@mailwoman/corpus 10.0.0 → 10.2.0

package/out/fr/recipes/fragment.js

@@ -3,87 +3,51 @@
  * @license AGPL-3.0
  * @author Teffen Ellis, et al.
  *
- *   `fr-fragment` slice recipe (#727 T2) — the HOUSE-NUMBER-LICENCE change.
+ *   Generates French street fragments without a postcode from BAN tuples.
  *
- *   The measured problem (T1c, `2026-07-16-t1c-fragment-board-verdict.md`): the shipped model scores
- *   **0.925** on `<n> Rue X` and **0.215** on `Rue X`. Same streets, same model; the only difference
- *   is a leading number. It has learned that a house number LICENSES a street reading at all — strip
- *   it and a designator-led phrase parses as a locality, designator included:
- *
- *   ```
- *   "Rue Montmartre"        -> locality="Rue Montmartre"
- *   "Allee Poque"           -> locality="Allee Poque"
- *   ```
- *
- *   `Rue` can only mean street in French. The model is not mislabelling an ambiguous toponym; it is
- *   mislabelling `Rue`.
- *
- *   WHY THE EXISTING RECIPE DOESN'T COVER THIS. {@link frBareStreetRecipe} (#251) mints
- *   `<n> Rue <name>, <City>` — the bare COMMA form, no postcode. It targets postcode-anchoring
- *   imbalance, and every row it emits still carries a house number AND a locality. It cannot teach
- *   the class above, because it never shows the model a street standing alone.
- *
- *   WHAT THIS MINTS. Five street forms — the first three carry NO house number and NO locality, which
- *   is the whole point — plus the counter-distribution:
- *
- *   1. `bare-street`         "Rue Montmartre"
- *   2. `street-particle`     "Rue de la Paix"
- *   3. `date-name`           "Allee du 11 Novembre 1918"
- *   4. `street-housenumber`  "12 Rue Montmartre"     — the anchor, so the licence isn't UNLEARNED
- *   5. `alnum-housenumber`   "12 bis Rue Montmartre"
- *   6. `bare-locality`       "Mery-sur-Oise"         — NEGATIVE: a bare toponym that IS a locality
- *
- *   The **admin/street homonym** class (`Rue de Rome`) is deliberately NOT a separate form. Homonym
- *   streets are already in the register and fall into 1–3 naturally; there is nothing different to
- *   teach about them, because the lesson is identical — the designator makes it a street. The
- *   fragment board scores them separately (they need measuring, not minting), and the recipe would
- *   need a second pass over the commune set to label them, buying a `synth_method` string and no
- *   training signal.
- *
- *   FORM 7 IS NOT OPTIONAL. T1c's standing prediction: the board's `bare-locality` cell reads 0.980
- *   for the WRONG REASON — the model calls everything without a house number a locality, and on bare
- *   localities that is accidentally right. Teach bare streets alone and the model has every incentive
- *   to flip that default rather than learn the distinction, trading a 0.215 for a 0.980. The slice
- *   must show BOTH bare forms so the discriminating evidence is the designator, which is the only
- *   thing that actually distinguishes them. This is the same counter-distribution principle
- *   {@link noStreetRecipe} established after synth-street pushed the model into "decompose mode".
- *
- *   SPLIT. `--exclude-surfaces` takes the fragment board's reserved surface list
- *   (`mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt`). Every listed surface is
- *   skipped — source-disjoint by normalized street SURFACE, never by record row. Row-disjoint leaks
- *   the surface across the boundary and measures memorization of `Rue de Rivoli` while claiming
- *   generalization to unseen streets. The recipe REFUSES to run without the list rather than
- *   silently minting a contaminated slice.
- *
- *   MIX. The slice is ~145K rows off a 120K-tuple draw; the intended corpus mix is **5–10%**, set at
- *   assembly time by slice weight rather than by row count. Keep the cap: a slice that fixes fragments
- *   by degrading full addresses has moved the failure, not fixed it — which is what
- *   `street-housenumber` / `alnum-housenumber` on the fragment board and the global parity floor are
- *   there to catch. `date-name` is ~0.5% of the slice because BAN only holds ~1,418 date-name streets
- *   after filtering and the tuple extractor already takes every one; if that class needs more, the
- *   change is slice weight, not invented data.
- *
- *   ⚠ Convention loss-mask: like {@link frBareStreetRecipe}, this recipe TEACHES FR `street_prefix`.
- *   The conventions loss-mask forbids it for FR and will `-inf` these gold labels (the v1.6.0 ~7M-loss
- *   blow-up). Disable that mask for any run including this slice.
+ *   The output mixes designator-led streets with bare commune names as counterexamples.
+ *   Some streets include a house number.
+ *   `--exclude-surfaces` keeps streets on the fragment evaluation board out of training. A model trained on
+ *   these rows needs the French `street_prefix` loss mask disabled.
  */
+import { titleCaseFR } from "@mailwoman/codex/fr";
+import { sample } from "@mailwoman/core/random";
 import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils";
 import { TextSpliterator } from "spliterator";
-import { decomposeFrStreet } from "#fr/adapters/ban/street-decompose";
-import { alignAndWrite, readTuples, sliceSourceID } from "#recipes/scaffold";
-import { pick } from "#synthesizers/utils";
+import { decomposeFrStreet } from "#fr/street-decompose";
+import { alignAndWrite, readTuples, recipeSourceID } from "#recipes/scaffold";
+import { defaultRecipeSource } from "#recipes/sources";
+import { SourceRegister } from "#registers";
+import { SurfaceOrigin } from "#types";
+/**
+ * The provenance for every row.
+ * Street and commune names both come from BAN.
+ */
 /**
- * House numbers, weighted toward the small values that dominate real BAN rows.
+ * Resolved once, because `source_id` includes the source as its prefix and a pair
+ * that disagreed would name a source no row of this output declares.
+ */
+const FR_FRAGMENT_SOURCE = defaultRecipeSource("synth-fr-fragment");
+const FR_FRAGMENT_PROVENANCE = {
+    register: SourceRegister.BaseAdresseNationale,
+    surface: SurfaceOrigin.Composed,
+    // The BAN record behind each row exists.
+    // The tuple file has no `sourceID` to identify it.
+    // The extraction must emit that ID to populate this field (#2359).
+    baseSourceID: null,
+};
+/**
+ * House numbers to sample from.
+ * Repeated entries weight the draw toward small values.
  */
 const HOUSE_NUMBERS = [
     1, 1, 2, 2, 3, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 16, 18, 20, 21, 24, 27, 30, 33, 42, 57, 68, 84, 102, 115, 140,
 ];
 /**
- * FR alphanumeric house-number forms. `bis`/`ter` are separated; a bare letter is suffixed.
+ * French repetition and letter suffixes for house numbers.
  */
 const ALNUM_SUFFIXES = ["bis", "ter", "A", "B"];
-// Recipe-fidelity: this accent-STRIPPING fold is the surface key every committed fr-fragment slice and the fragment
-// board's reserved list were built with — not the diacritic-KEEPING `foldNOSurface` the Norwegian slices share.
+// The reserved-surface list is stored in this folded form: accents stripped, lowercase, single spaces.
 const norm = (value) => value
     .normalize("NFD")
     .replaceAll(/[̀-ͯ]/g, "")
@@ -91,58 +55,21 @@ const norm = (value) => value
     .replaceAll(/\s+/g, " ")
     .trim();
 /**
- * French commune convention: capitalize each element, leave the joining particles lowercase. `saint-jean-de-luz` →
- * `Saint-Jean-de-Luz`, not `Saint-Jean-De-Luz`.
- *
- * Needed because BAN's sliced DBs keep only `locality_base` — normalized, lowercase, accent-stripped. Emitting that
- * verbatim would teach the counter-distribution that a lowercase accent-stripped string is a locality, which is not a
- * fact about French and would not match the fragment board (which reconstructs the same casing). The accents are gone
- * from the source and cannot be recovered here; the casing can.
- */
-const FR_LOWER = new Set([
-    "le",
-    "la",
-    "les",
-    "de",
-    "du",
-    "des",
-    "d",
-    "l",
-    "sur",
-    "sous",
-    "en",
-    "aux",
-    "au",
-    "et",
-    "lez",
-]);
-export function frTitleCase(value) {
-    const cap = (token, first) => !first && FR_LOWER.has(token) ? token : token.charAt(0).toUpperCase() + token.slice(1);
-    return value
-        .split(" ")
-        .map((word, wordIndex) => word
-        .split("-")
-        .map((bit, bitIndex) => cap(bit, wordIndex === 0 && bitIndex === 0))
-        .join("-"))
-        .join(" ");
-}
-/**
- * Does the street name carry a particle? Decides the particle vs bare classification.
+ * Matches common French particles in a street name.
  */
 const PARTICLE = /\b(de la|de l'|du|des|de|d'|le|la|les)\b/i;
 /**
- * Does the street name carry date material (a year, or a day + French month)?
+ * Matches a year or a French day-and-month phrase in a street name.
  */
 const DATEISH = /\b(1[0-9]|20)\d{2}\b|\b\d{1,2}\s+(janvier|f[ée]vrier|mars|avril|mai|juin|juillet|ao[ûu]t|septembre|octobre|novembre|d[ée]cembre)\b/i;
-// A quarter of house numbers carry an alphanumeric suffix (`8 bis`, `12B`).
+// The share of numbered rows whose house number has a suffix.
 const ALNUM_HOUSE_NUMBER_SHARE = 0.25;
 /**
- * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
- * and `description` below for the surface form it generates.
+ * The recipe registered with the corpus builder.
  */
 export const frFragmentRecipe = {
     name: "fr-fragment",
-    description: "FR street fragments with NO house number (#727 T2): the house-number-licence change — bare/particle/date-name/homonym + the bare-locality counter",
+    description: "FR street fragments with NO house number (#727 T2): the house-number-license change — bare/particle/date-name/homonym + the bare-locality counter",
     mode: "tuples",
     options: [
         {
@@ -163,22 +90,20 @@ export const frFragmentRecipe = {
         const excludePath = opts.excludeSurfaces;
         if (!excludePath) {
             throw new Error("fr-fragment: --exclude-surfaces is REQUIRED. Pass the fragment board's reserved list " +
-                "(mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt) or this slice trains on its own eval set. " +
+                "(mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt) or this recipe trains on its own eval set. " +
                 "Source-disjoint by street SURFACE is the split discipline; there is no safe default.");
         }
         const excluded = new Set();
         for await (const line of TextSpliterator.fromAsync(excludePath)) {
-            const trimmed = line.trim();
-            if (trimmed && !trimmed.startsWith("#")) {
-                excluded.add(trimmed);
+            if (!line.startsWith("#")) {
+                excluded.add(line);
             }
         }
         if (!excluded.size)
             throw new Error(`fr-fragment: --exclude-surfaces "${excludePath}" listed no surfaces`);
         const hnProb = opts.hnProb ?? 0.35;
         const bareLocalityProb = opts.bareProb ?? 0.25;
-        // The locality pool is harvested from the tuples themselves — every BAN row carries its
-        // commune, so the counter-distribution needs no second source.
+        // Commune names collected here feed the bare-locality counterexamples.
         const localities = new Set();
         let read = 0;
         let emitted = 0;
@@ -195,14 +120,12 @@ export const frFragmentRecipe = {
                 skipped++;
                 continue;
             }
-            // THE SPLIT. A surface on the fragment board never enters training.
             if (excluded.has(norm(fullStreet))) {
                 contaminated++;
                 continue;
             }
             const { prefix, street } = decomposeFrStreet(fullStreet);
-            // The failing class is the designator-led street. A no-prefix nom_voie ("La Ville Mois")
-            // is a different problem and would muddy the signal.
+            // The recipe keeps only streets that start with a designator such as "rue".
             if (!prefix || !street) {
                 skipped++;
                 continue;
@@ -212,9 +135,9 @@ export const frFragmentRecipe = {
             let raw = `${prefix} ${street}`;
             let klass = DATEISH.test(street) ? "date-name" : PARTICLE.test(street) ? "street-particle" : "bare-street";
             if (carriesNumber) {
-                const number = pick(HOUSE_NUMBERS, random);
+                const number = sample(HOUSE_NUMBERS, random);
                 const alnum = random() < ALNUM_HOUSE_NUMBER_SHARE;
-                const suffix = pick(ALNUM_SUFFIXES, random);
+                const suffix = sample(ALNUM_SUFFIXES, random);
                 const houseNumber = alnum
                     ? suffix === "bis" || suffix === "ter"
                         ? `${number} ${suffix}`
@@ -224,44 +147,41 @@ export const frFragmentRecipe = {
                 raw = `${houseNumber} ${prefix} ${street}`;
                 klass = alnum ? "alnum-housenumber" : "street-housenumber";
             }
-            const sourceID = sliceSourceID("synth-fr-fragment", { ...components, v: String(read) });
+            const sourceID = recipeSourceID(FR_FRAGMENT_SOURCE, { ...components, v: String(read) });
             if (alignAndWrite(write, {
                 raw,
                 components,
                 country: "FR",
                 locale: "fr-FR",
-                source: "synth-fr-fragment",
+                source: FR_FRAGMENT_SOURCE,
                 source_id: sourceID,
                 corpus_version: "0.9.4",
                 license: "Synthetic — fr-fragment; (street, commune) from BAN (Base Adresse Nationale, Licence Ouverte)",
-            }, `fr-fragment:${klass}`)) {
+            }, `fr-fragment:${klass}`, FR_FRAGMENT_PROVENANCE)) {
                 emitted++;
             }
             else {
                 skipped++;
             }
         }
-        // MARK: counter-distribution — bare localities
-        // Minted last so the locality pool is complete. Without these the model can satisfy every row
-        // above by flipping its default from "bare => locality" to "bare => street", which trades one
-        // broken prior for another and would show up as bare-locality collapsing on the board.
+        // MARK: Bare-locality counterexamples
+        // The count makes bare localities `bareLocalityProb` of all emitted rows.
         const pool = [...localities].toSorted();
         const wanted = Math.round((emitted / Math.max(1, 1 - bareLocalityProb)) * bareLocalityProb);
         for (let i = 0; i < wanted && pool.length; i++) {
-            // BAN gives `locality_base` normalized; restore the casing the fragment board also
-            // reconstructs, so train and eval show the model the same shape of French.
-            const name = frTitleCase(pick(pool, random));
-            const sourceID = sliceSourceID("synth-fr-fragment", { locality: name, v: `neg-${i}` });
+            // The fragment evaluation board uses title-cased commune names.
+            const name = titleCaseFR(sample(pool, random));
+            const sourceID = recipeSourceID(FR_FRAGMENT_SOURCE, { locality: name, v: `neg-${i}` });
             if (alignAndWrite(write, {
                 raw: name,
                 components: { locality: name },
                 country: "FR",
                 locale: "fr-FR",
-                source: "synth-fr-fragment",
+                source: FR_FRAGMENT_SOURCE,
                 source_id: sourceID,
                 corpus_version: "0.9.4",
                 license: "Synthetic — fr-fragment counter-distribution; commune from BAN (Licence Ouverte)",
-            }, "fr-fragment:bare-locality")) {
+            }, "fr-fragment:bare-locality", FR_FRAGMENT_PROVENANCE)) {
                 emitted++;
             }
             else {

package/out/fr/recipes/fragment.js.map

@@ -1 +1 @@
-{"version":3,"file":"fragment.js","sourceRoot":"","sources":["../../../lib/fr/recipes/fragment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoEG;AAEH,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AACpE,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAA;AAE7C,OAAO,EAAE,iBAAiB,EAAE,MAAM,mCAAmC,CAAA;AACrE,OAAO,EAAE,aAAa,EAAE,UAAU,EAAqB,aAAa,EAAE,MAAM,mBAAmB,CAAA;AAC/F,OAAO,EAAE,IAAI,EAAE,MAAM,qBAAqB,CAAA;AAE1C;;GAEG;AAEH,MAAM,aAAa,GAAG;IACrB,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG;CACrH,CAAA;AAED;;GAEG;AACH,MAAM,cAAc,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,GAAG,CAAC,CAAA;AAE/C,oHAAoH;AACpH,gHAAgH;AAChH,MAAM,IAAI,GAAG,CAAC,KAAa,EAAU,EAAE,CACtC,KAAK;KACH,SAAS,CAAC,KAAK,CAAC;KAChB,UAAU,CAAC,QAAQ,EAAE,EAAE,CAAC;KACxB,WAAW,EAAE;KACb,UAAU,CAAC,MAAM,EAAE,GAAG,CAAC;KACvB,IAAI,EAAE,CAAA;AAET;;;;;;;;GAQG;AACH,MAAM,QAAQ,GAAG,IAAI,GAAG,CAAC;IACxB,IAAI;IACJ,IAAI;IACJ,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,KAAK;IACL,GAAG;IACH,GAAG;IACH,KAAK;IACL,MAAM;IACN,IAAI;IACJ,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,KAAK;CACL,CAAC,CAAA;AAEF,MAAM,UAAU,WAAW,CAAC,KAAa;IACxC,MAAM,GAAG,GAAG,CAAC,KAAa,EAAE,KAAc,EAAU,EAAE,CACrD,CAAC,KAAK,IAAI,QAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAA;IAEvF,OAAO,KAAK;SACV,KAAK,CAAC,GAAG,CAAC;SACV,GAAG,CAAC,CAAC,IAAI,EAAE,SAAS,EAAE,EAAE,CACxB,IAAI;SACF,KAAK,CAAC,GAAG,CAAC;SACV,GAAG,CAAC,CAAC,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,EAAE,SAAS,KAAK,CAAC,IAAI,QAAQ,KAAK,CAAC,CAAC,CAAC;SACnE,IAAI,CAAC,GAAG,CAAC,CACX;SACA,IAAI,CAAC,GAAG,CAAC,CAAA;AACZ,CAAC;AAED;;GAEG;AACH,MAAM,QAAQ,GAAG,2CAA2C,CAAA;AAE5D;;GAEG;AACH,MAAM,OAAO,GACZ,qIAAqI,CAAA;AAEtI,4EAA4E;AAC5E,MAAM,wBAAwB,GAAG,IAAI,CAAA;AAErC;;;GAGG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAiB;IAC7C,IAAI,EAAE,aAAa;IACnB,WAAW,EACV,mJAAmJ;IACpJ,IAAI,EAAE,QAAQ;IACd,OAAO,EAAE;QACR;YACC,IAAI,EAAE,2BAA2B;YACjC,WAAW,EAAE,uFAAuF;SACpG;QACD;YACC,IAAI,EAAE,eAAe;YACrB,WAAW,EAAE,kFAAkF;SAC/F;QACD;YACC,IAAI,EAAE,iBAAiB;YACvB,WAAW,EAAE,qFAAqF;SAClG;KACD;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,WAAW,GAAG,IAAI,CAAC,eAAe,CAAA;QAExC,IAAI,CAAC,WAAW,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CACd,uFAAuF;gBACtF,4GAA4G;gBAC5G,sFAAsF,CACvF,CAAA;QACF,CAAC;QAED,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAU,CAAA;QAElC,IAAI,KAAK,EAAE,MAAM,IAAI,IAAI,eAAe,CAAC,SAAS,CAAC,WAAW,CAAC,EAAE,CAAC;YACjE,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,CAAA;YAE3B,IAAI,OAAO,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;gBACzC,QAAQ,CAAC,GAAG,CAAC,OAAO,CAAC,CAAA;YACtB,CAAC;QACF,CAAC;QAED,IAAI,CAAC,QAAQ,CAAC,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,oCAAoC,WAAW,sBAAsB,CAAC,CAAA;QAE1G,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAA;QAClC,MAAM,gBAAgB,GAAG,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAA;QAE9C,wFAAwF;QACxF,+DAA+D;QAC/D,MAAM,UAAU,GAAG,IAAI,GAAG,EAAU,CAAA;QAEpC,IAAI,IAAI,GAAG,CAAC,CAAA;QACZ,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,YAAY,GAAG,CAAC,CAAA;QAEpB,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,UAAU,CAAC,IAAI,CAAC,KAAM,CAAC,EAAE,CAAC;YACnD,IAAI,EAAE,CAAA;YACN,MAAM,UAAU,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YACpD,MAAM,QAAQ,GAAG,MAAM,CAAC,KAAK,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAEpD,IAAI,QAAQ,EAAE,CAAC;gBACd,UAAU,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;YACzB,CAAC;YAED,IAAI,CAAC,UAAU,EAAE,CAAC;gBACjB,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,oEAAoE;YACpE,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC;gBACpC,YAAY,EAAE,CAAA;gBAEd,SAAQ;YACT,CAAC;YAED,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,iBAAiB,CAAC,UAAU,CAAC,CAAA;YAExD,yFAAyF;YACzF,qDAAqD;YACrD,IAAI,CAAC,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC;gBACxB,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,MAAM,aAAa,GAAG,MAAM,EAAE,GAAG,MAAM,CAAA;YACvC,MAAM,UAAU,GAA2B,EAAE,aAAa,EAAE,MAAM,EAAE,MAAM,EAAE,CAAA;YAC5E,IAAI,GAAG,GAAG,GAAG,MAAM,IAAI,MAAM,EAAE,CAAA;YAC/B,IAAI,KAAK,GAAG,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,aAAa,CAAA;YAE1G,IAAI,aAAa,EAAE,CAAC;gBACnB,MAAM,MAAM,GAAG,IAAI,CAAC,aAAa,EAAE,MAAM,CAAC,CAAA;gBAC1C,MAAM,KAAK,GAAG,MAAM,EAAE,GAAG,wBAAwB,CAAA;gBACjD,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,EAAE,MAAM,CAAC,CAAA;gBAE3C,MAAM,WAAW,GAAG,KAAK;oBACxB,CAAC,CAAC,MAAM,KAAK,KAAK,IAAI,MAAM,KAAK,KAAK;wBACrC,CAAC,CAAC,GAAG,MAAM,IAAI,MAAM,EAAE;wBACvB,CAAC,CAAC,GAAG,MAAM,GAAG,MAAM,EAAE;oBACvB,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAA;gBAEjB,UAAU,CAAC,YAAY,GAAG,WAAW,CAAA;gBACrC,GAAG,GAAG,GAAG,WAAW,IAAI,MAAM,IAAI,MAAM,EAAE,CAAA;gBAC1C,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,mBAAmB,CAAC,CAAC,CAAC,oBAAoB,CAAA;YAC3D,CAAC;YAED,MAAM,QAAQ,GAAG,aAAa,CAAC,mBAAmB,EAAE,EAAE,GAAG,UAAU,EAAE,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;YAEvF,IACC,aAAa,CACZ,KAAK,EACL;gBACC,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,mBAAmB;gBAC3B,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,+FAA+F;aACxG,EACD,eAAe,KAAK,EAAE,CACtB,EACA,CAAC;gBACF,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,+CAA+C;QAC/C,8FAA8F;QAC9F,8FAA8F;QAC9F,uFAAuF;QACvF,MAAM,IAAI,GAAG,CAAC,GAAG,UAAU,CAAC,CAAC,QAAQ,EAAE,CAAA;QACvC,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,gBAAgB,CAAC,CAAC,GAAG,gBAAgB,CAAC,CAAA;QAE3F,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAChD,mFAAmF;YACnF,2EAA2E;YAC3E,MAAM,IAAI,GAAG,WAAW,CAAC,IAAI,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC,CAAA;YAC5C,MAAM,QAAQ,GAAG,aAAa,CAAC,mBAAmB,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,EAAE,CAAC,CAAA;YAEtF,IACC,aAAa,CACZ,KAAK,EACL;gBACC,GAAG,EAAE,IAAI;gBACT,UAAU,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;gBAC9B,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,mBAAmB;gBAC3B,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,kFAAkF;aAC3F,EACD,2BAA2B,CAC3B,EACA,CAAC;gBACF,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,IAAI,YAAY,EAAE,CAAC;YAClB,OAAO,CAAC,KAAK,CAAC,wBAAwB,YAAY,8DAA8D,CAAC,CAAA;QAClH,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAClC,CAAC;CACD,CAAA"}
+{"version":3,"file":"fragment.js","sourceRoot":"","sources":["../../../lib/fr/recipes/fragment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,qBAAqB,CAAA;AACjD,OAAO,EAAE,MAAM,EAAE,MAAM,wBAAwB,CAAA;AAC/C,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AACpE,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAA;AAE7C,OAAO,EAAE,iBAAiB,EAAE,MAAM,sBAAsB,CAAA;AACxD,OAAO,EAAE,aAAa,EAAE,UAAU,EAAqB,cAAc,EAAE,MAAM,mBAAmB,CAAA;AAChG,OAAO,EAAE,mBAAmB,EAAE,MAAM,kBAAkB,CAAA;AACtD,OAAO,EAAE,cAAc,EAAE,MAAM,YAAY,CAAA;AAC3C,OAAO,EAAE,aAAa,EAAE,MAAM,QAAQ,CAAA;AAEtC;;;GAGG;AACH;;;GAGG;AACH,MAAM,kBAAkB,GAAG,mBAAmB,CAAC,mBAAmB,CAAC,CAAA;AAEnE,MAAM,sBAAsB,GAAG;IAC9B,QAAQ,EAAE,cAAc,CAAC,oBAAoB;IAC7C,OAAO,EAAE,aAAa,CAAC,QAAQ;IAC/B,yCAAyC;IACzC,mDAAmD;IACnD,mEAAmE;IACnE,YAAY,EAAE,IAAI;CAClB,CAAA;AAED;;;GAGG;AACH,MAAM,aAAa,GAAG;IACrB,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG;CACrH,CAAA;AAED;;GAEG;AACH,MAAM,cAAc,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,GAAG,CAAC,CAAA;AAE/C,uGAAuG;AACvG,MAAM,IAAI,GAAG,CAAC,KAAa,EAAU,EAAE,CACtC,KAAK;KACH,SAAS,CAAC,KAAK,CAAC;KAChB,UAAU,CAAC,QAAQ,EAAE,EAAE,CAAC;KACxB,WAAW,EAAE;KACb,UAAU,CAAC,MAAM,EAAE,GAAG,CAAC;KACvB,IAAI,EAAE,CAAA;AAET;;GAEG;AACH,MAAM,QAAQ,GAAG,2CAA2C,CAAA;AAE5D;;GAEG;AACH,MAAM,OAAO,GACZ,qIAAqI,CAAA;AAEtI,8DAA8D;AAC9D,MAAM,wBAAwB,GAAG,IAAI,CAAA;AAErC;;GAEG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAiB;IAC7C,IAAI,EAAE,aAAa;IACnB,WAAW,EACV,mJAAmJ;IACpJ,IAAI,EAAE,QAAQ;IACd,OAAO,EAAE;QACR;YACC,IAAI,EAAE,2BAA2B;YACjC,WAAW,EAAE,uFAAuF;SACpG;QACD;YACC,IAAI,EAAE,eAAe;YACrB,WAAW,EAAE,kFAAkF;SAC/F;QACD;YACC,IAAI,EAAE,iBAAiB;YACvB,WAAW,EAAE,qFAAqF;SAClG;KACD;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,WAAW,GAAG,IAAI,CAAC,eAAe,CAAA;QAExC,IAAI,CAAC,WAAW,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CACd,uFAAuF;gBACtF,6GAA6G;gBAC7G,sFAAsF,CACvF,CAAA;QACF,CAAC;QAED,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAU,CAAA;QAElC,IAAI,KAAK,EAAE,MAAM,IAAI,IAAI,eAAe,CAAC,SAAS,CAAC,WAAW,CAAC,EAAE,CAAC;YACjE,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC3B,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,CAAA;YACnB,CAAC;QACF,CAAC;QAED,IAAI,CAAC,QAAQ,CAAC,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,oCAAoC,WAAW,sBAAsB,CAAC,CAAA;QAE1G,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAA;QAClC,MAAM,gBAAgB,GAAG,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAA;QAE9C,uEAAuE;QACvE,MAAM,UAAU,GAAG,IAAI,GAAG,EAAU,CAAA;QAEpC,IAAI,IAAI,GAAG,CAAC,CAAA;QACZ,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,YAAY,GAAG,CAAC,CAAA;QAEpB,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,UAAU,CAAC,IAAI,CAAC,KAAM,CAAC,EAAE,CAAC;YACnD,IAAI,EAAE,CAAA;YACN,MAAM,UAAU,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YACpD,MAAM,QAAQ,GAAG,MAAM,CAAC,KAAK,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAEpD,IAAI,QAAQ,EAAE,CAAC;gBACd,UAAU,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;YACzB,CAAC;YAED,IAAI,CAAC,UAAU,EAAE,CAAC;gBACjB,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC;gBACpC,YAAY,EAAE,CAAA;gBAEd,SAAQ;YACT,CAAC;YAED,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,iBAAiB,CAAC,UAAU,CAAC,CAAA;YAExD,4EAA4E;YAC5E,IAAI,CAAC,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC;gBACxB,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,MAAM,aAAa,GAAG,MAAM,EAAE,GAAG,MAAM,CAAA;YACvC,MAAM,UAAU,GAA2B,EAAE,aAAa,EAAE,MAAM,EAAE,MAAM,EAAE,CAAA;YAC5E,IAAI,GAAG,GAAG,GAAG,MAAM,IAAI,MAAM,EAAE,CAAA;YAC/B,IAAI,KAAK,GAAG,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,aAAa,CAAA;YAE1G,IAAI,aAAa,EAAE,CAAC;gBACnB,MAAM,MAAM,GAAG,MAAM,CAAC,aAAa,EAAE,MAAM,CAAC,CAAA;gBAC5C,MAAM,KAAK,GAAG,MAAM,EAAE,GAAG,wBAAwB,CAAA;gBACjD,MAAM,MAAM,GAAG,MAAM,CAAC,cAAc,EAAE,MAAM,CAAC,CAAA;gBAE7C,MAAM,WAAW,GAAG,KAAK;oBACxB,CAAC,CAAC,MAAM,KAAK,KAAK,IAAI,MAAM,KAAK,KAAK;wBACrC,CAAC,CAAC,GAAG,MAAM,IAAI,MAAM,EAAE;wBACvB,CAAC,CAAC,GAAG,MAAM,GAAG,MAAM,EAAE;oBACvB,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAA;gBAEjB,UAAU,CAAC,YAAY,GAAG,WAAW,CAAA;gBACrC,GAAG,GAAG,GAAG,WAAW,IAAI,MAAM,IAAI,MAAM,EAAE,CAAA;gBAC1C,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,mBAAmB,CAAC,CAAC,CAAC,oBAAoB,CAAA;YAC3D,CAAC;YAED,MAAM,QAAQ,GAAG,cAAc,CAAC,kBAAkB,EAAE,EAAE,GAAG,UAAU,EAAE,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;YAEvF,IACC,aAAa,CACZ,KAAK,EACL;gBACC,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,kBAAkB;gBAC1B,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,+FAA+F;aACxG,EACD,eAAe,KAAK,EAAE,EACtB,sBAAsB,CACtB,EACA,CAAC;gBACF,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,sCAAsC;QAEtC,0EAA0E;QAC1E,MAAM,IAAI,GAAG,CAAC,GAAG,UAAU,CAAC,CAAC,QAAQ,EAAE,CAAA;QACvC,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,gBAAgB,CAAC,CAAC,GAAG,gBAAgB,CAAC,CAAA;QAE3F,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAChD,gEAAgE;YAChE,MAAM,IAAI,GAAG,WAAW,CAAC,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC,CAAA;YAC9C,MAAM,QAAQ,GAAG,cAAc,CAAC,kBAAkB,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,EAAE,CAAC,CAAA;YAEtF,IACC,aAAa,CACZ,KAAK,EACL;gBACC,GAAG,EAAE,IAAI;gBACT,UAAU,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;gBAC9B,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,kBAAkB;gBAC1B,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,kFAAkF;aAC3F,EACD,2BAA2B,EAC3B,sBAAsB,CACtB,EACA,CAAC;gBACF,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,IAAI,YAAY,EAAE,CAAC;YAClB,OAAO,CAAC,KAAK,CAAC,wBAAwB,YAAY,8DAA8D,CAAC,CAAA;QAClH,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAClC,CAAC;CACD,CAAA"}

package/out/fr/recipes/lieudit.d.ts

@@ -3,30 +3,28 @@
  * @license AGPL-3.0
  * @author Teffen Ellis, et al.
  *
- *   `fr-lieudit` slice recipe — FR lieu-dit (hamlet/place) `dependent_locality` coverage
- *   (`.superpowers/sdd/deploc-world-survey.md`, FR section, 2026-07-22). Streams every BAN
- *   `adresses-<dept>.csv` dump under `--ban-dir` through `@mailwoman/ban/sdk`'s
- *   `extractBANAddrPoints`, which now surfaces a cleaned `lieuDit` per record (junk/dup filtering
- *   lives in `ban/sdk/extract.ts`'s `cleanLieuDit`, not duplicated here). Only rows carrying a clean
- *   lieu-dit survive into the pool; the existing `ban`/`synth-fr` sources and their emitted rows are
- *   untouched — this recipe reads the SAME raw CSVs but emits under its own source name.
+ * `fr-lieudit` recipe: FR lieu-dit (hamlet/place) `dependent_locality` coverage.
  *
- *   Mapping: lieu-dit → `dependent_locality`, commune → `locality`. Rendered to match the formatter's
- *   FR `place`-slot convention (`fix(formatter): render dependent_locality for neither-slot templates`,
- *   b1edc1b7, verified via a `formatAddress` smoke call): house+street on line 1, the lieu-dit ALONE on
- *   line 2, postcode+commune on line 3 — French postal convention (La Poste's line 5).
+ * Streams every BAN `adresses-<dept>.csv` dump under `--ban-dir` through `@mailwoman/ban/sdk`'s
+ * `extractBANAddrPoints`, which surfaces a cleaned `lieuDit` per record. Only rows carrying a clean
+ * lieu-dit survive into the pool.
  *
- *   ~1.69M clean rows survive the filter nationally (26M total BAN rows, 1.81M raw `nom_ld` fills, ~6.6%
- *   junk/dup). The pool is read in full (small string tuples only — no coordinates needed) and
- *   Fisher-Yates shuffled with the seeded PRNG before slicing to `--count`, rather than sampled WITH
- *   replacement — at a `--count` a sizeable fraction of the pool, with-replacement draws would produce a
- *   large duplicate rate (birthday-paradox math: ~190k expected collisions at count=800k over a 1.69M
- *   pool).
+ * The recipe maps lieu-dit to `dependent_locality` and commune to `locality`. It renders to match the
+ * formatter's FR `place`-slot convention. House and street go on line 1.
+ * The lieu-dit appears by itself on line 2. Postcode and commune go on line 3.
+ * This follows French postal convention (La Poste's line 5).
+ *
+ * The pool is read in full and Fisher-Yates shuffled with the seeded prng before slicing to
+ * `--count`. With-replacement draws at a large `--count` would produce a large duplicate rate.
+ *
+ * A row's jurisdiction comes from its INSEE commune code, as it does in the `ban` adapter reading
+ * the same files. The BAN fetcher downloads departments 971, 972, 973, 974 and 976 beside the
+ * metropolitan ones, so this recipe emits Guadeloupe, Martinique, French Guiana, Réunion and
+ * Mayotte rows, and labeling them `FR` would mislabel their country and their locale.
  */
 import type { CorpusRecipe } from "#recipes/scaffold";
 /**
- * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
- * and `description` below for the surface form it generates.
+ * Recipe registered with the corpus builder.
  */
 export declare const frLieuditRecipe: CorpusRecipe;
 //# sourceMappingURL=lieudit.d.ts.map

package/out/fr/recipes/lieudit.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"lieudit.d.ts","sourceRoot":"","sources":["../../../lib/fr/recipes/lieudit.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAaH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAA;AAgHrD;;;GAGG;AACH,eAAO,MAAM,eAAe,EAAE,YAwH7B,CAAA"}
+{"version":3,"file":"lieudit.d.ts","sourceRoot":"","sources":["../../../lib/fr/recipes/lieudit.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAiBH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAA;AA8GrD;;GAEG;AACH,eAAO,MAAM,eAAe,EAAE,YA8J7B,CAAA"}