semaSelector<typeof dialogMorfo>>[2]) =>
	semaSelector(dialogMorfo, 'content', matchers);

cascade: [
	// [data-dialog-content][data-event-family="commit"]
	{ selector: onContent({ eventFamily: 'commit' }), haptic: { kind: 'tap' } },

	// [data-dialog-content][data-event="signal-alert-close-fail"]
	{ selector: onContent({ eventName: 'signal-alert-close-fail' }), sound: { sampleUrl: '/fail.wav' } },

	// [data-dialog-content][data-event^="emerge-close"][data-event-family="emerge"]
	{
		selector: onContent({ eventNamePrefix: 'emerge-close', eventFamily: 'emerge' }),
		sound: { contour: 'descending', pitch: { op: 'add', value: -150 } }
	}
];
```

Compile-time guarantees:

- `partKebab` is checked against `morfo.parts[].kebab`.
- `eventName` is checked against `morfo.events[].name`.
- `eventFamily` / `eventIntent` are typed against the canonical sema unions.
- `state` / `aria` / `pseudo` accept plain strings (the data-attr vocabulary
  is per-component and not yet typed-derived).

Renaming a part or event in morfo breaks the cascade at type-check time, not
silently in production. **Hand-written selector strings in cascade rules are
a code smell** — review them as drift.

### Per-component packs — `sema/components/{name}.ts`

Each component ships its perceptual-defaults pack in
`src/uix/sema/components/{name}.ts`, symmetric to soma and eidos:

```ts
import { semaSelector } from '$uix/morfo';
import { dialogMorfo } from '$uix/morfo/components/dialog';
import type { Sema } from '../sema-map';
import { sound, soundSampleUrls } from '../sounds';

export const dialogSema: Sema = {
	name: 'dialog',
	preloadSamples: soundSampleUrls(['notification.ping']),
	cascade: [
		{
			// `[data-dialog-content][data-event-intent="threat"]`, built from
			// the morfo contract — a part rename breaks at compile time.
			selector: semaSelector(dialogMorfo, 'content', { eventIntent: 'threat' }),
			sound: sound('notification.ping'),
			haptic: { kind: 'error', pattern: [50, 80, 50, 80, 50] }
		}
	]
};
```

The app imports the packs it uses (tree-shakable):

```ts
import { dialogSema } from '$uix/sema/components/dialog';

defineEngineSemantic({
	components: [dialogSema],
	overrides: {
		cascade: [
			// App-level rules win over packs on specificity ties. Apps use the
			// SAME typed builder; instance scoping goes in `ancestor`.
			{
				selector: semaSelector(dialogMorfo, 'content', { ancestor: '#critical' }),
				sound: { sampleUrl: '/x.wav' }
			}
		]
	}
});
```

`preloadSamples` is concatenated across all packs and passed to
`SoundChannel.preloadSamples()` when the engine is created, together with every
sample URL the map itself can reach — so installing a pack at runtime
(`applySoundPack`) warms it too, instead of leaving the first play of each name
to pay its own round trip. Preloading does NOT wait for the unlock gesture:
`decodeAudioData` works on a suspended context, while awaiting `resume()` before
the first click stays pending forever and downloaded nothing.

With `preferences.sound === 'off'` it downloads nothing at all — a pack that
cannot be heard is bytes spent on silence. The request is REMEMBERED, not
dropped: the level is read at dispatch, so the first audible play flushes the
deferred warm, and turning sound on does not hand back a cold cache.

The global unlock listener is not registered in the channel's constructor; it is
installed only once there is an `AudioContext` to unlock.

### When a component deserves a pack — participation doctrine (2026-07-07)

`Morfo.expression` (`SemaExpressionMode`, [`morfo/types.ts`](../../src/uix/morfo/types.ts),
author decision 2026-05-26) declares HOW a sema-scoped morfo contributes its
perceptual signature. THIS section is the canonical doctrine; it was
canonized at the component-audit checkpoint (verdicts S3a/S11 — historical
record in [`docs/audit/components/_veredictos.md`](../audit/components/_veredictos.md)):
four legitimate stances, each with shipped exemplars —

1. **`'pack'`** — a cascade in `sema/components/{kebab}.ts` NAMES the sound
   per event. Justified when the pattern ADDS meaning beyond family + intent
   deltas (date-field naming a fall for its clear; menubar's in-place note:
   "subtle commit + tap haptic for high-frequency top-level triggers"). A pack
   that only restates the family default is not a pack — it is noise, and the
   type makes that cheap to see: one word, or nothing.
2. **`'family-default'`** — family base + intent deltas suffice. Legitimate
   when (a) the gesture is generic contact (button), (b) the book itself
   counsels restraint (command, citing ch. 22 §8: celebrating at the click is
   *"celebrate before time"* — the outcome fires downstream), or (c) a
   composed child supplies the character (field-langs: the embedded
   ToggleGroup's pack puts the tap). The reason is written IN the morfo.
3. **`'delegated'`** — composite; expression lives in the children's packs.
   The composite declares ONLY the events its children don't own (date-picker:
   its own `commit-reset`; selection/commit sound through the embedded
   date-field/calendar). The anti-duplication rationale is radio-cards'
   morfo header: *"declaring them here would duplicate the contract"*.
4. **Declared debt** — `events: []` with an honest "yet" comment (the generic
   Picker) is better than silence but MUST carry a deadline or become a
   decision: an undated "yet" is a hole in disguise.

Coherence is guarded: `morfo-vocabulary-check` fails when a pack file exists
and `expression` declares anything other than `'pack'`, and warns when a pack
exists with no `expression` at all (verdict S11d).

### The sound pack — `SEMA_MAP.sounds`

**A component names a sound. It never writes a parameter.** The whole authoring
surface of the sound channel is one word, and only when the component differs
from its family's default:

```ts
{ selector: onTrigger({ eventName: 'contact-activate' }), sound: 'snap' }
```

`Sema['cascade'].sound` accepts a `SoundName` or `SILENT`, **and nothing else** —
the law is a type, not a convention. The same holds for the morfo's per-event
`overrides.sound`, for `overrides.cascade` (the app's door) and for the per-emit
option. Nobody authors a signature inline, not even the product: bringing a new
sound means REGISTERING it (declare the name, define it once) and then naming it.

#### The two kinds of entry

The pack has BASES (`tick`) and VARIANTS (`tick.threat`). Resolution is two
lookups: `name.intent`, then bare `name`, then silence. So a pack that ships only
bases works whole — the intent is ignored where no variant exists — and a pack
that cares adds variants only where it matters.

The default pack is **pure synthesis**: it works offline, pays no fetch and
cannot 404. A sample entry (`wav` / `mp3` / `ogg` / `aac`) carries its URL plus a
synthetic FALLBACK, so a missing file degrades to a designed sound rather than to
silence. Whether a name is a recipe or a file is not the component's business.

##### Two ways to DELIVER a file-based pack

The delivery is not part of the contract. A pack is
`Partial<Record<SoundName, NamedSound>>` either way, hot-swappable either way,
and a name that is not a `SoundName` is a compile error either way.

- **One file per name.** The pack points at URLs (`sample('/sounds/tick.mp3', …)`).
  The engine's cache is keyed by URL, so each file is fetched and decoded once;
  the browser caches each independently, and `preload: false` on an entry keeps
  it out of the warm-up so rare sounds load on demand.
- **One file for the whole pack.** A JSON of base64 payloads keyed by sound
  name, turned into a pack by `packFromBase64()` — the entries become `data:`
  URIs. Measured on `static/sounds/ui-inline.json`: **one** request instead of
  fifteen, and 69.082 B under brotli against 79.332 B for the loose files,
  because base64's +33 % is undone by compression while `audio/mpeg` is not
  compressed by anyone. The cost is granularity: the pack caches as one object,
  so one changed sound re-downloads all of them.

A `data:` URI is decoded IN PLACE — never fetched. A `connect-src 'self'` CSP
BLOCKS `fetch('data:…')` (the directive matches by scheme, and `'self'` does not
cover `data:`) while letting a same-origin path through, so routing an inline
pack through `fetch` made it degrade silently to synthesis in exactly the
environment that made someone inline it. It is also ~6x slower.

#### Designed to be told apart

`sound-names.test.ts` fails when two entries are not separated on at least TWO
perceptual axes — a fifth of pitch, a third of brightness, half again as long,
a different contour, double the level, or 0.2 of roughness.

That guard exists because its absence was measured, by ear, by the author: a
previous catalogue had five names that were the same 700 Hz note at five
volumes, and every test passed. A snapshot could not have caught it — each value
was exactly what the table said. **The missing property was a RELATION between
entries**, and that is what the guard asserts.

#### Verbs, not components

Names describe an occurrence, never a widget: `open`, `tick`, `alert`. A library
organised by component (`button_hard`, `panel_expand`, `window_open`) needs
TRANSLATING into this vocabulary, and several of its files will compete for one
name — that is normal, and the choosing is where a pack author's ear works. See
`packs/ui-mp3.ts` for a worked example, including what it could not use.

#### Samples: what a file can and cannot carry

`playSample` reads **exactly two** fields of the resolved signature: `sampleUrl`
and `gain`. There is no `playbackRate` and no `detune`. So over a recording the
intent can only move the LEVEL.

Consequence, and it is physics rather than policy: **a sample that must carry
evaluative weight needs one file per intent** — `tick.mp3` AND
`tick.threat.mp3`, not one file bent two ways. With synthesis the problem does
not arise, because each variant is designed whole.

#### Silence is a value, not a name

`SILENT` is THE canonical silence — one value for the whole system, exported
from `$uix/sema`. Set a channel slice to it and **the channel that owns the
slice ignores it and dispatches nothing downstream**: no signature is
resolved, no engine is called, no context is opened.

It admits no family, no verb and no intent, because there is nothing to
modulate. That is why it is not a key: a per-family silencer would need to
know which base it is cancelling, which is how the catalogue ended up with
three of them.

```ts
// a pack rule that does not speak in sound
{ selector: onProvider({ eventName: 'emerge-open' }), sound: SILENT }
```

**Absorbing by construction.** Intent deltas do not apply on top of `SILENT`
(layer 2 cannot un-silence). A later override CAN lift it by REPLACING the
slice with a full signature — declaring sound is a decision, not a delta.

**Why it is a value and not a zero.** Until 2026-08-06 silence was arithmetic:
`{family}.silent` subtracted the family's base gain so neutral landed at 0.
A zero gain is a NUMBER, and later layers move numbers. Measured on a
"silenced" toggle: an intent delta that raises `roughness` past the AM
threshold (`risk` +0.2, `threat` +0.4) turned the signature into a raw tremolo
at **−13.9 dBFS** (`risk`) and **−6.7 dBFS** (`threat`) — louder than a real
button press at −9.3. The engine had already written the rule it was breaking:
*«muting DROPS the channel rather than scaling to zero, because a `0` still
buzzes»* (`engine.ts`).

The grammar and the ban on silence-as-arithmetic are enforced by
`src/uix/sema/sounds-grammar.test.ts`, which reads the families from `event.ts`
and the base gains from `sema-map.ts` rather than restating them; the runtime
half (`SILENT` reaches no engine) is pinned in `chans/sound.test.ts`. The guard
exists because `form.*` named seven form controls truthfully on 2026-05-19, was
extended to menus and trees a week later, and reached 50 packs — one of which
was the Form — before anyone noticed, since no check ever looked at these keys.

### Channels and signatures

| Channel  | Consumed slice                                                      | Behavior when not applicable                                     |
| -------- | ------------------------------------------------------------------- | ---------------------------------------------------------------- |
| `visual` | `effective.hold` for the hold                                       | Falls back to the family table (SEMA_DURATIONS) or `defaultHold` |
| `sound`  | `effective.sound` (skips if `'sound'` not in activeChannels)        | no-op                                                            |
| `haptic` | `effective.haptic` (Vibration API, honors `prefers-reduced-motion`) | no-op if there is no `navigator.vibrate`                         |
| (custom) | declaration merging of `SemaChannelSignatures`                      | read by the registered channel                                   |

### The visual channel's attr namespace

`VisualChannel.prepare()` projects **only** attributes under the
`data-event-*` prefix:

| Attr                   | When                  |
| ---------------------- | --------------------- |
| `data-event`           | always                |
| `data-event-id`        | always                |
| `data-event-phase`     | always (`'active'`)   |
| `data-event-family`    | if `signal.family`    |
| `data-event-intent`    | if `signal.intent`    |
| `data-event-direction` | if `signal.direction` |

**Rule**: the channel never touches state attrs (`data-state`,
`data-intent`, `data-disabled`, ...). State is managed by the runtime/morfo.
Reason: state is persistent and a signal is transient; reusing the same name
would force the channel to erase state on cleanup (or into fragile
save/restore if state mutates during the hold).

For CSS:

- `[data-event-intent='risk']` → reacts to the transient occurrence's intent
- `[data-intent='risk']` → reacts to the component's persistent state

Both can coexist on the same element with distinct semantics.

### Hold — the visual channel's resolution chain

The `VisualChannel` resolves its `hold` (how long the `data-event-*` live in
the DOM) in this order:

1. `signal.hold` — imperative per-call override.
2. `effective.hold` — composed by the resolver from
   `resolveHoldsByIntent(family, intent)` (`holds.ts`,
   `SEMA_HOLDS_BY_INTENT` — the ONE canonical hold source; the per-family
   `SEMA_MAP.families[*].hold` field was REMOVED 2026-07-06 because it
   duplicated this table and had drifted from the book's regions).
3. The same canonical table, consulted directly by the channel — defense
   when a custom map's resolver path didn't compose a hold.
4. The channel's global `defaultHold` (`brief`, 240ms by default).

See `holds.ts` (`SEMA_HOLDS_BY_INTENT`) for the concrete family+intent values.

**The hold timer runs on the managed scheduler, not on raw `setTimeout`.**
The engine receives `timers` (in `ActiveUix` it is `uix.timers`) and forwards
it to the three channels: the `VisualChannel`'s hold, the `HapticChannel`'s
`delay` and the `SoundChannel`'s earcon duration are scheduled via
`uix.timers.schedule(...)` (the `semaDelay` helper in
`src/uix/sema/timers.ts`). That makes perceptual timing cancelable on
`dispose`, observable, and deterministic under a fake clock in tests. It only
falls back to `setTimeout` when a channel is built without a scheduler
(direct unit tests); production always injects the scheduler.

> **Overlay openings — `sequence: 'post'`, not `'pre'`.** An appearance event
> whose provider sets `open` in the HANDLER (Popover, Dialog and Drawer's
> `emerge-open`) MUST declare `sequence: 'post'`. With `'pre'` the
> runtime awaits the emit — and therefore the ~240ms hold — BEFORE the
> handler, gating the content mount behind the hold: the overlay opens late
> and its first render lands inside the hold's `setTimeout` turn (the
> "setTimeout handler took N ms" violation). Same doctrine as the
> checkbox-lag fix. Closing (`emerge-close`) stays `'pre'`: there the element exists
> and the signal MUST precede the unmount.

```ts
// Per-signal override
semantic.emit({ ..., hold: 1200 })

// Override the visual channel's global default
const semantic = new EngineSemantic({ dom, visual: { defaultHold: 400 } })

// Disable visual (DOM-less environments)
const semantic = new EngineSemantic({ visual: false })

// Enable the built-in sound channel (opt-in: audible side effect)
const semantic = new EngineSemantic({ dom, sound: true })

// With SoundChannel options
const semantic = new EngineSemantic({
	dom,
	sound: { masterGain: 0.6 }
})

// Enable the built-in haptic channel (opt-in: device feedback)
const semantic = new EngineSemantic({ dom, haptic: true })

// With HapticChannel options
const semantic = new EngineSemantic({
	dom,
	haptic: { masterIntensity: 0.7 }
})

// Register custom channels (a11y, voice, future)
class A11yChannel implements Channel {
	readonly id = 'a11y'
	async handle(signal, effective) { /* live-region updates, etc. */ }
}
semantic.register(new A11yChannel())
```

### Override layers — recipes

```ts
import { dialogSema } from '$uix/sema/components/dialog';

const semantic = new EngineSemantic({
	dom,
	sound: true,
	haptic: true,

	// Layer 5a — component packs (defaults shipped with each component)
	components: [dialogSema /* , toastSema, drawerSema, … */],

	overrides: {
		// Layer 4 — targeted SEMA_MAP edits, valid app-wide
		runtime: {
			'families.commit.base.sound.pitch': 850,
			'intents.threat.deltas.haptic.intensity': 0.4
		},

		// Layer 5b — app CSS-like rules, matched against signal.target
		// AFTER the component packs
		cascade: [
			{
				// Typed builder here too — `ancestor` carries the instance id
				// (it lives outside the morfo's contract: plain string by design).
				selector: semaSelector(dialogMorfo, 'content', {
					eventIntent: 'threat',
					ancestor: '#delete-confirm-dialog'
				}),
				sound: { sampleUrl: '/sounds/scary.wav', gain: 0.25 },
				haptic: { kind: 'error', pattern: [50, 80, 50, 80, 50] }
			},
			{
				// Targets NO morfo-emitted attr (a global reduce gate over the
				// engine's own `data-event-*` stamps) — legitimately hand-written;
				// the builder mandate covers morfo-targeting selectors only.
				selector: ':root[data-sound="reduce"] [data-event-phase="active"]',
				priority: 100,
				sound: { gain: 0.05 },
				channels: ['sound']
			}
		]
	}
});
```

```ts
// Layer 3 — per-event override declared in the component's own morfo.
// Propagates via SomaRuntime → SemanticSignal → resolver. The app can
// still override from the cascade (layers 5a/5b).

// src/uix/morfo/components/dialog.ts
{
	name: 'signal-alert-close-fail',
	semantic: {
		family: 'signal',
		verb: 'alert',
		target: v.partRef('content'),
		sequence: 'pre',
		intent: 'threat',
		// Per-event silencing: no sound, to avoid competing with the live region
		channels: ['haptic'],
		// Per-event override: the Dialog's own sample
		overrides: {
			haptic: { kind: 'error', pattern: [60, 80, 60, 80, 60] }
		}
	}
}
```

### Specificity between rules

Like CSS:

- **Computed selector specificity** — IDs × 100, attributes × 10, classes ×
  10, pseudo-classes × 10, elements × 1.
- **Rules apply in ascending specificity order** — the last applied wins each
  pointwise conflict.
- **Tie** → declaration order. Component packs appear BEFORE
  `overrides.cascade`, so app rules win ties.
- **`priority?: number`** — overrides the computed value for cases that must
  win without counting attributes (typically accessibility, `priority: 100+`).

To vary by event, intent, family, name, etc. — everything goes through the
typed builder's matchers, which stamp the `[data-event-*]` tokens. (The
hand-written examples this section used to show targeted `[data-toast-root]`
— a part that never existed in the toast morfo: exactly the silent drift
`semaSelector` turns into a compile error.)

```ts
// Vary by the event's intent
{ selector: semaSelector(toastMorfo, 'item', { eventIntent: 'threat' }), haptic: { kind: 'error' } }

// Vary by family
{ selector: semaSelector(toastMorfo, 'item', { eventFamily: 'signal' }), sound: { gain: 0.4 } }

// Vary by exact event name (typed against the morfo's declared events)
{ selector: semaSelector(toastMorfo, 'item', { eventName: 'emerge-dismiss' }), sound: { sampleUrl: '/dismiss.wav' } }

// Vary by event prefix
{ selector: semaSelector(dialogMorfo, 'content', { eventNamePrefix: 'emerge-close' }), sound: { contour: 'descending' } }
```

### SoundChannel — doctrine here, machine in `$sound`

> **The Web Audio machinery does NOT live in sema** (since 2026-07-30). The
> context lifecycle, the synthesis graph, the unlock-on-gesture and the sample
> path were extracted to the art [`$sound`](../../src/arts/sound/README.md)
> (`EngineSound`) — they were ~63% of a 407-line "channel" and are machinery,
> not perceptual doctrine. Same move `$motion` made out of eidos, for the same
> reason and with the same result: the art owns the RUNTIME, the layer keeps its
> DATA and doctrine.

What the channel keeps — all of it doctrine:

- the `prepare` gate (does this signal admit sound at all?),
- the per-channel reduction policy (BK-REDUCTIONS): `off` silences (meaning
  migrates via `SEMA_MIGRATION.sound` → presence / live region), `reduce`
  attenuates,
- **its voice** — `SEMA_SOUND_VOICE` in `sounds.ts`, the synthesis calibration
  sema's vocabulary was tuned against, registered on the engine at channel
  construction (redesign 2026-07-31, D-SR.3 in `PLAN-sound-redesign.md`; it was
  baked into the art as constants before). Earcons play on the engine's **`ui`
  bus** with that voice — so UI-sound policy (BK-REDUCTIONS, resolved here and
  handed over as `gainScale`) can never touch the content's volume; the `ui`
  bus itself is the art's graph-side lever, with no direct prefs wiring today,
- and the division of labour: **the channel resolves the LEVEL, the engine
  applies the gain**.

Everything else — `SOUNDS` (the named catalogue), the gesture resolvers, the
cascade — is unchanged and still sema's.

**How it reaches the engine.** `EngineSemantic` takes a `soundEngine` option and
forwards it. Both boot modes fill it without the integrator doing anything:
standalone `ActiveUix` creates the engine and passes it in (exposing it as
`uix.sound`); in attach, `defineEngineSemantic()` declares `sound` as a service
dependency and forwards whatever `defineUixServices()` registered. The engine
arrives as a structural port, the `MotionDom` / `SceneDom` pattern — sema never
reaches for it. (**Precision**: the *instance* is injected, but the module graph
is not free of the art — `chans/sound.ts` imports `createEngineSound` as a value
for the private-engine fallback, so an app that builds an `EngineSemantic` still
bundles the synthesizer with sound off. See `PLAN-sound-engine.md` §16.)
Without an injected engine the channel creates a private one and owns
its lifecycle — *whoever creates, disposes*; a shared engine is never closed by
a consumer. That last rule is what lets a page rebuild its `EngineSemantic`
freely: the sema studio does it on every draft edit, and the context survives.

**Why one engine matters.** Browsers cap concurrent `AudioContext`s and the
autoplay unlock gesture is per-context, so a second context leaves one of the
two mute. An app that also plays content (a media player, a waveform) must take
`uix.sound` — or pass `soundEngine` to its own `EngineSemantic` — instead of
opening its own. The art warns when a second context goes live.

Behaviour, unchanged by the extraction:

- **Prepare-time priming**, no constructor side effect: the context is created +
  resumed in `prepare()` when the signal admits `sound`, synchronously inside
  the user gesture.
- The unlock listener (`pointerdown` / `mousedown` / `click` / `touchstart` /
  `keydown`) is registered only AFTER the context exists, through the injected
  DOM surface. An engine that never plays installs no global listeners.
- Synthesis: two oscillators (sine + a fifth) → biquad lowpass (centroid) →
  `roughness > 0.2` inserts a tremolo stage → ADSR-lite envelope; `contour`
  (`flat` / `ascending` / `descending` / `arc` / `bell`) rides `osc.detune`.
  The tremolo is a gain node IN SERIES at `1 - depth`, modulated on its own
  `.gain`, so the factor peaks at 1 and the trill scales WITH the envelope.
  It used to be patched onto `envelope.gain`, where a connection ADDS to the
  automation instead of scaling it: the depth stayed constant while the
  envelope moved, closing every `signal` note on a non-zero sample (an audible
  tick) and inflating a `commit.subtle` charged with `threat` to seven times
  its declared level. Pinned by `engine-sound.test.ts`.
- A signature carrying `sampleUrl` plays the sample (with an `AudioBuffer`
  cache) and **falls back to synthesis** on fetch / decode failure.
- Any failure is absorbed — sema is ornamental.

The prepare-time priming pattern applies in general to any channel whose
backend has a "first time must happen inside a gesture" restriction: audio,
vibration, fullscreen, clipboard write. Documented as **convention 12** in
[`guia-semantica-historica.md`](../decisions/guia-semantica-historica.md)
(historical seed).

### Error policy

- Errors in NON-visual channels are logged but never propagate. Sema is
  ornamental: an audio-context or vibration-API failure must not abort the
  provider's operation.
- If the visual channel throws, the `emit` Promise rejects. The caller
  decides.
- In fire-and-forget (`void semantic.emit(...)`), a visual rejection
  propagates as an unhandled promise — a conscious policy.

## Perceptual arbitration, reductions and accessibility (C-series, 2026-07-04)

Landed from the Sema↔book audit (`sema-findings.md`, a process registry since
removed from the tree); anchors in [`book-map.md`](../book-map.md). All
opt-outable, all covered by unit tests.

### Frequency memory — `engine.frequencyMemory` (default on)

`BK-FREQ-MEMORY` (ch. 32 §11): _"la gramática debe tener memoria de frecuencia"_.
Repeated occurrences of the **same** event identity (`signal.name`, else
`family:intent`) attenuate their non-visual channels — `sound.gain` and
`haptic.intensity` — past a small threshold, down to a floor, so the 50th
autosave doesn't sound like the first. `threat` is exempt; the visual hold is
never touched. A window of silence resets the key to full. Distinct from the
pointer-rate throttle (that samples ONE continuous gesture; this damps _repeated_
discrete signals). Disable with `frequencyMemory: false`.

### Dominance arbiter — `engine.dominance` (default on)

`BK-DOMINANCE` (ch. 30 §7): when signals overlap, the engine ranks each active
occurrence (evaluable `commit`/`signal` > structural; higher activation from
intent; recency breaks ties for the newcomer) and **mutes the non-visual
channels** of an incoming signal a still-active occurrence out-ranks — "lo
evaluable calla a lo estructural". `threat` is never muted; the visual hold and
the `announce` channel survive (a dominated event still reads structurally and
remains accessible). Occurrences stay active for a short overlap window. Disable
with `dominance: false`.

### Announce channel — `engine.announce` (opt-in)

`BK-SIGNAL-A11Y` / `BK-A11Y-CRITICAL`: the accessible live region is now a
first-class **channel** (`AnnounceChannel`, id `announce`), opt-in like sound /
haptic. It reads the human-facing text from `signal.message` (never inferred from
the technical event name), derives priority from intent (`threat`/`loss` →
`assertive`, else `polite`), and materializes it either through an injected
announcer (`{ announce }`, e.g. `ActiveUix.announce`, one shared region) or a
self-owned pair of live regions built from the injected `dom`. Fire-and-forget,
never rejects. Supersedes `book-deviations.md` D.8's deferral.

**Live-region doctrine — who owns what (AUX-1 / SEM-1, 2026-07-11).** The
framework has ONE sink, two emitters, and an app-level component:

- `uix.announce` (ActiveUix) owns the SHARED pair of live regions — the
  low-level sink (its zwsp-toggle re-announce is an implementation detail of
  that sink, not a third policy).
- The two framework **emitters** both prefer delegating to it: the soma
  runtime's a11y commitments (`a11ySemantic.requiresLiveRegion` →
  `sources.announce`) and this channel; its self-owned clear-then-set regions
  are ONLY the no-uix fallback.

  The two emitters serve **different publics, and that is why nothing is
  announced twice**: the runtime path covers components built on soma, and this
  channel covers an app driving `EngineSemantic` WITHOUT soma, which writes its
  own signal and puts the text in `signal.message` itself. The soma runtime
  deliberately does NOT forward `message` into the perceptual signal — its
  a11y commitment is already discharged through `sources.announce`.

  ✅ **The delegation is now a literal** (S-19, fixed 2026-08-13). `AnnounceFn`
  takes the priority POSITIONALLY, like the sink and like soma's source, so
  `{ announce: uix.announce }` compiles as written — no adapter. It used to take
  an options bag, and forcing the wiring through a cast handed the sink an
  OBJECT where it reads a priority: `liveRegionIds[obj]` is `undefined`, so every
  announcement, `threat` included, landed in the polite region and stopped
  interrupting.

  ✅ **Both roots wire it by DEFAULT** (S-19(ii), signed 2026-08-13). The
  composition root is the only place that holds both ends — the shared sink
  and the engine — and at engine-construction time `uix.announce` does not
  exist yet, so any wiring an app could write fell back to the channel's
  self-owned regions: a SECOND live pair, the very thing this section forbids.
  `ActiveUix` therefore registers the channel post-construction with a
  late-bound closure, in standalone and attach alike. Default ON because
  announce is SUBSTITUTION, not ornament (§channels) — the a11y announcer
  ships ambient, the field norm (Angular CDK `LiveAnnouncer`, React Aria).
  Opt out with `events: {{ announce: false }}`; an explicitly passed
  `announce` wins (the root never overwrites an existing channel). Soma
  components still announce exactly once: their runtime keeps `message` out
  of the perceptual signal, so the channel no-ops for them.
- The soma `<Announce>` component is different in kind: an app-authored,
  declarative region for app-level messages — not a vehicle for framework
  event announcements.

Priority is ONE policy everywhere (SEM-1 — the runtime used to add a
`family === 'signal'` conjunct, making a threat-tinted commit announce
polite on one path and assertive on the other): the evaluative INTENT alone
— `threat` / `loss` → `assertive`, else `polite` — implemented identically
in `priorityForIntent` (`chans/announce.ts`) and the runtime's a11y step
(`runtime.svelte.ts`), pinned by tests on both sides
(`announce.test.ts` · `runtime.svelte.test.ts` SEM-1).

### Per-channel reductions — `engine.preferences`

`BK-REDUCTIONS` (ch. 32 §12 / ch. 33 §7): a symmetric reduction story per
channel. `SemaPreferences { sound?, haptic?: 'full' | 'reduce' | 'off' }` is read
at dispatch (back it with reactive state to make it live). Sound gains a
reduce/off path (attenuate master gain / no-op) symmetric to haptic; haptic
_also_ keeps honoring `prefers-reduced-motion`. For sound, `off` governs the
NETWORK as well as the output: no sample is downloaded while it holds (see
`preloadSamples` above). Visual-motion reduction stays in
eidos CSS; when a channel is `off`, meaning migrates via the table below.

### Migration table — `sema/migration.ts`

`BK-A11Y-MIGRATION` (ch. 33): _"cuando un canal falla, migra el significado, no
el adorno"_ — as **data**, not just prose. `SEMA_MIGRATION` gives per-modality
fallback surfaces (motion → state/focus/text; color → shape/icon/text; sound →
presence/text/liveRegion; haptic → motion/shape/state); `SEMA_MIGRATION_BY_FAMILY`
refines the cases the book calls out (signal-sound → live region; commit-color →
footprint). `resolveFallback(modality, family?)` is the accessor a11y / eidos
consult.

### Frame-intent guardrail (morfo)

`BK-FRAME-NO-INTENT` (Apéndice A): a structural family (`contact` / `emerge` /
`shift` / `sustain` / `delegate`) that declares a non-neutral intent must justify
it with `MorfoEventSemantic.intentRationale`, or `validateMorfo` rejects it. This
makes the "appearance that is the warning" exception ([book-updates
A-1](../audit/book-updates.md)) explicit and auditable rather than a silent
loophole. `handle` is exempt (the book puts intent on `handle.drop`);
`commit`/`signal` require intent anyway.

## Continuous components (drag, swipe, hold)

Most events are discrete: one gesture, one emit, one hold. Direct-manipulation
components — Slider, Drawer, and any future knob / swipe surface — are
**continuous**: the pointer moves at sample rate (dozens to hundreds of events
per second) while the value updates the whole time. Emitting one perceptual
signal per pointer sample would flood the channels and desynchronise the hold.
The rules below keep a continuous gesture perceptually legible without breaking
the emit contract.

### One door: `runtime.trigger(eventName, opts?)`

There is no `emitEvent`. Providers reach the engine through a single method,
`runtime.trigger(name)`, which resolves the morfo's compiled action, applies
polymorphic overrides, runs the handler in the declared `sequence`, awaits the
perceptual emit, and honours the event's a11y commitments. It returns a
`Promise` that represents **the whole occurrence, including the visual hold** —
the same sequential-strict window described in _Promise semantics_.

That Promise is the control the caller uses to opt into or out of the hold:

- `void runtime.trigger('handle-drag')` — **fire-and-forget**. The signal
  starts; the caller does not wait for the hold. Correct for high-frequency
  emits, where blocking the gesture loop on a ~240ms hold would be absurd.
- `await runtime.trigger('emerge-close', …)` — **blocking**. The caller waits for
  the hold to finish before its next step. Correct when a structural change must
  observe the resolved signal (e.g. an `emerge-close` whose element unmounts after
  the pulse — see the `sequence: 'post'` note under _Hold_).

Slider and Drawer go through `trigger` exclusively — `handle-pick`,
`handle-drag`, `commit-set` on the slider; `handle-pick`,
`handle-drag-progress`, `handle-drop` on the drawer. Every continuous one is
`void`.

### `coincident` vs `post` for a moving value

A continuous gesture has two temporally distinct moments, and the morfo encodes
each with its `sequence`:

- **The move** (`handle-drag`, `handle-drag-progress`) is `sequence: 'coincident'` —
  the perceptual signal and the value update are indivisible. The user _is_ the
  value changing, so the signal fires alongside the mutation, neither
  anticipating nor trailing it. (`coincident` is emit-then-handler, like `pre`;
  the two are just declared indivisible.)
- **The commit** (`commit-set`) is `sequence: 'post'` — the handler settles the
  resolved value first, then the signal celebrates it. The move is felt
  continuously; the commit is felt once, after. (`post` is handler-then-emit;
  see the `sequence` comment in `runtime.svelte.ts` and the overlay note under
  _Hold_.)

Slider is the worked example: its morfo declares `handle-pick` / `handle-drag`
as `handle` · `coincident` and `commit-set` as `commit` · `post`.

### Throttle continuous pointer emits; leave keyboard alone

`onpointermove` fires far faster than the eye or ear resolves. Continuous emits
are therefore **coalesced to animation frames** and floored to a
component-tuned minimum interval — both via `ActiveDom`'s `requestFrame` (the
sanctioned rAF vehicle; never a raw `requestAnimationFrame` or `setTimeout`):

- Slider's `queueHandleDrag` stores the pending value, schedules one
  `requestFrame`, and inside it skips the emit when less than
  `SliderProvider.DRAG_SIGNAL_MS` has elapsed since the last signal.
- Drawer's `flushDragProgress` does the same with `DRAG_PROGRESS_SIGNAL_MS`,
  deliberately set _below_ rAF cadence (sparser than one signal per frame).

The concrete millisecond floors are perceptual tuning, not architecture — they
live at those constants in the providers, never copied here.

Keyboard is the opposite case and is **not** throttled: Slider's `onkeydown`
calls `commitValue()` → `trigger('commit-set')` on every arrow / Page / Home /
End press. OS key-repeat is already human-paced and every repeat is a discrete
committed value, so one emit per keydown is correct. Throttling is a property of
_sub-perceptual pointer sampling_, not of continuous components in general.

### The per-emit payload owns the primitives; the cascade only adds character

`handle-drag` carries a per-emit payload through `trigger`'s `overrides` —
since the 2026-08-06 redesign, a sound NAME (`'step'`: the gesture sounds by
REPETITION, so the cadence of the drag is the cadence of the ratchet) plus the
HAPTIC primitives derived from pointer velocity. The gesture resolvers that
used to synthesise pitch/gain/contour from position+velocity died with that
trade-off and were retired on 2026-08-12 (ledger
`docs/process/AUDIT-docs-code-ledger.md` §D10) — nothing authors sound
parameters per-emit any more. For the live payload to survive frame after
frame, the component's sema pack for the continuous event sets **only**
`channels: ['sound', 'haptic']` and overrides nothing else
(`sema/components/slider.ts`).

> A cascade rule for a continuous event MUST NOT set `sound` (it would replace
> the per-emit name on every frame) nor the per-intent haptic `intensity` (it
> would flatten the live velocity curve to a constant buzz). The rule enables
> the channels and adds character — it never fixes the gesture's payload.
> (Same division as _Override layers_ and the canon rule "cascade rules add
> character, never the intent's evaluative profile".)

## The per-family intent policy — `SEMA_FAMILY_POLICY`

> **Canonical:** [`CANON.md`](../CANON.md) §4 is the single source for this
> policy. The shape below mirrors `SEMA_FAMILY_POLICY` in `types.ts`; if they
> ever disagree, the code + canon win.

The doctrine on when intent is mandatory lives in a const in
`src/uix/sema/types.ts`. Each family declares **two independent axes**:

```ts
export const SEMA_FAMILY_POLICY = {
	contact: { intentRequirement: 'optional', intentGuidance: 'discouraged' },
	commit: { intentRequirement: 'required', intentGuidance: 'expected' },
	signal: { intentRequirement: 'required', intentGuidance: 'expected' },
	handle: { intentRequirement: 'optional', intentGuidance: 'contextual' },
	emerge: { intentRequirement: 'optional', intentGuidance: 'contextual' },
	shift: { intentRequirement: 'optional', intentGuidance: 'contextual' },
	sustain: { intentRequirement: 'optional', intentGuidance: 'contextual' },
	delegate: { intentRequirement: 'optional', intentGuidance: 'contextual' }
} as const;
```

- **`intentRequirement`** (`'required' | 'optional' | 'forbidden'`) — the
  compile-time gate that shapes the discriminated union. `'required'` (only
  `commit` and `signal`) → `intent` is MANDATORY in `MorfoEventSemantic`.
  `'forbidden'` is reserved; no family uses it today.
- **`intentGuidance`** (`'expected' | 'contextual' | 'discouraged'`) — a
  doctrinal hint with no type effect. `commit`/`signal` are `'expected'`;
  `contact` is `'discouraged'` (book ch. 22 §11: "strong intent should not
  live in the contact"); the rest are `'contextual'`.

**How it is enforced:**

1. **Compile time** — `SemaEvent` and `MorfoEventSemantic` are discriminated
   unions derived from the `intentRequirement` axis. Flipping a family from
   `'optional'` to `'required'` forces every morfo of that family to declare
   an intent or fail the typecheck.

2. **Runtime** — `validateSemaEvent` throws when an event of a family with
   `intentRequirement: 'required'` is built without intent (a defense against
   malformed morfos or external inputs).

**Why this policy replaced the valenced/transitional split:**

The original canonical doctrine assumed only valenced families (contact,
commit, signal, handle) could declare intent. Transitional ones (emerge,
shift, sustain) were intent-less by definition. UX reality contradicted it: a
Dialog opening to confirm a destructive delete carries threat in its very
appearance. The policy distinguishes by the practical NEED for intent, not by
taxonomic category.

The valenced/transitional distinction still exists as a classification, but
it no longer dictates the intent rules — the policy does.

## The canonical verb vocabulary (`SEMA_VERBS`)

> **Canonical:** [`CANON.md`](../CANON.md) §6 +
> [`verbs.ts`](../../src/uix/sema/verbs.ts) are the source of truth for the
> verb vocabulary. This section documents how sema _consumes_ it (validation,
> naming shapes).

Cross-component action verbs grouped by family. The canon lives in
[`verbs.ts`](../../src/uix/sema/verbs.ts) and reflects the book _Diseñando lo
que ocurre_ ch. 22–29 (families) + ch. 10 (intents).
`morfo.events[].semantic.verb` MUST be in this vocabulary;
`morfo.events[].name` MUST follow the `{family}-{verb}[-{nuance}]` shape
(guaranteed by `validateMorfo`) so sema/sound/haptic can subscribe by verb and
eidos can write transversal selectors (`[data-event^="emerge-dismiss"]`).

> **The literal table used to live here, and it rotted.** It was missing
> `commit.unselect` and `handle.zoom` — both live in `verbs.ts` and both used
> by real morfos. A copy of a closed set is a copy that goes stale; the
> generated enumeration in
> [`canon/vocabularies.md`](../canon/vocabularies.md#sema-verbs-by-family)
> is regenerated from the const by `npm run docs:vocabularies` and guarded for
> freshness by `npm run docs:check`. Removed 2026-08-06.


**Verbs that look like one family but belong to another** per the canon:
`select`/`toggle`/`acknowledge` are `commit` (they fix state); `edit` is
expressed as `shift.enter-mode` (it changes the regime — there is no `edit`
verb in handle).

Defined in [`verbs.ts:SEMA_VERBS`](../../src/uix/sema/verbs.ts).

### Naming shapes

A `morfo.events[].name` takes ONE canonical shape — the name declares the
family, and `validateMorfo` rejects any that does not:

```ts
// {family}-{verb}[-{nuance}] — head is the family, then the canonical verb,
// then the nuance when the component needs one.
'emerge-dismiss'; // family=emerge, verb=dismiss
'emerge-dismiss-outside'; // + nuance (outside)
'commit-toggle'; // family=commit, verb=toggle
'commit-save'; // family=commit, verb=save
```

`validateEventName(name)` parses a name — it still accepts the retired
bare-verb shape — and returns `{ family, verb, variant, matchesCanonical }`.
Used by
`scripts/morfo-vocabulary-check.ts` (`npm run morfo:vocabulary`), which
hard-fails when a morfo's declared `semantic.verb` is not in its family's
canon, and soft-warns when the event NAME doesn't fit
`{family}-{verb}[-{nuance}]` even though the verb is canonical. A temporary
allowlist in the script covers deliberate divergences.

```ts
import { validateEventName } from '$uix/sema';

validateEventName('commit-toggle');
// { name: 'commit-toggle', head: 'commit', matchesCanonical: true,
//   variant: 'toggle', family: 'commit', verb: 'toggle' }

validateEventName('dismiss-outside');
// { name: 'dismiss-outside', head: 'dismiss', matchesCanonical: true,
//   variant: 'outside', family: 'emerge', verb: 'dismiss' }

validateEventName('frob-glob');
// { name: 'frob-glob', head: 'frob', matchesCanonical: false,
//   variant: 'glob', family: undefined, verb: undefined }
```

## Relationship with Morfo and Soma

- `Morfo` declares the component's semantic events in `morfo.events`
- `Provider` decides when they occur and calls `semantic.emit(...)`
- `SomaRuntime` orchestrates the `prewrite -> emit -> handler -> effects`
  sequence
- `Sema` contributes the vocabulary, normalization and domain validation, and
  publishes the occurrences

## Dependencies

- `EngineSemantic` writes no attributes directly. It orchestrates channel
  hooks (`prepare`, `handle`, `cleanup`) without knowing the DOM attrs.
- Each channel manages its own modality:
  - `VisualChannel.prepare()` projects `data-event-*` and then holds the
    perceptual window.
  - `DomSignalProjector` is the DOM writer used by the visual channel; it
    writes through the `ActiveDom` received from `ActiveUix`.
  - Other channels (sound, haptic) access their respective APIs
    (`AudioContext`, `navigator.vibrate`, etc.).
- In normal use, `ActiveUix` injects the `ActiveDom` into `EngineSemantic`.
  Using Sema directly outside `ActiveUix` must pass an explicit
  `dom/projector` or choose a documented degradation.

## Architecture rule

`Morfo` authorizes the component's semantics.

`Sema` defines the canonical vocabulary and dispatches signals to channels.

`Provider` decides when to emit.

Each `Channel` materializes the signal in its modality.

See [`architecture/overview.md`](./overview.md) §2.bis for the cross-layer
view. The original Spanish API conventions (single-event for instantaneous
operations, intent ↔ visual token resolution, sound prepare-time priming)
survive as a historical seed in
[`guia-semantica-historica.md`](../decisions/guia-semantica-historica.md);
the ruling vocabulary is [`CANON.md`](../CANON.md).