@mailwoman/corpus 9.3.0 → 10.0.0

package/out/fr/recipes/fragment.js

@@ -0,0 +1,277 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-fragment` slice recipe (#727 T2) — the HOUSE-NUMBER-LICENCE change.
+ *
+ *   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.
+ */
+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";
+/**
+ * House numbers, weighted toward the small values that dominate real BAN rows.
+ */
+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.
+ */
+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.
+const norm = (value) => value
+    .normalize("NFD")
+    .replaceAll(/[̀-ͯ]/g, "")
+    .toLowerCase()
+    .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.
+ */
+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)?
+ */
+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`).
+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.
+ */
+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",
+    mode: "tuples",
+    options: [
+        {
+            flag: "--exclude-surfaces <path>",
+            description: "REQUIRED. The fragment board's reserved surface list; every listed street is skipped.",
+        },
+        {
+            flag: "--hn-prob <n>",
+            description: "Share of rows carrying a house number (default 0.35 — the anchor, not the point)",
+        },
+        {
+            flag: "--bare-prob <n>",
+            description: "Share of NO-house-number rows that are bare LOCALITIES (default 0.25 — the counter)",
+        },
+    ],
+    async run(opts, write) {
+        const random = makeMulberry32(opts.seed);
+        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. " +
+                "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 (!excluded.size)
+            throw new Error(`fr-fragment: --exclude-surfaces "${excludePath}" listed no surfaces`);
+        const hnProb = opts.hnProb ?? 0.35;
+        const bareLocalityProb = opts.bareProb ?? 0.25;
+        // The locality pool is harvested from the tuples themselves — every BAN row carries its
+        // commune, so the counter-distribution needs no second source.
+        const localities = new Set();
+        let read = 0;
+        let emitted = 0;
+        let skipped = 0;
+        let contaminated = 0;
+        for await (const tuple of readTuples(opts.input)) {
+            read++;
+            const fullStreet = String(tuple.street ?? "").trim();
+            const locality = String(tuple.locality ?? "").trim();
+            if (locality) {
+                localities.add(locality);
+            }
+            if (!fullStreet) {
+                skipped++;
+                continue;
+            }
+            // THE SPLIT. A surface on the fragment board never enters training.
+            if (excluded.has(norm(fullStreet))) {
+                contaminated++;
+                continue;
+            }
+            const { prefix, street } = decomposeFrStreet(fullStreet);
+            // The failing class is the designator-led street. A no-prefix nom_voie ("La Ville Mois")
+            // is a different problem and would muddy the signal.
+            if (!prefix || !street) {
+                skipped++;
+                continue;
+            }
+            const carriesNumber = random() < hnProb;
+            const components = { street_prefix: prefix, street };
+            let raw = `${prefix} ${street}`;
+            let klass = DATEISH.test(street) ? "date-name" : PARTICLE.test(street) ? "street-particle" : "bare-street";
+            if (carriesNumber) {
+                const number = pick(HOUSE_NUMBERS, random);
+                const alnum = random() < ALNUM_HOUSE_NUMBER_SHARE;
+                const suffix = pick(ALNUM_SUFFIXES, random);
+                const houseNumber = alnum
+                    ? suffix === "bis" || suffix === "ter"
+                        ? `${number} ${suffix}`
+                        : `${number}${suffix}`
+                    : String(number);
+                components.house_number = houseNumber;
+                raw = `${houseNumber} ${prefix} ${street}`;
+                klass = alnum ? "alnum-housenumber" : "street-housenumber";
+            }
+            const sourceID = sliceSourceID("synth-fr-fragment", { ...components, v: String(read) });
+            if (alignAndWrite(write, {
+                raw,
+                components,
+                country: "FR",
+                locale: "fr-FR",
+                source: "synth-fr-fragment",
+                source_id: sourceID,
+                corpus_version: "0.9.4",
+                license: "Synthetic — fr-fragment; (street, commune) from BAN (Base Adresse Nationale, Licence Ouverte)",
+            }, `fr-fragment:${klass}`)) {
+                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.
+        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}` });
+            if (alignAndWrite(write, {
+                raw: name,
+                components: { locality: name },
+                country: "FR",
+                locale: "fr-FR",
+                source: "synth-fr-fragment",
+                source_id: sourceID,
+                corpus_version: "0.9.4",
+                license: "Synthetic — fr-fragment counter-distribution; commune from BAN (Licence Ouverte)",
+            }, "fr-fragment:bare-locality")) {
+                emitted++;
+            }
+            else {
+                skipped++;
+            }
+        }
+        if (contaminated) {
+            console.error(`fr-fragment: skipped ${contaminated} rows whose street surface is reserved by the fragment board`);
+        }
+        return { read, emitted, skipped };
+    },
+};
+//# sourceMappingURL=fragment.js.map

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

@@ -0,0 +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"}

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

@@ -0,0 +1,32 @@
+/**
+ * @copyright Sister Software
+ * @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.
+ *
+ *   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).
+ *
+ *   ~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).
+ */
+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.
+ */
+export declare const frLieuditRecipe: CorpusRecipe;
+//# sourceMappingURL=lieudit.d.ts.map

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

@@ -0,0 +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"}

package/out/fr/recipes/lieudit.js

@@ -0,0 +1,211 @@
+/**
+ * @copyright Sister Software
+ * @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.
+ *
+ *   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).
+ *
+ *   ~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).
+ */
+import { extractBANAddrPoints } from "@mailwoman/ban/sdk";
+import { formatAddress } from "@mailwoman/codex/address-format";
+import { COUNTRY_SURFACE_FORMS } from "@mailwoman/codex/country";
+import { dataRootPath } from "@mailwoman/core/data-root";
+import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils";
+import { join } from "path-ts";
+import { Globerator } from "spliterator/node/fs";
+import { stableSourceID } from "#adapters/utils";
+import { decomposeFrStreet } from "#fr/adapters/ban/street-decompose";
+import { pick } from "#synthesizers/utils";
+import { alignRow } from "#utils";
+/**
+ * Matches the `ban` adapter's Tier-B election for BAN data.
+ */
+const DEFAULT_LICENSE = "Licence Ouverte 2.0";
+/**
+ * Enumerate `adresses-<dept>.csv[.gz]` files in `banDir`, ONE path per département. Excludes the `merged`/`france`
+ * aggregates (they duplicate the per-département rows) and, when both a `.csv` and a `.csv.gz` exist for the same dept
+ * (observed on disk for 13/2A/48/69/75 — a stale re-fetch artifact), prefers the uncompressed `.csv` — mirrors
+ * `ban/scripts/build-address-point-slice.ts`'s `departementFiles`, which hit and fixed this exact double-count trap
+ * first.
+ */
+async function departementFiles(banDir) {
+    const byDept = new Map();
+    for (const name of await Globerator.from("*", { cwd: banDir, absolute: false }).toSorted()) {
+        const m = /^adresses-(.+?)\.csv(\.gz)?$/.exec(name);
+        if (!m)
+            continue;
+        const dept = m[1];
+        if (dept === "merged" || dept === "france")
+            continue;
+        const existing = byDept.get(dept);
+        if (!existing || (existing.endsWith(".gz") && !name.endsWith(".gz"))) {
+            byDept.set(dept, join(banDir, name));
+        }
+    }
+    return [...byDept.keys()].toSorted().map((dept) => byDept.get(dept));
+}
+/**
+ * Stream every département file, keeping only rows with a clean `lieuDit` (junk/dup filtering lives in `ban/sdk`).
+ */
+async function readLieuDitPool(banDir) {
+    const files = await departementFiles(banDir);
+    if (!files.length) {
+        throw new Error(`No BAN adresses-<dept>.csv files found in ${banDir} — fetch BAN first (\`mailwoman corpus fetch ban\`).`);
+    }
+    const pool = [];
+    let scanned = 0;
+    for (const path of files) {
+        let deptCount = 0;
+        for await (const rec of extractBANAddrPoints(path)) {
+            scanned++;
+            if (!rec.lieuDit || !rec.city)
+                continue;
+            pool.push({
+                numero: rec.numero,
+                rep: rec.rep,
+                street: rec.street,
+                postcode: rec.postcode,
+                locality: rec.city,
+                dependentLocality: rec.lieuDit,
+            });
+            deptCount++;
+        }
+        console.error(`  ${path}: ${deptCount.toLocaleString()} clean lieu-dit rows`);
+    }
+    console.error(`  scanned ${scanned.toLocaleString()} BAN rows across ${files.length} départements → pool=${pool.length.toLocaleString()}`);
+    return pool;
+}
+/**
+ * `house_number` = `numero` + folded `rep` ("10 bis"), matching the `ban` adapter's own composition.
+ */
+function composeHouseNumber(numero, rep) {
+    return rep ? `${numero} ${rep}` : numero;
+}
+/**
+ * Fisher-Yates shuffle, in place, with the recipe's seeded PRNG — reproducible sampling without replacement.
+ */
+function shuffleInPlace(arr, random) {
+    for (let i = arr.length - 1; i > 0; i--) {
+        const j = Math.floor(random() * (i + 1));
+        [arr[i], arr[j]] = [arr[j], arr[i]];
+    }
+}
+/**
+ * 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.
+ */
+export const frLieuditRecipe = {
+    name: "fr-lieudit",
+    description: "FR lieu-dit rows: BAN nom_ld → dependent_locality (commune → locality), lieu-dit on its own line",
+    mode: "generate",
+    options: [
+        {
+            flag: "--ban-dir <dir>",
+            description: "BAN adresses-<dept>.csv directory. Default $MAILWOMAN_DATA_ROOT/corpus/sources/ban",
+        },
+        {
+            flag: "--country-fraction <f>",
+            description: "Fraction of rows that append an explicit 'France' surface form + a `country` component. Default 0",
+        },
+    ],
+    async run(opts, write) {
+        const random = makeMulberry32(opts.seed);
+        const source = opts.sourceName ?? "synth-fr-lieudit";
+        const count = opts.count ?? 800_000;
+        const banDir = opts.banDir ?? dataRootPath("corpus", "sources", "ban");
+        const countryFraction = opts.countryFraction ?? 0;
+        if (!(countryFraction >= 0 && countryFraction <= 1)) {
+            throw new Error(`--country-fraction must be in [0, 1], got ${countryFraction}`);
+        }
+        const pool = await readLieuDitPool(banDir);
+        if (!pool.length) {
+            throw new Error(`No clean lieu-dit rows found under ${banDir} — see ban/sdk/extract.ts's cleanLieuDit filter.`);
+        }
+        shuffleInPlace(pool, random);
+        const selected = pool.slice(0, Math.min(count, pool.length));
+        let emitted = 0;
+        let skipped = 0;
+        let countryAppended = 0;
+        for (const t of selected) {
+            const house = composeHouseNumber(t.numero, t.rep);
+            const decomposed = decomposeFrStreet(t.street);
+            const components = {
+                house_number: house,
+                dependent_locality: t.dependentLocality,
+                locality: t.locality,
+            };
+            if (decomposed.prefix) {
+                components.street_prefix = decomposed.prefix;
+            }
+            if (decomposed.street) {
+                components.street = decomposed.street;
+            }
+            if (t.postcode) {
+                components.postcode = t.postcode;
+            }
+            // The envelope form: house+street line, the lieu-dit ALONE on its own line, postcode+commune line. That is
+            // La Poste's line 5, and it is what `FR`'s layout prints — this recipe used to restate it.
+            let raw = formatAddress(components, "FR");
+            if (!raw) {
+                skipped++;
+                continue;
+            }
+            // Country-append (the fr-admin-split #728 pattern, generalized): ~`countryFraction` of the time
+            // append an explicit "France" surface form onto the trailing (postcode+commune) line + a
+            // `country` component — the model relearns to emit country WHEN present without over-firing it
+            // on the (still-majority) country-less rows. `countryFraction <= 0` (the default) never draws
+            // from `random`, so the byte-stream is unaffected when the flag is unset.
+            if (countryFraction > 0 && random() < countryFraction) {
+                const forms = COUNTRY_SURFACE_FORMS.FR;
+                const form = pick(forms, random);
+                raw = `${raw}, ${form}`;
+                components.country = form;
+                countryAppended++;
+            }
+            const sourceID = stableSourceID(source, {
+                street: t.street,
+                house_number: house,
+                dependent_locality: t.dependentLocality,
+                locality: t.locality,
+                postcode: t.postcode ?? undefined,
+            });
+            const canonical = {
+                raw,
+                components,
+                country: "FR",
+                locale: "fr-FR",
+                source,
+                source_id: sourceID,
+                corpus_version: "",
+                license: DEFAULT_LICENSE,
+            };
+            const aligned = alignRow(canonical);
+            if (aligned.kind !== "labeled" || !aligned.row) {
+                skipped++;
+                continue;
+            }
+            write(JSON.stringify({ ...aligned.row, synth_method: source, synth_base_id: null }) + "\n");
+            emitted++;
+        }
+        console.error(`  emitted=${emitted} skipped=${skipped} country-appended=${countryAppended} pool=${pool.length}`);
+        return { emitted, skipped };
+    },
+};
+//# sourceMappingURL=lieudit.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"lieudit.js","sourceRoot":"","sources":["../../../lib/fr/recipes/lieudit.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAA;AACzD,OAAO,EAAE,aAAa,EAAE,MAAM,iCAAiC,CAAA;AAE/D,OAAO,EAAE,qBAAqB,EAAE,MAAM,0BAA0B,CAAA;AAChE,OAAO,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAA;AACxD,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AACpE,OAAO,EAAE,IAAI,EAAwB,MAAM,SAAS,CAAA;AACpD,OAAO,EAAE,UAAU,EAAE,MAAM,qBAAqB,CAAA;AAEhD,OAAO,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAA;AAChD,OAAO,EAAE,iBAAiB,EAAE,MAAM,mCAAmC,CAAA;AAErE,OAAO,EAAE,IAAI,EAAE,MAAM,qBAAqB,CAAA;AAE1C,OAAO,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAA;AACjC;;GAEG;AACH,MAAM,eAAe,GAAG,qBAAqB,CAAA;AAc7C;;;;;;GAMG;AACH,KAAK,UAAU,gBAAgB,CAAC,MAAuB;IACtD,MAAM,MAAM,GAAG,IAAI,GAAG,EAAkB,CAAA;IAExC,KAAK,MAAM,IAAI,IAAI,MAAM,UAAU,CAAC,IAAI,CAAC,GAAG,EAAE,EAAE,GAAG,EAAE,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC;QAC5F,MAAM,CAAC,GAAG,8BAA8B,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QAEnD,IAAI,CAAC,CAAC;YAAE,SAAQ;QAEhB,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,CAAE,CAAA;QAElB,IAAI,IAAI,KAAK,QAAQ,IAAI,IAAI,KAAK,QAAQ;YAAE,SAAQ;QAEpD,MAAM,QAAQ,GAAG,MAAM,CAAC,GAAG,CAAC,IAAI,CAAC,CAAA;QAEjC,IAAI,CAAC,QAAQ,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC;YACtE,MAAM,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC,CAAA;QACrC,CAAC;IACF,CAAC;IAED,OAAO,CAAC,GAAG,MAAM,CAAC,IAAI,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,GAAG,CAAC,IAAI,CAAE,CAAC,CAAA;AACtE,CAAC;AAED;;GAEG;AACH,KAAK,UAAU,eAAe,CAAC,MAAuB;IACrD,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAAC,MAAM,CAAC,CAAA;IAE5C,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC;QACnB,MAAM,IAAI,KAAK,CACd,6CAA6C,MAAM,sDAAsD,CACzG,CAAA;IACF,CAAC;IAED,MAAM,IAAI,GAAmB,EAAE,CAAA;IAC/B,IAAI,OAAO,GAAG,CAAC,CAAA;IAEf,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QAC1B,IAAI,SAAS,GAAG,CAAC,CAAA;QAEjB,IAAI,KAAK,EAAE,MAAM,GAAG,IAAI,oBAAoB,CAAC,IAAI,CAAC,EAAE,CAAC;YACpD,OAAO,EAAE,CAAA;YAET,IAAI,CAAC,GAAG,CAAC,OAAO,IAAI,CAAC,GAAG,CAAC,IAAI;gBAAE,SAAQ;YAEvC,IAAI,CAAC,IAAI,CAAC;gBACT,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,GAAG,EAAE,GAAG,CAAC,GAAG;gBACZ,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,QAAQ,EAAE,GAAG,CAAC,QAAQ;gBACtB,QAAQ,EAAE,GAAG,CAAC,IAAI;gBAClB,iBAAiB,EAAE,GAAG,CAAC,OAAO;aAC9B,CAAC,CAAA;YAEF,SAAS,EAAE,CAAA;QACZ,CAAC;QAED,OAAO,CAAC,KAAK,CAAC,KAAK,IAAI,KAAK,SAAS,CAAC,cAAc,EAAE,sBAAsB,CAAC,CAAA;IAC9E,CAAC;IAED,OAAO,CAAC,KAAK,CACZ,aAAa,OAAO,CAAC,cAAc,EAAE,oBAAoB,KAAK,CAAC,MAAM,wBAAwB,IAAI,CAAC,MAAM,CAAC,cAAc,EAAE,EAAE,CAC3H,CAAA;IAED,OAAO,IAAI,CAAA;AACZ,CAAC;AAED;;GAEG;AACH,SAAS,kBAAkB,CAAC,MAAc,EAAE,GAAkB;IAC7D,OAAO,GAAG,CAAC,CAAC,CAAC,GAAG,MAAM,IAAI,GAAG,EAAE,CAAC,CAAC,CAAC,MAAM,CAAA;AACzC,CAAC;AAED;;GAEG;AACH,SAAS,cAAc,CAAI,GAAQ,EAAE,MAAoB;IACxD,KAAK,IAAI,CAAC,GAAG,GAAG,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzC,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CACvC;QAAA,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAE,EAAE,GAAG,CAAC,CAAC,CAAE,CAAC,CAAA;IACvC,CAAC;AACF,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,MAAM,eAAe,GAAiB;IAC5C,IAAI,EAAE,YAAY;IAClB,WAAW,EAAE,kGAAkG;IAC/G,IAAI,EAAE,UAAU;IAChB,OAAO,EAAE;QACR;YACC,IAAI,EAAE,iBAAiB;YACvB,WAAW,EAAE,oFAAoF;SACjG;QACD;YACC,IAAI,EAAE,wBAAwB;YAC9B,WAAW,EAAE,mGAAmG;SAChH;KACD;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,MAAM,GAAG,IAAI,CAAC,UAAU,IAAI,kBAAkB,CAAA;QACpD,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,OAAO,CAAA;QACnC,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,YAAY,CAAC,QAAQ,EAAE,SAAS,EAAE,KAAK,CAAC,CAAA;QACtE,MAAM,eAAe,GAAG,IAAI,CAAC,eAAe,IAAI,CAAC,CAAA;QAEjD,IAAI,CAAC,CAAC,eAAe,IAAI,CAAC,IAAI,eAAe,IAAI,CAAC,CAAC,EAAE,CAAC;YACrD,MAAM,IAAI,KAAK,CAAC,6CAA6C,eAAe,EAAE,CAAC,CAAA;QAChF,CAAC;QAED,MAAM,IAAI,GAAG,MAAM,eAAe,CAAC,MAAM,CAAC,CAAA;QAE1C,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CAAC,sCAAsC,MAAM,kDAAkD,CAAC,CAAA;QAChH,CAAC;QAED,cAAc,CAAC,IAAI,EAAE,MAAM,CAAC,CAAA;QAC5B,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC,CAAA;QAE5D,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,eAAe,GAAG,CAAC,CAAA;QAEvB,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;YAC1B,MAAM,KAAK,GAAG,kBAAkB,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,GAAG,CAAC,CAAA;YACjD,MAAM,UAAU,GAAG,iBAAiB,CAAC,CAAC,CAAC,MAAM,CAAC,CAAA;YAE9C,MAAM,UAAU,GAA0C;gBACzD,YAAY,EAAE,KAAK;gBACnB,kBAAkB,EAAE,CAAC,CAAC,iBAAiB;gBACvC,QAAQ,EAAE,CAAC,CAAC,QAAQ;aACpB,CAAA;YAED,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC;gBACvB,UAAU,CAAC,aAAa,GAAG,UAAU,CAAC,MAAM,CAAA;YAC7C,CAAC;YAED,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC;gBACvB,UAAU,CAAC,MAAM,GAAG,UAAU,CAAC,MAAM,CAAA;YACtC,CAAC;YAED,IAAI,CAAC,CAAC,QAAQ,EAAE,CAAC;gBAChB,UAAU,CAAC,QAAQ,GAAG,CAAC,CAAC,QAAQ,CAAA;YACjC,CAAC;YAED,2GAA2G;YAC3G,2FAA2F;YAC3F,IAAI,GAAG,GAAG,aAAa,CAAC,UAAU,EAAE,IAAI,CAAC,CAAA;YAEzC,IAAI,CAAC,GAAG,EAAE,CAAC;gBACV,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,gGAAgG;YAChG,yFAAyF;YACzF,+FAA+F;YAC/F,8FAA8F;YAC9F,0EAA0E;YAC1E,IAAI,eAAe,GAAG,CAAC,IAAI,MAAM,EAAE,GAAG,eAAe,EAAE,CAAC;gBACvD,MAAM,KAAK,GAAG,qBAAqB,CAAC,EAAE,CAAA;gBACtC,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,EAAE,MAAM,CAAC,CAAA;gBAChC,GAAG,GAAG,GAAG,GAAG,KAAK,IAAI,EAAE,CAAA;gBACvB,UAAU,CAAC,OAAO,GAAG,IAAI,CAAA;gBAEzB,eAAe,EAAE,CAAA;YAClB,CAAC;YAED,MAAM,QAAQ,GAAG,cAAc,CAAC,MAAM,EAAE;gBACvC,MAAM,EAAE,CAAC,CAAC,MAAM;gBAChB,YAAY,EAAE,KAAK;gBACnB,kBAAkB,EAAE,CAAC,CAAC,iBAAiB;gBACvC,QAAQ,EAAE,CAAC,CAAC,QAAQ;gBACpB,QAAQ,EAAE,CAAC,CAAC,QAAQ,IAAI,SAAS;aACjC,CAAC,CAAA;YAEF,MAAM,SAAS,GAAiB;gBAC/B,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM;gBACN,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,EAAE;gBAClB,OAAO,EAAE,eAAe;aACxB,CAAA;YAED,MAAM,OAAO,GAAG,QAAQ,CAAC,SAAS,CAAC,CAAA;YAEnC,IAAI,OAAO,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,GAAG,EAAE,CAAC;gBAChD,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,OAAO,CAAC,GAAG,EAAE,YAAY,EAAE,MAAM,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,GAAG,IAAI,CAAC,CAAA;YAE3F,OAAO,EAAE,CAAA;QACV,CAAC;QAED,OAAO,CAAC,KAAK,CAAC,aAAa,OAAO,YAAY,OAAO,qBAAqB,eAAe,SAAS,IAAI,CAAC,MAAM,EAAE,CAAC,CAAA;QAEhH,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAC5B,CAAC;CACD,CAAA"}

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

@@ -0,0 +1,38 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-order` slice recipe — French reversed-order coverage (#560). Reads REAL OpenAddresses FR
+ *   tuples (`fr/countrywide.csv` from the cached zip), then for each row picks — by
+ *   `--reversed-fraction` (default 0.5) — whether to render in CANONICAL French order
+ *   (number-street, postcode-city) or one of four REVERSED / postcode-first variants the
+ *   v4.4.0→v0.5.0 regression exposed (the model misses house_number in every reversed one):
+ *
+ *   - A: "47110 Sainte-Livrade-sur-Lot, 69 Allée du Bugatel" (postcode city, HN street)
+ *   - B: "Sainte-Livrade-sur-Lot, 47110, 619 Impasse de la Rose" (city, postcode, HN street)
+ *   - C: "Sainte-Livrade-sur-Lot 59 bis Rue des Ecuries 47110" (city HN street postcode — NO commas)
+ *   - D: "47110, 6 rue de la république, Sainte-Livrade-sur-Lot" (postcode, HN street, city)
+ *
+ *   Sub-modes ride alongside order: `bis`/`ter`/`quater` ordinal suffixes in house_number, and
+ *   ALL-CAPS locality. `--golden` emits a held-out reversed-order eval slice with a different
+ *   seed.
+ *
+ *   The inline synthesis (the OA-CSV reader, the ordinal/all-caps tables, the canonical + reversed
+ *   renderers) is ported faithfully from scripts/build-fr-order-slice.mjs. This is a
+ *   `generate`-mode recipe that still reads REAL tuples off disk — `--count` bounds the OUTPUT, not
+ *   the input. The passed `random` (the framework LCG) is consumed in the exact call order the
+ *   legacy script used.
+ *
+ *   NOT ported (diagnostic-only, no effect on emitted bytes): the post-run `runSpanCheck` self-check
+ *   (it reads the finished file back with a separate PRNG and prints to stderr; the recipe's output
+ *   stream is still open during `run`), and the dead `renderReversed` helper (the legacy `main`
+ *   inlined the variant logic and never called it).
+ */
+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.
+ */
+export declare const frOrderRecipe: CorpusRecipe;
+//# sourceMappingURL=order.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"order.d.ts","sourceRoot":"","sources":["../../../lib/fr/recipes/order.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAQH,OAAO,EAAgB,KAAK,YAAY,EAAE,MAAM,mBAAmB,CAAA;AA6EnE;;;GAGG;AACH,eAAO,MAAM,aAAa,EAAE,YA8H3B,CAAA"}