@adia-ai/gen-ui 0.8.62 → 0.8.64

package/CHANGELOG.md

@@ -1,5 +1,41 @@
 # Changelog — @adia-ai/gen-ui
 
+## [0.8.64] — 2026-09-13
+
+### Fixed
+
+- fix: generateUIStream() now wires currentCanvas through to buildCanvasDiffPrompt() the same way generatePro() does, instead of discarding it, so a thinking-mode iteration turn diffs against the prior canvas instead of regenerating fresh every time (ticket 10032) (local ticket 10032)
+- fix: generateUIStream() now merges its diff-shaped LLM response against the prior canvas via mergeCanvasDiff before repair, instead of treating the diff as the complete canvas, so components the LLM left out of a partial-diff response above the 300-component threshold are carried forward instead of silently dropped, matching generatePro's mergeCanvasDiffDetailed behavior (ticket 10050) (local ticket 10050)
+
+### Maintenance
+- **`compose/` touched in this release window** (2 file(s), e.g. `core/generator.js`): carried by the entries above.
+- **`corpus/` touched in this release window** (1 file(s), e.g. `chunks/_source-hashes.json`): carried by the entries above.
+
+## [0.8.63] — 2026-09-12
+
+### Changed
+
+- chore: BREAKING, remove the deprecated `./compose/evals` export alias (ticket 4299, ADR-0098 retirement path) (#4386)
+    The one-release deprecation window LLD-3751 opened for the `@adia-ai/gen-ui/compose/evals`
+    subpath alias (introduced by commit `76789d172`, first carried by `0.8.61`, after `0.8.60`
+    shipped with the original non-aliased path) has closed: `0.8.62`, the release after `0.8.61`,
+    is published. The `./compose/evals` export entry is removed from the engine
+    `package.json`'s `exports` map outright, no successor alias; `@adia-ai/gen-ui/evals` is the
+    sole subpath. This completes the MIGRATION GUIDE lifecycle LLD-3751 started rather than
+    opening a new one: `.claude/docs/MIGRATION GUIDE.md` gains a "Shipped" entry documenting the
+    closure (the original alias-introduction entry stays as historical record, matching this
+    repo's own admin-page/admin-scroll ADR-0098 removal precedent), and its generated
+    `packages/web-components/MIGRATION.md` copy is resynced. No in-repo consumer still imported
+    the old subpath.
+
+### Fixed
+
+- fix: core/generator.js's streaming path now substitutes fallbackMessage() when validateSchema flags isRawText, matching generate-pro.js/generate-thinking.js, so a raw-text leak surviving the repair loop on that path no longer reaches the canvas verbatim (ticket 10028) (local ticket 10028)
+- fix: every producer strategy (zettel, chunk-zettel, free-form-composer, monolithic) now imports one shared `SURFACE_ID` constant instead of three disagreeing literals (`'default'`, `'main'`, or none at all), closing a hole where a zero-named-surface or mismatched-surfaceId doc fell outside a2ui-root's single-surface inference and tore down its own throwaway surface, taking `<a2ui-root>` itself out of the DOM (ticket 10037) (local ticket 10037)
+
+### Maintenance
+- **`corpus/` touched in this release window** (2 file(s), e.g. `chunks/_output-hashes.json`): carried by the entries above.
+
 ## [0.8.62] — 2026-09-10
 
 ### Changed

package/compose/core/generator.js

@@ -38,6 +38,7 @@ import {
   buildRepairPrompt,
   generateSuggestions,
   foldInAntiPatterns,
+  fallbackMessage,
 } from '../strategies/monolithic/_shared.js';
 import { validateAndRepair } from '../shared/validate-and-repair.js';
 import { generateInstant }  from '../strategies/monolithic/generate-instant.js';
@@ -269,11 +270,19 @@ export async function generateUI({ intent, engine: engineName = 'monolithic', mo
  * @yields {object}
  */
 export async function* generateUIStream({ intent, executionId, llmAdapter, model, search, currentCanvas, context, signal }) {
-  // currentCanvas accepted for API parity with generateUI. The streaming
-  // path's downstream consumers don't yet read it, but parity prevents the
-  // client from having to branch its call shape per mode. Tier-2 will
-  // wire it through the same way generatePro does.
-  void currentCanvas;
+  // ticket 10032: normalize the client-owned canvas payload into a flat
+  // prior-components array, the same shape collapse generateUI() does
+  // (above) before dispatching to generatePro. Pre-fix this path discarded
+  // `currentCanvas` entirely (`void currentCanvas;`), so thinking-mode
+  // iteration never diffed against the prior canvas at all.
+  let priorComponentsFromPayload = null;
+  if (currentCanvas) {
+    if (Array.isArray(currentCanvas.components)) {
+      priorComponentsFromPayload = currentCanvas.components;
+    } else if (Array.isArray(currentCanvas.messages)) {
+      priorComponentsFromPayload = currentCanvas.messages.flatMap((m) => m?.components || []);
+    }
+  }
   // ── Intent gate: reject non-UI prompts before entering the pipeline ──
   // Same iteration bypass as generateUI() — modifiers don't need to clear the
   // conversational gate when there's a prior turn to refine.
@@ -479,7 +488,19 @@ export async function* generateUIStream({ intent, executionId, llmAdapter, model
 
   // ── Stage 4: Generate via LLM stream ──
   const systemPrompt = await buildSystemPrompt(catalogContext, patterns, researchContext, intent, null, context);
-  const chatMessages = buildChatMessages(intent, previousExecId || executionId);
+  // ticket 10032: when the client sent a canvas payload, diff against it
+  // (same helper generatePro uses) instead of the legacy store-lookup path
+  // buildChatMessages() takes: the payload is the source of truth (it may
+  // be running stateless, across a server restart, or in a different tab).
+  const hasPriorCanvasFromPayload = Array.isArray(priorComponentsFromPayload) && priorComponentsFromPayload.length > 0;
+  const chatMessages = hasPriorCanvasFromPayload
+    ? [{
+        role: 'user',
+        content: buildCanvasDiffPrompt(intent, priorComponentsFromPayload, {
+          originalIntent: store.getOriginalIntent(previousExecId || executionId) || null,
+        }),
+      }]
+    : buildChatMessages(intent, previousExecId || executionId);
 
   yield { type: 'status', stage: 'generate', message: 'Streaming from LLM...' };
 
@@ -516,6 +537,23 @@ export async function* generateUIStream({ intent, executionId, llmAdapter, model
 
   // ── Stage 5: Parse final + validate ──
   let messages = parseA2UIResponse(fullText, { executionId, mode: 'stream', intent });
+
+  // ticket 10050: above the 300-component threshold, buildCanvasDiffPrompt
+  // (monolithic/_shared.js) switches to asking for a sparse diff instead of
+  // the full canvas, the same threshold generatePro uses. Unlike generatePro
+  // (generate-pro.js's own mergeCanvasDiffDetailed call), this path never
+  // reconstituted the diff against the prior canvas, so components the LLM
+  // left out of the diff were silently dropped instead of carried forward.
+  // Merge here, before repair, so validation/storage below see the same
+  // full canvas generatePro would produce.
+  if (hasPriorCanvasFromPayload && messages && messages.length > 0) {
+    for (const msg of messages) {
+      if (msg.type === 'updateComponents' && msg.components) {
+        msg.components = mergeCanvasDiff(priorComponentsFromPayload, msg.components);
+      }
+    }
+  }
+
   engine.submitStage(executionId, 'generate', {
     messages, source: 'llm-stream', confidence: 0.8,
   });
@@ -551,6 +589,17 @@ export async function* generateUIStream({ intent, executionId, llmAdapter, model
   });
   messages = repairResult.messages;
   let validation = repairResult.validation;
+  // ticket 10028: this streaming path parses raw LLM text via
+  // parseA2UIResponse the same way generate-pro.js/generate-thinking.js do,
+  // so it is exposed to the same raw-text-leak surface `isRawText` flags,
+  // but unlike those two engines, it never substituted fallbackMessage(),
+  // so a leak surviving the repair loop here would have reached the wire
+  // verbatim instead of the same honest failure card every other engine
+  // shows.
+  if (validation.isRawText) {
+    messages = fallbackMessage('Generation produced unstructured raw text instead of a UI', { executionId, intent, mode: 'stream' });
+    validation = validateSchema(messages, { intent, context });
+  }
   if (repairResult.attempts > 0) {
     yield { type: 'status', stage: 'validate', message: `Score ${validation.score}/100 — repaired in ${repairResult.attempts} attempt(s)` };
   }

package/compose/strategies/_shared/surface-id.js

@@ -0,0 +1,19 @@
+// gh#10037: the ONE canonical surfaceId every gen-ui producer strategy must
+// target for its single-surface output. Before this module existed, three
+// different literals were in play across the roster:
+//   - monolithic (generate-instant.js, _shared.js): 'default'
+//   - zettel's chunk-refiner.js (the chunk-zettel iteration path): 'main'
+//   - free-form-composer/transpile.js: 'main'
+//   - zettel's generator-adapter.js (pre-fix): no surfaceId at all
+//
+// A doc's messages must agree on which surface they target across turns:
+// handleAuto.js reads the PRIOR turn's own surfaceId back off
+// state.allMessages to decide whether to bracket a regenerate
+// (beginSurfaceUpdate) or fall back to a full replaceDoc. A mismatched or
+// missing surfaceId between two turns' engines defeats that entirely: the
+// new doc never declares the surface the prior turn created, so
+// a2ui-root.js's replaceDoc() tears the prior surface down as undeclared
+// (packages/gen-ui/a2ui/renderer.js's #deleteSurface). One shared constant,
+// imported everywhere a producer strategy builds an A2UI message, closes
+// that gap for every present and future tier at once.
+export const SURFACE_ID = 'default';

package/compose/strategies/free-form-composer/transpile.js

@@ -53,6 +53,8 @@
  * - Structural substitutions (swap Card → Section, etc.).
  */
 
+import { SURFACE_ID } from '../_shared/surface-id.js';
+
 const VALID_LAYOUTS = new Set(['column', 'row', 'grid']);
 
 /**
@@ -190,7 +192,7 @@ export function transpilePlan(plan, chunkLookup) {
   return {
     messages: [{
       type: 'updateComponents',
-      surfaceId: 'main',
+      surfaceId: SURFACE_ID,
       components: [root, ...flatComponents],
     }],
     warnings,

package/compose/strategies/monolithic/_shared.js

@@ -16,6 +16,7 @@ import { composeSubtasks } from '../../../retrieval/intent/decomposer.js';
 import { getWiringCatalog } from '../../../retrieval/wiring-catalog.js';
 import { getComponentData } from '../../../retrieval/component-catalog.js';
 import { assessIterationFidelity } from '../../shared/iteration-fidelity.js';
+import { SURFACE_ID } from '../_shared/surface-id.js';
 
 // ── v0.9 .a2ui.json sidecar → prompt-catalog adapter ─────────────
 //
@@ -1217,14 +1218,14 @@ function wrapAsMessages(parsed) {
     if (parsed.length > 0 && validTypes.has(parsed[0].type)) return parsed;
     // Bare components array
     if (parsed.length > 0 && parsed[0].id && parsed[0].component) {
-      return [{ type: 'updateComponents', surfaceId: 'default', components: parsed }];
+      return [{ type: 'updateComponents', surfaceId: SURFACE_ID, components: parsed }];
     }
     return parsed;
   }
   if (parsed && typeof parsed === 'object' && parsed.type === 'updateComponents') return [parsed];
   if (parsed && typeof parsed === 'object' && parsed.type === 'wireComponents') return [parsed];
   if (parsed && typeof parsed === 'object' && parsed.id && parsed.component) {
-    return [{ type: 'updateComponents', surfaceId: 'default', components: [parsed] }];
+    return [{ type: 'updateComponents', surfaceId: SURFACE_ID, components: [parsed] }];
   }
   return null;
 }
@@ -1531,7 +1532,7 @@ export function fallbackMessage(reason, opts = {}) {
 
   const updateMsg = {
     type: 'updateComponents',
-    surfaceId: 'default',
+    surfaceId: SURFACE_ID,
     // Marker so validators / scorers / consumers can detect that this is a
     // generation-failure surface (NOT a successful UI). Without this, the
     // validator was scoring fallbacks ~89/100 because they're structurally
@@ -1558,7 +1559,7 @@ export function fallbackMessage(reason, opts = {}) {
 
   const wireMsg = {
     type: 'wireComponents',
-    surfaceId: 'default',
+    surfaceId: SURFACE_ID,
     actions,
   };
 

package/compose/strategies/monolithic/generate-instant.js

@@ -12,6 +12,7 @@ import { assessClarity } from '../../../retrieval/intent/clarity.js';
 import { feedbackStore } from '../../../retrieval/feedback/feedback-store.js';
 import { store, engine } from '../../core/state.js';
 import { isRecording } from '../../../retrieval/feedback/dialog-recorder.js';
+import { SURFACE_ID } from '../_shared/surface-id.js';
 import {
   generateSuggestions, foldInAntiPatterns } from './_shared.js';
 
@@ -156,7 +157,7 @@ export async function generateInstant({ intent, executionId, storeId, analysis,
   const messages = [
     {
       type: 'updateComponents',
-      surfaceId: 'default',
+      surfaceId: SURFACE_ID,
       components,
       ...(isPlaceholderStub ? {
         _fallback: true,
@@ -169,7 +170,7 @@ export async function generateInstant({ intent, executionId, storeId, analysis,
   if (bestMatch?.wiring) {
     messages.push({
       type: 'wireComponents',
-      surfaceId: 'default',
+      surfaceId: SURFACE_ID,
       ...(bestMatch.wiring.data_sources ? { data: { sources: bestMatch.wiring.data_sources } } : {}),
       ...(bestMatch.wiring.controllers ? { state: { controllers: bestMatch.wiring.controllers } } : {}),
       ...(bestMatch.wiring.actions ? { actions: bestMatch.wiring.actions } : {}),

package/compose/strategies/registry.js

@@ -18,6 +18,11 @@
  *   { executionId, messages, validation, suggestions?, strategy?, ... }
  */
 
+// Pure constant, no node:fs/node:path deps, safe to static-import even
+// though the strategies that use it (zettel, chunk-zettel) are lazy-loaded
+// below for browser-safety reasons that don't apply to this module.
+import { SURFACE_ID } from './_shared/surface-id.js';
+
 // Zettel is lazy-loaded — it transitively imports node:fs / node:path / node:url
 // (strategies/zettel/composition-library.js), which Vite externalizes in the browser.
 // Static-importing it here would break browser loads of core/generator.js.
@@ -324,10 +329,15 @@ async function generateChunkZettelAdapter(ctx) {
   // use it directly; the pre-fix fallback (a single node carrying an
   // UNREGISTERED `component: 'article'` type) only fires if transpilation
   // itself somehow produced nothing, which should not happen in practice.
+  // gh#10037: this adapter's two branches used to omit `surfaceId`
+  // entirely, the same defect fixed in zettel/generator-adapter.js: a
+  // doc naming zero surfaces falls outside a2ui-root.js's single-surface
+  // inference and tears down its own throwaway surface on regenerate,
+  // taking the <a2ui-root> host element with it.
   const messages = result.template
-    ? [{ type: 'updateComponents', components: result.template }]
+    ? [{ type: 'updateComponents', surfaceId: SURFACE_ID, components: result.template }]
     : result.html
-      ? [{ type: 'updateComponents', components: [{ id: 'chunk-root', component: 'article', html: result.html }] }]
+      ? [{ type: 'updateComponents', surfaceId: SURFACE_ID, components: [{ id: 'chunk-root', component: 'article', html: result.html }] }]
       : [];
 
   // Full schema/catalog/anti-pattern conformance — the same final pass

package/compose/strategies/zettel/chunk-refiner.js

@@ -27,10 +27,10 @@ import {
   searchChunksAsync,
 } from '../../../corpus/scripts/chunk-library.js';
 import { composeFromPlan } from './chunk-composer.js';
+import { SURFACE_ID } from '../_shared/surface-id.js';
 
 const DEFAULT_MAX_ATTEMPTS = 2;
 const PRE_SEARCH_LIMIT = 30;
-const SURFACE_ID = 'main';
 
 const VALID_OP_TYPES = new Set(['rebindSlot', 'appendToSlot', 'removeFromSlot', 'replacePage']);
 

package/compose/strategies/zettel/generator-adapter.js

@@ -1,11 +1,11 @@
 /**
- * Generator adapter for zettel MCP — conforms to the eval harness contract:
+ * Generator adapter for zettel MCP: conforms to the eval harness contract:
  *
  *   generate({ intent, mode, llmAdapter, sessionId, signal }) -> { messages, validation, strategy, ...extra }
  *
  * Two reasoning layers (post-§195, v0.5.6):
- *   1. Retrieval (always available) — keyword-rank the corpus, resolve top composition.
- *   2. LLM synthesis (when llmAdapter provided AND retrieval weak) — bridge to
+ *   1. Retrieval (always available) , keyword-rank the corpus, resolve top composition.
+ *   2. LLM synthesis (when llmAdapter provided AND retrieval weak) , bridge to
  *      chunk-zettel's `composeFromIntent` (the §37 successor codepath) which
  *      produces a single-shot html composition from the chunk corpus.
  *
@@ -13,19 +13,19 @@
  * are recorded for analytics + drift tracking, but turn≥2 takes the same code
  * path as turn 1 (fresh retrieval → fresh synthesis). Full history-aware
  * iteration (multi-turn refinement modifying an existing canvas) lives in the
- * `chunk-zettel` engine via `chunk-refiner.js::refineFromIntent` — wire it via
+ * `chunk-zettel` engine via `chunk-refiner.js::refineFromIntent` , wire it via
  * `engine: 'chunk-zettel'` rather than `engine: 'zettel'`.
  *
  * The prior fragment-graph iteration codepath (`synthesizeComposition` with
  * `historySummary`) was retired in §37 (2026-05-12) when fragments retired.
- * §195 (v0.5.6) cleaned up the dangling caller — turn≥2 no longer throws +
+ * §195 (v0.5.6) cleaned up the dangling caller , turn≥2 no longer throws +
  * auto-fires a `iteration-synthesis-failure` issue on every multi-turn turn.
  *
  * Strategy labels in the return:
- *   - composition-match          — fresh retrieval, strong match, emitted verbatim
- *   - composition-synthesized    — chunk-zettel single-shot synthesis
- *   - fragment-candidates        — retrieval weak + no LLM, returning atoms only
- *   - synthesis-failed           — chunk-zettel tried and failed validation
+ *   - composition-match          , fresh retrieval, strong match, emitted verbatim
+ *   - composition-synthesized    , chunk-zettel single-shot synthesis
+ *   - fragment-candidates        , retrieval weak + no LLM, returning atoms only
+ *   - synthesis-failed           , chunk-zettel tried and failed validation
  */
 import {
   getComposition,
@@ -38,6 +38,7 @@ import {
   getTurns,
 } from './session-store.js';
 import { validateSchema } from '@adia-ai/a2ui/validate';
+import { SURFACE_ID } from '../_shared/surface-id.js';
 
 // Lazy-load the chunk-synthesizer bridge. It imports chunk-corpus data
 // via composition-library's dual-mode loader, which is already top-level
@@ -51,13 +52,13 @@ async function bridgeToChunkSynthesis({ intent, llmAdapter, signal }) {
   // to the synthesis-branch's prior `{ messages, template, synthesis }`
   // contract. TKT-0009: prefer the real component array (`result.template`)
   // over the pre-fix fallback of wrapping raw HTML in a single node with an
-  // UNREGISTERED `component: 'article'` type — that fallback made this
+  // UNREGISTERED `component: 'article'` type , that fallback made this
   // caller's validate-and-repair pass below structurally unable to ever
   // report a clean pass.
   const messages = result.template
-    ? [{ type: 'updateComponents', components: result.template }]
+    ? [{ type: 'updateComponents', surfaceId: SURFACE_ID, components: result.template }]
     : result.html
-      ? [{ type: 'updateComponents', components: [{ id: 'chunk-root', component: 'article', html: result.html }] }]
+      ? [{ type: 'updateComponents', surfaceId: SURFACE_ID, components: [{ id: 'chunk-root', component: 'article', html: result.html }] }]
       : [];
   return {
     messages,
@@ -74,7 +75,7 @@ async function bridgeToChunkSynthesis({ intent, llmAdapter, signal }) {
 // Composition library auto-loads at module import time via top-level `await
 // loadAll()` in composition-library.js (§72 dual-mode loader, commit
 // `76dbcff2`). The previous `ensureBooted()` here called `loadAll()` WITHOUT
-// `await`, which after §72 made it async — clearing the compositions map +
+// `await`, which after §72 made it async , clearing the compositions map +
 // racing the synchronous `searchAll` below. Symptom (caught in §87):
 // eval:diff zettel coverage dropped 5% → 1% because the race-winner pattern
 // emitted composition-match only for the intent whose `searchAll` happened
@@ -83,7 +84,7 @@ async function bridgeToChunkSynthesis({ intent, llmAdapter, signal }) {
 // Fix: rely entirely on the top-level await. Tests that want a forced reload
 // can call `loadAll()` directly with `await` (it's idempotent).
 
-// Retrieval score threshold — above this we trust the match and emit verbatim;
+// Retrieval score threshold , above this we trust the match and emit verbatim;
 // below, fall through to LLM synthesis (creative composition from fragments).
 // Calibrated on the 100-intent held-out set:
 //   strong matches: login-form=24 (exact), signup-form=27, chart-dashboard=48, pricing-tiers=54
@@ -95,9 +96,23 @@ async function bridgeToChunkSynthesis({ intent, llmAdapter, signal }) {
 // threshold maximized verbatim-cache hits at the cost of repetitive output.
 const STRONG_MATCH_THRESHOLD = 40;
 
+// gh#10037: every emitted `updateComponents` message must name a `surfaceId`
+// (the shared `SURFACE_ID` imported above, so every tier agrees across
+// turns). Without it, a2ui-root.js's replaceDoc() has zero named
+// surfaceIds to infer a sole surface from (its inference only covers the
+// single-named-surface case, not zero), so the message ends up applied
+// against surfaceId `undefined`, renderer.js synthesizes a throwaway
+// surface rooted at the <a2ui-root> element itself for that write, and
+// replaceDoc's own tear-down sweep then deletes it as undeclared, which
+// removes a2ui-root from the DOM (renderer.js's `deleteSurface` calls
+// `surface.root.remove()`). This is the same class of bug as gh#2426's
+// fix in a2ui-root.js, reintroduced here on the zettel producer side of
+// the contract instead of the consumer side.
+
 function toUpdateComponentsMessages(template) {
   return [{
     type: 'updateComponents',
+    surfaceId: SURFACE_ID,
     components: template.map((n) => {
       const {
         id, component, children,
@@ -114,12 +129,12 @@ export async function generateZettel({ intent, mode = 'instant', llmAdapter = nu
   // ── Session iteration note (post-§195, v0.5.6) ──
   // Fragment-graph history-aware iteration (`synthesizeComposition` with
   // `historySummary`) was retired in §37 when fragments retired. Turn≥2 now
-  // takes the same path as turn 1 — fresh retrieval → fresh chunk-synthesis.
+  // takes the same path as turn 1 , fresh retrieval → fresh chunk-synthesis.
   // For true history-aware iteration (modify-an-existing-canvas), wire the
   // request via `engine: 'chunk-zettel'` which uses `chunk-refiner.js`.
   //
   // Prior turns are still recorded (analytics + drift tracking via `getDrift`)
-  // but the iteration BRANCH is gone — no more spurious
+  // but the iteration BRANCH is gone , no more spurious
   // `iteration-synthesis-failure` auto-fires on every multi-turn request.
   const priorTurns = sessionId ? getTurns(sessionId) : [];
 
@@ -127,10 +142,10 @@ export async function generateZettel({ intent, mode = 'instant', llmAdapter = nu
   const composition = hits.find((h) => h.type === 'composition');
   const strongMatch = composition && composition.score >= STRONG_MATCH_THRESHOLD;
 
-  // ── Strong retrieval match: emit verbatim (turn 1 only — iteration branch above handles repeats) ──
+  // ── Strong retrieval match: emit verbatim (turn 1 only , iteration branch above handles repeats) ──
   if (strongMatch) {
     // On a static deploy the composition was registered from `_index.json`
-    // for SCORING only — its `template` is fetched lazily here, right before
+    // for SCORING only , its `template` is fetched lazily here, right before
     // we resolve it. No-op in Node / Vite (template already populated). This
     // keeps the static page-load to a single index request; only the
     // actually-emitted match pays for its body. 2026-06-10.
@@ -142,7 +157,7 @@ export async function generateZettel({ intent, mode = 'instant', llmAdapter = nu
     // is looser (it can't see templates in `_index.json`), so a handful of
     // body-less partials (auth-card-header, onb-step-*, …) COULD surface as a
     // strong match on a static deploy. If hydration produced no usable
-    // template, treat it as a non-match and fall through — keeping the
+    // template, treat it as a non-match and fall through , keeping the
     // static result identical to what Node would emit (synthesis / atoms),
     // never an empty verbatim surface.
     if (!Array.isArray(template) || template.length === 0) {
@@ -179,7 +194,7 @@ export async function generateZettel({ intent, mode = 'instant', llmAdapter = nu
   if (llmAdapter && mode !== 'instant-only') {
     try {
       const synth = await bridgeToChunkSynthesis({ intent, llmAdapter, signal });
-      // Full schema/catalog/anti-pattern conformance (validate-only — see
+      // Full schema/catalog/anti-pattern conformance (validate-only , see
       // validate-and-repair.js; TKT-0009 made this meaningful for the first
       // time by giving synth.messages a real component graph instead of an
       // opaque HTML wrapper).

package/corpus/chunks/_index.json

@@ -1,6 +1,6 @@
 {
   "schema": "genui-chunk-index@2",
-  "captured_at": "2026-09-09T16:07:52.181Z",
+  "captured_at": "2026-09-13T22:07:39.419Z",
   "total_instances": 746,
   "unique_names": 591,
   "by_kind": {

package/corpus/chunks/_output-hashes.json

@@ -1,7 +1,7 @@
 {
   "schema": "genui-chunk-output-hashes@1",
   "algo": "sha256",
-  "capturedAt": "2026-09-10T23:55:57.033Z",
+  "capturedAt": "2026-09-13T22:07:39.419Z",
   "aggregateDigest": "9eaca6f62e9dc45f78ebb2d22889b53071405a51a9bcd4410c737f076e57e8d3",
   "harvestEnv": {
     "nodeMajor": "24",

package/corpus/chunks/_source-hashes.json

@@ -283,7 +283,7 @@
     "catalog/ui-patterns/v050-marketing-blocks/v050-marketing-blocks.html": "7833b21ee84643978cd3ca0e6ff20caa20ec8cbc9a55e371aa295a532f8ed48a",
     "catalog/ui-patterns/v050-nav-blocks/v050-nav-blocks.contents.html": "86f492db19b8aafc2bf25620f52377cd7a6cb92cfa2d0d3bc9dd4c5ade3e88c6",
     "catalog/ui-patterns/v050-nav-blocks/v050-nav-blocks.html": "196862f06f495547d85f542e6a0eee152ab40a40ca969c93b0cbd0733de5a5e2",
-    "packages/gen-ui/a2ui/catalog/tier-index.json": "e373144f8bc693ed6cba3d910352b6513a8bd948e7383331bceb5d151c74d815",
+    "packages/gen-ui/a2ui/catalog/tier-index.json": "850ec77e8b2f33ea9756e22e27976cd1011ccce618d92b7e9ecf0461501302bd",
     "packages/web-components/components/accordion/accordion.examples.html": "0218a098bc9db38f6ad5a73b7fbbd8755882d92f47790488928384b79b2e7868",
     "packages/web-components/components/action-list/action-list.examples.html": "abcb4897fc8143a4c32b15bf82eb3340304cfa22d8498198328c1570102d7bc2",
     "packages/web-components/components/adia-mark/adia-mark.examples.html": "2124ad37660d157e7a85fdc57eb7725ac02d629a76a47a58ff53703ed5b06238",
@@ -611,7 +611,7 @@
     "site/pages/getting-started/usage.html": "23f477ad6e505876e36d6c774cd49acc3bb483f7e4872ea8d6d5261fdab02b4f",
     "site/pages/guides/agent-skills.html": "647864bb9c6e43ed123c95b48201f3e55a5bb26e8a77cd0ee65ad361df47bd97",
     "site/pages/guides/bundle-loading.html": "dc08497d9621c090e1581289ebf9a706ae05421a55dc9f20841c107969b3b829",
-    "site/pages/guides/bundle-sizes.html": "1f649468e9b5be807bec4e72d2898a6f71b779346b483b802ab73eb4811d16a8",
+    "site/pages/guides/bundle-sizes.html": "b8b067e5e3c89bfb6ea809a971a5a7d13816b27f550a8a9e391dafa2d4421551",
     "site/pages/guides/creating-components.html": "f34020a12d91a0239b6303e32b3c39ecc47d7a96ccfa7ff10fc12e313065a5d6",
     "site/pages/guides/css-architecture.html": "42b4258cefc5f98d3b1297362aa090fe1310b92032871fc3144b33a42f03f331",
     "site/pages/guides/data-flow.html": "c1965674271df66cfebf95eb012048996fc8c844179fcb1816d49aae98621219",

package/corpus/chunks/playground-chat.json

@@ -23,7 +23,7 @@
       "agent"
     ]
   },
-  "captured_at": "2026-09-07T13:06:17.368Z",
+  "captured_at": "2026-09-13T22:07:39.419Z",
   "template": [
     {
       "id": "text",

package/corpus/manifest.json

@@ -1,5 +1,5 @@
 {
-  "timestamp": "2026-09-09T16:07:52.181Z",
+  "timestamp": "2026-09-13T22:07:39.419Z",
   "pages": 221,
   "chunks": 591,
   "skipped": 1680

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@adia-ai/gen-ui",
-  "version": "0.8.62",
+  "version": "0.8.64",
   "description": "AdiaUI generative-UI system — compose strategies, retrieval, the harvested training corpus, and catalog-aware + LLM-judge validation. Emits A2UI protocol messages; pairs with @adia-ai/a2ui (the protocol runtime). Folded from @adia-ai/a2ui-{compose,retrieval,corpus} + the catalog/semantic halves of the former @adia-ai/a2ui-validator (ADR-0048).",
   "type": "module",
   "main": "./compose/index.js",
@@ -45,11 +45,6 @@
       "import": "./compose/transpiler/transpiler.js",
       "default": "./compose/transpiler/transpiler.js"
     },
-    "./compose/evals": {
-      "types": "./evals/harness.d.ts",
-      "import": "./evals/harness.mjs",
-      "default": "./evals/harness.mjs"
-    },
     "./evals": {
       "types": "./evals/harness.d.ts",
       "import": "./evals/harness.mjs",