fix(recipes): recalibre CONFIDENCE_THRESHOLD pour le nouveau classifieur spaCy

0.75 (calibre a l'origine contre node-nlp) laissait de vrais verdicts
corrects sur des clauses sans ancre NER (rien sur quoi retomber) sous le
seuil : melt scorait 0.68 sur "jusqu'a ce que le beurre ait disparu dans
la poele" (le cas motivant tout ce pipeline), preheat 0.52 sur "mettre la
poele sur feu vif" — tous deux corrects, tous deux rejetes a 0.75.

Recalibre a 0.45 : marge confortable au-dessus du bruit (texte anglais
via le classifieur francais score ~0.04, indiscernable du hasard sur ~26
classes) et sous les deux cas ci-dessus. Confirme par
calibrate-tech-step-threshold.ts contre TECH_STEP_EVAL_DATASET (F1
plafonne a 0.987 des 0.45, reste plat jusqu'a 0.95 — 0.45 est deja le
seuil le plus bas qui capture tout le gain disponible).

Retire l'etape CI temporaire de calibration (ci.yml) une fois la valeur
choisie.

Verifie : lint + build complets du monorepo, 27/27 pytest du service,
sweep de seuils + verification manuelle contre le corpus reel en local
(services Python, sans Postgres) et en CI. La suite Mocha complete reste
a confirmer sur ce commit (executee en CI, pas localement — pas de
Postgres disponible dans cet environnement).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

.github/workflows/ci.yml

@@ -98,11 +98,6 @@ jobs:
 
       - run: pnpm install --frozen-lockfile
       - run: pnpm --filter api exec prisma migrate deploy
-      # TEMPORAIRE — recalibration de CONFIDENCE_THRESHOLD (tech-step-matcher.ts)
-      # contre le vrai service Python démarré ci-dessus. À retirer une fois
-      # la valeur choisie et reportée dans le code (voir le plan de migration).
-      - run: pnpm --filter api exec prisma db seed
-      - run: pnpm --filter api exec tsx src/scripts/calibrate-tech-step-threshold.ts
       - run: pnpm --filter api test
 
   intent-service-test:

apps/api/src/lib/recipe-matching/tech-step-matcher.ts

@@ -242,18 +242,31 @@ export function splitIntoClauses(
  * `TECH_STEP_TRAINING_DATA` — see `test/tech-step-matcher.test.ts` for the
  * cases this threshold was picked to pass.
  *
- * Raised from `0.65` after finding real (non-adversarial) misclassified
+ * Recalibrated to `0.45` for the migration off `node-nlp` to
- * clauses that scored just above the old threshold — e.g. English recipe
+ * `services/tech-step-intent-service` (spaCy `textcat`, exclusive classes
- * text run through the French classifier (which must find *nothing*,
+ * over ~26 techniques) — its score distribution is meaningfully different
- * confirmed by `recipe-translation.test.ts`'s own locale-isolation test)
+ * from node-nlp's own classifier, and the old `0.75` (tuned against
- * scored `0.69` for `boil`, essentially classifier noise on
+ * node-nlp) left genuine, correct verdicts on *anchor-less* clauses
- * out-of-vocabulary input rather than a real, confident verdict. The
+ * (nothing to fall back to — see this file's doc comment, point 3) below
- * clauses this threshold exists to actually trust score far higher in
+ * the bar: e.g. `melt` scored `0.68` on "jusqu'à ce que le beurre ait
- * practice (`0.91`–`1.0` for the real corrected cases found this session)
+ * disparu dans la poêle" (the exact motivating no-keyword case this
- * — `0.75` sits comfortably above the noise floor and below every genuine
+ * pipeline exists for), `preheat` scored `0.52` on "mettre la poêle sur
- * match seen so far.
+ * feu vif" — both the classifier's own confident best guess, both correct,
+ * both wrongly discarded at `0.75`. `0.45` clears both with margin while
+ * staying far above the noise floor: English recipe text run through the
+ * French classifier (must find *nothing*, confirmed by
+ * `recipe-translation.test.ts`'s own locale-isolation test) scored `0.04`
+ * for every technique — indistinguishable from the ~1/26 uniform baseline
+ * over this many exclusive classes, nowhere near `0.45`. Cross-checked
+ * against `apps/api/src/scripts/calibrate-tech-step-threshold.ts`'s sweep
+ * over `TECH_STEP_EVAL_DATASET`: aggregate F1 climbs to its plateau
+ * (`0.987`) starting exactly at `0.45` and stays flat through `0.95`, so
+ * this is the lowest threshold that already captures every gain available
+ * from trusting the classifier more — a higher value would only ever
+ * discard more anchor-less true positives like the two above, never buy
+ * back any precision.
  */
-export const CONFIDENCE_THRESHOLD = 0.75;
+export const CONFIDENCE_THRESHOLD = 0.45;
 
 /**
  * One clause's full classification detail — the finer-grained sibling of