@hyperfixi/testing-framework 2.5.1 → 2.6.0

package/src/multilingual/validators/execution-validator.ts

@@ -0,0 +1,510 @@
+/**
+ * R2 — execution smoke validator (CORRECTNESS_RELIABILITY_PLAN.md §8).
+ *
+ * Parse-level ratchets (R0 action-set fidelity, R1 role fidelity) cannot see a
+ * parse that EXECUTES wrongly: the session-5 probe found translations whose
+ * parse looks plausible but whose runtime behavior diverges from the English
+ * reference — a dropped destination role lands the effect on `me` instead of
+ * `#menu`, a mis-built AST throws `Unknown command: into`, a quoted literal
+ * keeps its quotes in the DOM. This validator executes a CURATED subset of
+ * corpus patterns in jsdom (semantic parse → buildAST → runtime.execute →
+ * dispatch the trigger event) and compares the resulting DOM effects against
+ * the English reference's effects.
+ *
+ * Effect signature: a before/after diff of every element's classes, attributes,
+ * inline style, and leaf text. Selectors/classes/attribute names are code (not
+ * translated), so signatures are directly comparable across languages. A
+ * pattern scores 1 when its signature EXACTLY matches the en reference's
+ * (binary — extra effects are as wrong as missing ones). Trapped runtime
+ * errors are recorded for diagnostics but are NOT part of the match: their
+ * attribution depends on unhandled-rejection timing (racy), while the effect
+ * snapshot is synchronous and deterministic. A mis-built AST that damages
+ * behavior diverges in its effects regardless.
+ *
+ * Determinism: a fresh JSDOM document and a fresh Runtime per execution, no
+ * network, no timers beyond a fixed settle window; the probe measured two full
+ * sweeps byte-identical. Patterns whose en reference errors or produces no
+ * effects are excluded (no usable reference — same semantics as R0/R1).
+ */
+
+import { JSDOM } from 'jsdom';
+import { parseSemantic, buildAST } from '@lokascript/semantic';
+import { getAllPatterns, getTranslationsByLanguage } from '@hyperfixi/patterns-reference';
+import type { LanguageCode } from '../types';
+
+/**
+ * The curated execution subset: simple, deterministic, fixture-friendly
+ * single-handler click patterns. Expand only after the loop is proven stable
+ * (an unstable ratchet is worse than none). `set-text-basic` (#id.innerText)
+ * is deliberately absent: jsdom does not implement innerText, so its en
+ * reference write is an inert expando (empty signature) — a harness
+ * limitation, not a runtime gap. `toggle-visibility` and `set-attribute`
+ * joined in wave 2b once @attr emitted canonical attributeAccess nodes.
+ */
+export const EXECUTION_SUBSET: readonly string[] = [
+  'add-class-basic',
+  'add-class-to-other',
+  'remove-class-basic',
+  'remove-class-from-all',
+  'toggle-class-basic',
+  'toggle-class-on-other',
+  'put-content-basic',
+  'show-element',
+  'hide-element',
+  'increment-counter',
+  'decrement-counter',
+  'modal-open',
+  'tabs-basic',
+  'set-text-possessive-dot',
+  'set-inner-html-possessive-dot',
+  'set-style',
+  'closest-ancestor',
+  // Expansion wave 1 (session 8): multi-command click patterns — sync,
+  // network/timer-free sequences of 2–4 commands. Same eligibility bar as the
+  // original subset: the en reference must execute with a non-empty effect
+  // signature in the current runtime. Eleven candidates were probed; the two
+  // then-excluded for unusable EN references (modal-close-button,
+  // make-toast-element) joined in waves 3/3b once their en parses were fixed.
+  'tabs-content',
+  'accordion-exclusive',
+  // Expansion wave 2 (session 9, post en-conditional work): control-flow click
+  // patterns whose en reference now parses (the if/else conditional fold) and
+  // executes (propertyAccess evaluator; matches/exists/is-empty condition
+  // forms). Probed through this validator: each produces a non-empty,
+  // deterministic effect signature. `unless-condition` stays out — `unless` is
+  // deliberately NOT folded (see semantic-parser.tryParseConditionalBlock), so
+  // its flat parse still errors at runtime.
+  'if-condition',
+  'if-matches',
+  'if-exists',
+  'modal-close-backdrop',
+  // Expansion wave 2b (session 9): the @attr family — the semantic value
+  // converter / expression parser now emit canonical attributeAccess nodes for
+  // `@attr`, so these en references execute (set → setAttribute, toggle →
+  // attribute toggle). set-text-basic stays out: its en reference writes
+  // `innerText`, which jsdom does not implement (the write becomes an inert
+  // expando, empty signature) — a harness limitation, not a runtime gap; it
+  // works in real browsers.
+  'set-attribute',
+  'toggle-visibility',
+  'tabs-aria',
+  // Expansion wave 2c (session 9): `halt the event` no longer collapses to a
+  // bare halt (the semantic halt mapper preserves its patient, so the handler
+  // CONTINUES after preventDefault/stopPropagation), and `make a <div.card/>`
+  // executes (element literals carry their markup on `raw`; #container added
+  // to the fixture as the put destination).
+  'halt-propagation',
+  'make-element',
+  // Expansion wave 2d (session 9): `next .dropdown-menu` folds to a positional
+  // call expression (next('.dropdown-menu')) the runtime's positional
+  // expressions evaluate — previously it mangled into `next.dropdown - menu`
+  // and the toggle target evaluated to NaN. `.dropdown-menu` added to the
+  // fixture (appended last; existing snapshot indexes unchanged).
+  'dropdown-toggle',
+  // Expansion wave 3 (session 10): positional-phrase patients/destinations —
+  // the pattern matcher now captures `closest <sel>` and article-led
+  // `the next <sel>` as positional expressions (previously the hide/show
+  // command dropped from the body entirely). The en reference for
+  // modal-close-button now hides the enclosing .modal (PATTERN_SETUP gives
+  // #btn a .modal ancestor, mirroring the real-page structure).
+  'modal-close-button',
+  // Expansion wave 3b (session 10): the at-end-of positional put — en put
+  // patterns for `at end of` / `at start of`, the parseBodyWithClauses
+  // end-noun guard (the `end` in `at end of` is a position noun, not the
+  // block terminator), putMapper reads `manner` (also fixing the latent
+  // before/after→into bug), core PutCommand accepts the multi-word modifier
+  // keys, and contextReference body/document/window resolve. The en
+  // reference appends the toast at end of body: one clean effect line.
+  'make-toast-element',
+  // Expansion wave 4 (the deferred S1 follow-up): the first non-click,
+  // event-reading cell. `on success put event.detail.message into #sr-announce
+  // set @role to "alert" on #sr-announce` exercises (1) a custom event name
+  // carried as an expression role — buildEventHandler now binds it instead of
+  // defaulting to `click`; (2) the `set @attr … on <scope>` plumb landed in S1;
+  // (3) per-cell trigger support (PATTERN_TRIGGER) dispatching a CustomEvent
+  // with a `detail` payload. The en reference writes the announcement text into
+  // #sr-announce and sets role=alert on it: two clean effect lines.
+  'announce-screen-reader',
+  // Expansion wave 5 (session 12): `remove me` — the bare self-removal positional.
+  // Discovery probe (every non-subset, non-network/timer/behavior pattern executed
+  // against this fixture, then all 23 translations checked against the en effect
+  // signature) found ten patterns with a clean non-empty en effect.
+  'remove-element',
+  // Expansion wave 6 (session 13): the six wave-5 "worklist" candidates that, when
+  // RE-GROUNDED against a freshly `populate`d patterns.db, in fact match the en effect
+  // signature in ALL 23 languages — so each keeps avgExecutionFidelity at 1.0. The
+  // wave-5 worklist's per-language divergence counts were measured against a STALE
+  // committed patterns.db snapshot (the committed copy lags the current dicts — see
+  // patterns-reference/CLAUDE.md); e.g. its ms `next-element` carried untranslated
+  // `to`/`next` (`apabila click … to next <li/>`), but a fresh populate emits the
+  // localized `apabila click tambah .highlight ke seterusnya <li/>`, which executes
+  // identically to en. Re-grounded counts (fresh db, all 24 priority langs): these six
+  // diverge in 0/23; the wave-5 worklist said next-element 1, toggle-aria-expanded 2,
+  // set-opacity 4, set-transform 4, accordion-toggle 6, caret-var-on-target 23 — all
+  // stale. No fixture/setup/trigger change needed; each en reference produces a clean
+  // non-empty signature against the existing fixture (next/closest positionals fall
+  // back to `me` consistently across every language; set *opacity/*transform write
+  // inline style; caret-var-on-target clears #btn text — the undefined `^count` resolves
+  // the same way in every language).
+  'next-element',
+  'toggle-aria-expanded',
+  'set-opacity',
+  'set-transform',
+  'accordion-toggle',
+  'caret-var-on-target',
+  // Expansion wave 7 (session 13): `multiple-events` — `on click or
+  // keypress[key=="Enter"] toggle .active`. The third wave-5 worklist divergence,
+  // now FIXED (semantic): a scoped or-clause excision pre-pass + the ja `または`→or
+  // tokenizer fix + hi/bn OR_KEYWORDS entries. It diverged in 7 languages
+  // (ja,ko,it,hi,tr,bn,qu — the translated `or` became a phantom body command or
+  // mangled into a selector); now all 23 match the en click effect (toggles .active
+  // on #btn). R2 stays 1.0.
+  'multiple-events',
+  // Expansion wave 8 (session 14): `put-before` / `put-after` — `put "<p>New</p>"
+  // before/after me`, the last two wave-5 worklist divergences (14 langs each), now
+  // FIXED. The position word (`before`/`after` + translations) is now captured as the
+  // put command's `manner` role (→ the DOM-insert modifier): per-language put-before/
+  // after patterns for SVO/SOV (#516) + handcrafted high-priority VSO put-event
+  // patterns for ar/tl/uk (the fused VSO event pattern consumed the position word as
+  // a plain destination marker and dropped manner). All 23 langs now insert the <p>
+  // before/after #btn (matching en); R2 stays 1.0.
+  'put-before',
+  'put-after',
+  // Expansion wave 9 (R2-coverage sweep): a fresh discovery probe (every
+  // non-subset, non-async pattern's en reference executed against the fixture,
+  // then all 23 translations checked against the en signature) surfaced three
+  // patterns whose en reference produces a clean, deterministic, non-empty
+  // effect AND that already match in ALL 23 languages — pure coverage, no
+  // parser/dict fix needed:
+  //  - chained-access-possessive-dot: `set my.parentElement.style.display to
+  //    "none"` — a synchronous chained-member write on the parent (.card).
+  //  - hide-with-transition / show-with-transition: `hide me with *opacity` /
+  //    `show #modal with *opacity`. Despite the names these are SYNCHRONOUS
+  //    hide/show STRATEGIES (the runtime writes display + a data-original-display
+  //    marker on hide, adds the visibility class on show) — no timer, no
+  //    animation frame, so the signature is deterministic (verified across two
+  //    back-to-back runs). Distinct from the async `transition … over Nms`
+  //    family the subset still excludes.
+  'chained-access-possessive-dot',
+  'hide-with-transition',
+  'show-with-transition',
+  // Expansion wave 10 (the SOV literal-role-extraction arc, PRs #560/#561):
+  // the two R2 blockers whose SOV translations silently dropped a bare literal
+  // role, now capturing it in all 23 languages:
+  //  - append-content: `append "<li>Item</li>" to #list` — the fronted content
+  //    literal bound to a bogus `event` role in the body-clause marker lookup
+  //    (ja/tr runtime "append requires content", bn silent no-op). Fixture adds
+  //    `<ul id="list">` (appended last, indices preserved).
+  //  - increment-by-amount: `increment #score by 10` — the trailing bare amount
+  //    was unconsumed by every fused event pattern and defaulted to 1 in the
+  //    SOV 6 + th (invisible to R0/R1: fillSchemaDefaults injects
+  //    quantity:literal=1 into the role signature — only R2 sees the value).
+  //    Fixture adds `<div id="score">0</div>` (appended last).
+  'append-content',
+  'increment-by-amount',
+];
+
+/**
+ * Shared fixture: every element the curated subset's selectors reference.
+ * The SAME fixture (plus the same per-pattern setup) is used for the en
+ * reference and every translation, so signatures are comparable.
+ */
+const FIXTURE_HTML = `<!DOCTYPE html><html><body>
+  <div class="card"><button id="btn">Click</button></div>
+  <div id="item"></div>
+  <div id="menu"></div>
+  <div id="panel"></div>
+  <div id="modal"></div>
+  <div id="output"></div>
+  <div id="counter">0</div>
+  <ul><li class="items active"></li><li class="items active"></li></ul>
+  <div class="tab active"></div>
+  <div class="tab"></div>
+  <div class="tab-panel"></div>
+  <div class="tab-panel"></div>
+  <div class="accordion-item open"></div>
+  <div id="container"></div>
+  <div class="dropdown-menu"></div>
+  <div id="sr-announce"></div>
+  <ul id="list"></ul>
+  <div id="score">0</div>
+</body></html>`;
+
+/** Per-pattern fixture preconditions (applied identically for every language). */
+const PATTERN_SETUP: Record<string, (doc: Document) => void> = {
+  // `remove .highlight from me` needs the class present to have an effect.
+  'remove-class-basic': doc => doc.getElementById('btn')!.classList.add('highlight'),
+  // `show #modal` needs the modal hidden.
+  'show-element': doc => {
+    (doc.getElementById('modal') as HTMLElement).style.display = 'none';
+  },
+  // `add .open to closest .accordion-item` needs #btn inside one; the
+  // fixture's standalone `.accordion-item.open` is the exclusivity victim.
+  'accordion-exclusive': doc => doc.querySelector('.card')!.classList.add('accordion-item'),
+  // `if target matches .modal-backdrop hide .modal-backdrop` — the dispatched
+  // click's target is #btn, so the condition only fires when #btn IS the
+  // backdrop (the real-page case is a click landing on the backdrop itself).
+  'modal-close-backdrop': doc => doc.getElementById('btn')!.classList.add('modal-backdrop'),
+  // `hide closest .modal` needs #btn inside a .modal (the real-page case is a
+  // close button inside the modal it closes), and `remove .modal-open from
+  // body` needs the class present so the body write is scoreable.
+  'modal-close-button': doc => {
+    doc.querySelector('.card')!.classList.add('modal');
+    doc.body.classList.add('modal-open');
+  },
+};
+
+/**
+ * Per-pattern trigger override. The subset is overwhelmingly `on click`, so the
+ * default (dispatched in `executeInner`) is a bare `Event('click')`. A handler
+ * that listens for a different event — or that READS the triggering event (e.g.
+ * `on success put event.detail.message …`) — declares the event name and, when
+ * needed, a `detail` payload here. With a `detail`, a `CustomEvent` is
+ * dispatched so `event.detail.*` resolves; the payload is identical for the en
+ * reference and every translation (it is harness-supplied, not translated), so
+ * effect signatures stay comparable across languages.
+ */
+interface PatternTrigger {
+  readonly event: string;
+  readonly detail?: unknown;
+}
+const PATTERN_TRIGGER: Record<string, PatternTrigger> = {
+  // `on success put event.detail.message into #sr-announce set @role to "alert"
+  // on #sr-announce` — a custom event carrying the announcement text. The fixed
+  // message lands in #sr-announce's text; role=alert is set on the same node.
+  'announce-screen-reader': { event: 'success', detail: { message: 'Saved successfully' } },
+};
+
+/** Result of executing one pattern translation. */
+export interface ExecutionResult {
+  codeExampleId: string;
+  language: LanguageCode;
+  /** Sorted effect lines (see snapshot/diff). Empty on error or no effect. */
+  effects: string[];
+  /** Parse/build/runtime error, when any. */
+  error?: string;
+  /** 1 = exact signature match vs the en reference, 0 = mismatch. Undefined for en itself or when no usable reference exists. */
+  executionFidelity?: number;
+}
+
+/**
+ * Serialize the identity-relevant state of every element under <body>.
+ * Keyed by #id when present, else tag[document-order-index]; the fixture is
+ * identical across languages, so keys are comparable.
+ */
+function serializeElement(el: Element): string {
+  const attrs = Array.from(el.attributes)
+    .filter(a => a.name !== 'class' && a.name !== 'style')
+    .map(a => `${a.name}=${a.value}`)
+    .sort()
+    .join(',');
+  const classes = Array.from(el.classList).sort().join(' ');
+  const style = (el as HTMLElement).getAttribute('style') ?? '';
+  // Leaf text only — container text would duplicate every child mutation.
+  const text =
+    el.childNodes.length === 0 || (el.childNodes.length === 1 && el.firstChild?.nodeType === 3)
+      ? (el.textContent ?? '')
+      : '';
+  return `cls[${classes}] attr[${attrs}] style[${style}] text[${text}]`;
+}
+
+function snapshot(document: Document): Map<string, string> {
+  const out = new Map<string, string>();
+  // body participates under its own stable key: body-targeted effects
+  // (`add .modal-open to body`, modal-open/modal-close-button) must be
+  // visible in the signature now that the runtime resolves `body`. Before
+  // wave 3b these writes fell back to `me` (visible by accident); a correct
+  // body write was invisible and a dropped one unscoreable.
+  out.set('body', serializeElement(document.body));
+  document.body.querySelectorAll('*').forEach((el, i) => {
+    const key = el.id ? `#${el.id}` : `${el.tagName.toLowerCase()}[${i}]`;
+    out.set(key, serializeElement(el));
+  });
+  return out;
+}
+
+/** Sorted, stable list of per-element changes between two snapshots. */
+function diffSnapshots(before: Map<string, string>, after: Map<string, string>): string[] {
+  const effects: string[] = [];
+  for (const [k, v] of after) {
+    const b = before.get(k);
+    if (b === undefined) effects.push(`+${k} ${v}`);
+    else if (b !== v) effects.push(`Δ${k} ${v}`);
+  }
+  for (const k of before.keys()) if (!after.has(k)) effects.push(`-${k}`);
+  return effects.sort();
+}
+
+/** How long to let the dispatched handler settle (ms). The subset contains no
+ * waits/transitions/fetches, so this only needs to drain micro/macrotasks. */
+const SETTLE_MS = 20;
+
+/** Per-execution hard timeout — a hung execution must not hang the sweep. */
+const EXECUTION_TIMEOUT_MS = 5000;
+
+/**
+ * R2 execution validator. `initialize()` MUST complete before `execute()`:
+ * it bootstraps jsdom globals and only then loads `@hyperfixi/core`, whose
+ * module evaluation touches `Element`/`document`.
+ */
+export class ExecutionValidator {
+  private core: {
+    Runtime: new () => { execute(ast: unknown, ctx: unknown): Promise<unknown> };
+    createContext: (el: HTMLElement) => unknown;
+  } | null = null;
+
+  /** Errors thrown inside dispatched listeners surface as unhandled rejections
+   * (the handler is async); trap them into the current execution's sink. */
+  private listenerErrors: string[] = [];
+  private trapInstalled = false;
+
+  async initialize(): Promise<void> {
+    if (this.core) return;
+
+    // Bootstrap DOM globals BEFORE importing @hyperfixi/core (its dist
+    // evaluates `Element` at module load). The per-execution fresh document
+    // replaces these for each run.
+    this.installGlobals(new JSDOM(FIXTURE_HTML));
+
+    const core = await import('@hyperfixi/core');
+    this.core = {
+      Runtime: core.Runtime as unknown as new () => {
+        execute(ast: unknown, ctx: unknown): Promise<unknown>;
+      },
+      createContext: core.createContext as unknown as (el: HTMLElement) => unknown,
+    };
+
+    if (!this.trapInstalled) {
+      process.on('unhandledRejection', (reason: unknown) => {
+        const msg = reason instanceof Error ? reason.message : String(reason);
+        this.listenerErrors.push(msg);
+      });
+      this.trapInstalled = true;
+    }
+  }
+
+  getName(): string {
+    return 'ExecutionValidator';
+  }
+
+  private installGlobals(dom: JSDOM): void {
+    const g = globalThis as Record<string, unknown>;
+    g.window = dom.window;
+    g.document = dom.window.document;
+    g.Event = dom.window.Event;
+    g.CustomEvent = dom.window.CustomEvent;
+    g.HTMLElement = dom.window.HTMLElement;
+    g.Element = dom.window.Element;
+    g.Node = dom.window.Node;
+    g.MutationObserver = dom.window.MutationObserver;
+    g.getComputedStyle = dom.window.getComputedStyle;
+  }
+
+  /**
+   * Execute one pattern translation in a fresh fixture and return its effect
+   * signature. Never throws; failures come back as `error`.
+   */
+  async execute(codeExampleId: string, code: string, lang: string): Promise<ExecutionResult> {
+    const run = this.executeInner(codeExampleId, code, lang);
+    const timeout = new Promise<ExecutionResult>(resolve =>
+      setTimeout(
+        () =>
+          resolve({
+            codeExampleId,
+            language: lang as LanguageCode,
+            effects: [],
+            error: `execution timed out (${EXECUTION_TIMEOUT_MS}ms)`,
+          }),
+        EXECUTION_TIMEOUT_MS
+      )
+    );
+    return Promise.race([run, timeout]);
+  }
+
+  private async executeInner(
+    codeExampleId: string,
+    code: string,
+    lang: string
+  ): Promise<ExecutionResult> {
+    if (!this.core) await this.initialize();
+    const base: ExecutionResult = {
+      codeExampleId,
+      language: lang as LanguageCode,
+      effects: [],
+    };
+
+    const dom = new JSDOM(FIXTURE_HTML);
+    this.installGlobals(dom);
+    const document = dom.window.document;
+    PATTERN_SETUP[codeExampleId]?.(document);
+    const btn = document.getElementById('btn')!;
+    this.listenerErrors = [];
+
+    // The runtime logs every failing command via console — across a full
+    // multi-language sweep that is pure noise; silence it for the execution.
+    const saved = { log: console.log, warn: console.warn, error: console.error };
+    console.log = console.warn = console.error = () => {};
+    try {
+      const parsed = parseSemantic(code, lang);
+      if (!parsed.node || parsed.confidence < 0.5) {
+        return { ...base, error: `parse failed (confidence ${parsed.confidence?.toFixed(2)})` };
+      }
+      const built = buildAST(parsed.node);
+      if (!built.ast) return { ...base, error: 'buildAST returned no AST' };
+
+      const runtime = new this.core!.Runtime();
+      const ctx = this.core!.createContext(btn as unknown as HTMLElement);
+      // Installs the event handler (or runs bare commands immediately).
+      await runtime.execute(built.ast, ctx);
+
+      const before = snapshot(document);
+      const trigger = PATTERN_TRIGGER[codeExampleId];
+      const triggerEvent =
+        trigger?.detail !== undefined
+          ? new dom.window.CustomEvent(trigger.event, { bubbles: true, detail: trigger.detail })
+          : new dom.window.Event(trigger?.event ?? 'click', { bubbles: true });
+      btn.dispatchEvent(triggerEvent);
+      await new Promise(r => setTimeout(r, SETTLE_MS));
+      const after = snapshot(document);
+
+      const effects = diffSnapshots(before, after);
+      if (this.listenerErrors.length > 0) {
+        return { ...base, effects, error: `runtime: ${this.listenerErrors.join('; ')}` };
+      }
+      return { ...base, effects };
+    } catch (e: unknown) {
+      return { ...base, error: e instanceof Error ? e.message : String(e) };
+    } finally {
+      console.log = saved.log;
+      console.warn = saved.warn;
+      console.error = saved.error;
+      dom.window.close();
+    }
+  }
+}
+
+/** The subset's source text per language: en from code_examples.raw_code,
+ * other languages from pattern_translations. */
+export async function loadExecutionSubset(
+  languages: readonly LanguageCode[]
+): Promise<Map<LanguageCode, Map<string, string>>> {
+  const subset = new Set(EXECUTION_SUBSET);
+  const result = new Map<LanguageCode, Map<string, string>>();
+
+  for (const lang of languages) {
+    const byId = new Map<string, string>();
+    if (lang === 'en') {
+      const patterns = await getAllPatterns();
+      for (const p of patterns) {
+        if (subset.has(p.id)) byId.set(p.id, p.rawCode);
+      }
+    } else {
+      const translations = await getTranslationsByLanguage(lang, 1000);
+      for (const t of translations) {
+        if (subset.has(t.codeExampleId)) byId.set(t.codeExampleId, t.hyperscript);
+      }
+    }
+    result.set(lang, byId);
+  }
+  return result;
+}

package/src/multilingual/validators/parse-validator.ts

@@ -4,7 +4,9 @@
 
 import { MultilingualHyperscript } from '@hyperfixi/core/multilingual';
 import type { SemanticNode } from '@lokascript/semantic';
+import { fillSchemaDefaults } from '@lokascript/semantic';
 import type { PatternTranslation, ParseResult, Validator } from '../types';
+import { collectActions, collectActionsMultiset, collectRoleSignature } from '../fidelity';
 
 /**
  * Parse Validator
@@ -77,6 +79,18 @@ export class ParseValidator implements Validator<ParseResult[]> {
       const command = semanticNode.action;
       const roles = this.extractRoles(semanticNode);
 
+      // R1 measurement normalization: materialize schema role DEFAULTS (e.g.
+      // increment.quantity → 1, toggle.destination → me) on this throwaway parse
+      // before signature collection. The en pattern path materializes them but the
+      // SOV path doesn't, so an SOV parse would otherwise read as dropping a role it
+      // defaults identically at runtime — a false-positive in role recall. Applied
+      // uniformly here (en reference AND every translation), it only removes that
+      // false-positive. NOT done in `parse()` itself, which would make the renderer
+      // emit the materialized defaults and break round-trips. Mutation is safe: this
+      // node is the validator's own and is used only for the signatures below
+      // (actions/multiset are role-agnostic, so they're unaffected).
+      fillSchemaDefaults(semanticNode);
+
       return {
         pattern,
         success: true,
@@ -84,6 +98,14 @@ export class ParseValidator implements Validator<ParseResult[]> {
         roles,
         confidence: pattern.confidence, // Use pattern confidence
         parser: 'semantic',
+        // Structural signature for cross-language fidelity scoring (see fidelity.ts).
+        actionSignature: collectActions(semanticNode),
+        // R0-precision multiset signature — duplicates preserved so a phantom
+        // command duplicated ahead of a real one (`[toggle, toggle]`) is visible.
+        actionMultisetSignature: collectActionsMultiset(semanticNode),
+        // R1 role signature (role name + value type per command) — collected here
+        // because live nodes carry roles as a ReadonlyMap that serializes to {}.
+        roleSignature: collectRoleSignature(semanticNode),
         duration: performance.now() - startTime,
       };
     } catch (error) {