Auditoría clean-room de todo ActiveUIX (excluido `web/`), componente a
componente. Lo que sale de aquí no es una lista de bugs: es un patrón.
El framework validaba que lo escrito fuese VÁLIDO, no que lo declarado
se CUMPLIESE — y sus guards fallaban ABIERTOS.
## El colapso de las uniones de props (95 → 0)
Un `Props` de eidos es `{ …props propias… } & <atributos nativos>`.
Cuando el elemento declara un atributo homónimo, la intersección funde
ambos y una unión estrecha contra el `string` nativo COLAPSA a `string`.
Causa: `Without<T, U> = Omit<T, keyof U>` invocado como `Without<T, {}>`
— `Omit<T, never>`, un no-op — 433 veces en soma; sólo 3 con argumento
real. Invisible para `svelte-check`: ensanchar un tipo no es un error,
es una garantía perdida.
Medido: 95 props en 72 componentes. `<Avatar color="nonsense">`
compilaba. `ComboboxInput.size` chocaba con el `<input size>` numérico
y era inusable. Migrado con codemod sobre AST (nunca regex) a
`Own & Omit<Nativos, keyof Own>`: 92 tipos en 73 ficheros + carousel a
mano. `check` no se movió.
Garantía nueva: `eidos/prop-surface.test.ts` (PROP-1) compara los
literales de la anotación del autor contra los de la propiedad pública.
Verificado que falla reintroduciendo el defecto.
## Los cuatro guards que fallaban abiertos
- `translations:check` crasheaba en CADA ejecución de su historia — un
stripper de comentarios borraba `//` dentro de strings. Sustituido por
import dinámico. Al arrancar destapó 8 slots `texts` sin traducción.
- `soma-attr-audit` agotaba el timeout de 5 s: sin veredicto, verde por
omisión.
- `component-audit` D-7.4 hacía `continue` mudo cuando el tipo no
resolvía. Ahora resuelve con el checker de TypeScript
(`scripts/prop-unions.ts`): puntos ciegos de 124 → 3.
- `component-audit` R-1.1: el regex casaba `[data-motion='reduce']` y
daba PASS por el motivo equivocado.
Regla adoptada: un guard que no puede evaluar TIENE que decirlo. El
informe lleva ahora bloque «Not verified» y recuento en el resumen.
## D-1 · tooltip y D-2 · card, cableados
`tooltip` declaraba 3 eventos `emerge` que nadie emitía. Ahora emiten;
`present` pasa a `sequence: 'post'` — con `'pre'` el hold de ~240 ms
gateaba el montaje del propio overlay.
`card` declaraba `commit-select` sin emisor posible (scope sin soma).
Puente headless en `soma/components/card/` con la forma ya establecida
por `menu-dial` / `onion-menu`: eidos posee estado y render, soma posee
sólo el `SomaRuntime` que emite.
## Documentación: 22 mentiras corregidas
`docs/` afirmaba guards inexistentes (`NO_MISSING_PROVIDER_TESTS`),
APIs con firma equivocada y un modelo de Motion que el código no
implementa. Corregido en CANON, arquitectura, glosario, theming/motion,
checklist y los README de `motion` / `callout` / `arts/motion`.
## Además
- CardGroup: la descripción se metía en la primera celda del grid.
- Motion: `data-state` siempre estampado, salida real en `leave()`,
token fantasma `--motion-stagger-each-default` eliminado.
- `engine-motion`: `handoffState` Map → WeakMap (fuga por nodo).
- `mockup`: primitivo crudo → token de rol (R-4.6).
- 5 catálogos de traducción que faltaban.
Handoff: `docs/process/CONTINUE-audit-2026-07-29.md`.
Batería: check 0 errores en src · vitest server 3701/3701 ·
docs:check 0/0 · component:audit 161 PASS / 2 NEEDS-WORK.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two streams, split by what the animation touches:
STREAM A — decorative backgrounds → the pack tier
- arts/scene: a consolidated scene runtime ($scene) that owns, once, the
citizenship every ad-hoc background reinvented or skipped (frame loop,
off-view pause, DPR cap, mandatory reduced-motion, WebGL context
loss/restore, scene budget, teardown). SceneDom port (adom satisfies it),
webgl/webgl2/canvas2d drivers + a custom-pipeline extension (vertexShader +
draw + glContext.depth/dprCap) for real geometry (beam, particles, dither,
grid, eter, pixel-blast, hyperspeed). 32 effects as shared resources.
- src/packs/ambient: the first pack — <Ambient effect="…"> mounts a registered
effect; the P contract (P-1..P-6) guarded by scripts/packs-check.ts; colors
are token-aware (P-4). One-way dependency, removable-by-construction.
- resolveToken extended to semantic color slots (--color-{role}-{slot}) so
consumers resolve theme tokens to concrete colors (the P-4 half).
STREAM B — animations over real text → canon
- Six components (count-up + text-{gradient,circular,blur,focus,scramble}):
each a morfo + eidos recipe (where there's styling) + demo. CountUp is a
service component (counts through uix.format.numbers). The five Text* are
passive decoratives. Upgrades over the seeds: SR hardening (real text
visually-hidden + aria-hidden decoration), a11y fix (no fake role=button),
measurement discipline (cached rects via dom.measure, no reflow storm),
reduced-motion, ecosystem citizenship (eidos.dom/timers, no raw platform).
- MorfoElement gains 'p'.
DOCS
- docs/architecture/packs.md (pack tier, admission rule, P contract, Aura
promotion path); docs/decisions/design-text-effects.md (the family design
record) + indexed in decisions.md / README.md; glossary entries
(scene/Ambient/Aura/text effects); scene README custom-pipeline + authoring
bridge; motion-guide content-effects note; strata tables acknowledge packs.
Gates: component:audit 141/0/0 · docs:check 0/0 · scene tests 9/9 · packs:check
0/36 · check 0 own errors. Verified in browser (32 effects mount+compile;
6 text components SSR+hydrate, CountUp re-formats by locale live, TextGradient
resolves token stops to OKLCH via var()).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
src/uix/eidos/THEMING.md (1499 L, Spanish) translated to English as
docs/theming/reference.md, same s1-s38 numbering: the mental model,
s1.bis (theming lives in eidos — the 2-of-3 derivation and the canonical
split), the CSS layers, the 7 token layers, the 9 roles, the size canon
(universal 1:1 + label step-down + container cap + the icon scale), the
naming conventions, the s7/s8/s9/s13/s15/s17/s18 stubs repointed into the
book (canon/tsc, theming/guide, theming/notes, theming/changelog,
theming/motion), runtime overrides + builders, bundle/purge, validation
tooling, s14 motion (the pickup lift), s16 anti-patterns, s19 variants
canon, and the s20-s38 standing-decision stubs. One internal
reconciliation applied: the s35 stub cited the control/compact/dense size
archetypes that s5 declares superseded by the universal 1:1 — aligned.
One stale v1 DEMO_AUTHORING s12.8 citation in s5 replaced by the v2 s6
parity rule. Stub at the old path carries the full s1-s38 map (the most
sN-cited doc in the repo: code comments, CLAUDE.md, RFCs). Corpus links
swept (CANON, docs map, comparison, glossary, eidos chapter, canon/tsc,
theming/guide + notes). docs:check 0 errors (260 docs).
F7.3 complete: docs/canon/ (tsc, recipe-contract) + docs/theming/
(reference, guide, notes, channels, motion, motion-guide, changelog).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
src/uix/eidos/TSC.md (334 L) and RECIPE_CONTRACT.md (131 L), both Spanish,
translated to English as docs/canon/tsc.md + docs/canon/recipe-contract.md
(the E2 visual canon inside the book tree). Thin stubs at the old paths
(code comments and component-audit R-4.x provenance keep resolving);
corpus links swept (docs map E2 rows, authoring anti-copy law,
building-a-component phase 5, comparison, getting-started, glossary,
eidos chapter). docs:check 0 errors.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
src/uix/soma/SOMA_ARCHITECTURE.md (1058 L, Spanish) translated to English
as docs/architecture/soma-architecture.md, same s1-s17 numbering: layer
architecture, design principles, the closed six-piece model + SomaRuntime,
component model (picker composition), runtime parts, the layers inventory,
the Soma class + date/time domain + statics convention, the reactive
system, internal helpers, data-* contracts, IDs, barrels, boundaries,
directory structure, anti-patterns, current shape, stability rule,
checklist. The frozen per-provider test list (dated 2026-05-15, ~140 L)
became the timeless fact: the NO_MISSING_PROVIDER_TESTS guard + the test
tree ARE the coverage inventory (testing-and-tooling aligned). Stub with
the full sN map at the old path; corpus links swept.
F7.2 is complete: docs/architecture/ now holds the whole E1 stratum in
English (7 chapters, ~5.4k lines), with thin stubs + sN maps next to the
code. docs:check 0 errors (251 docs).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
src/uix/eidos/README.md (1020 L, Spanish) translated to English: the
ActiveEidos runtime (config/patch, themes, themeSource, CSS contract,
setCssVariables, persistence envelope, resolveToken), canonical size +
transversal primitives, what eidos consumes from morfo/soma/sema, the
--* token rule, the THEMING sN map, typographic vertebration (two
anchors + alias chain + R-2.7), the 2-of-3 rule, disciplined-option-C
API conventions (7+4 rules, Toast special case), the selector-drift
defense, and the picker patterns P-1..P-5. Two already-decided
reconciliations folded in: themes/fonts.css superseded (font-faces live
in generated/base.css) and the dated 2026-05-21 picker block merged into
the picker-patterns intro — its dead PENDIENTES.md pointer replaced by a
TODO(reconcile) note (norms N-6/N-7 orphaned by d68d2c45), which also
clears one docs:check warn (13 -> 12). Thin stub at the old path keeps
THEMING/TSC/motion pointers next to the code; corpus links swept.
docs:check 0 errors.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
User decision (2026-07-02): the corpus stops being a MAP over dispersed
layer docs — the layer reference MOVES into docs/ with book order, in
English, superseding the kickoff-era 'hybrid structure' decision. The 189
component READMEs stay in-place (E5, linked).
- docs/process/PLAN-docs-book.md: the phase-7 plan — target tree
(architecture/ canon/ theming/ rfcs/ guides/ decisions/), the per-doc
migration pattern (translate -> Write new path -> stub at old path ->
link sweep -> docs:check -> commit per batch), batches F7.1-F7.7 with
volumes, and the known risks (sN citations get a map in the stub; the
legacy RFC renames finally become safe because the stub keeps the old
name alive for provenance citations).
- Pilot migrated end-to-end: docs/architecture/active-uix.md (English,
frontmatter, links repointed) with a thin stub at
src/uix/active-uix/README.md; docs/README map, glossary and
active_architecture s0 repointed. Validates the pattern for F7.2+.
docs:check: 0 errors (244 docs).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Companion to the entry point: an agent or new session can look up the framework's
coined terms (morfo, soma, sema, eidos, archetype, runtime part, Presence,
polymorphic close, hold, cascade, TSC, recipe, variant, role, scaling, …) in one
place, each with a pointer to its authoritative doc. The semantic subset
(family / intent / verb / channel) points at CANON.md instead of restating the
values, so it cannot drift.
Wired into docs/README.md (E0 stratum + reading order).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
dev