* feat(api): remplace la détection des tech steps par un pipeline NLP (node-nlp)
Le matching par regex ne généralisait jamais au-delà de son propre
vocabulaire — une étape décrivant la fonte du beurre comme "jusqu'à ce
que le beurre ait disparu dans la poêle" ne contient aucun verbe sur
lequel une regex pourrait s'ancrer, alors que le sens est sans
ambiguïté.
Nouveau pipeline en 3 étapes (TechStepClassifierService, node-nlp
4.27.0 — la 5.x est encore alpha, non retenue) :
1. NER (entités enum) trouve les mentions candidates + leur position
exacte, à partir de listes de synonymes (tech-step-training-data.ts)
plutôt que de regex écrites à la main. ner.threshold: 1 (exact,
après normalisation) — le défaut à 0.8 faisait matcher "faire" (verbe
auxiliaire omniprésent) contre "frire" par pure proximité de chaîne.
2. La description est découpée en clauses autour de ces candidats
(splitIntoClauses, pure/testable sans modèle).
3. Le NlpManager classe chaque clause individuellement, entraîné sur
des phrases qui n'emploient jamais le verbe de la technique — c'est
ce qui apporte la compréhension du sens. En dessous de
CONFIDENCE_THRESHOLD (0.65, ajusté empiriquement), retombe sur la
technique impliquée par l'ancre NER plutôt que d'abandonner un match
clairement ancré sur un mot-clé.
TechStepMapping (table de regex par technique/locale) supprimée —
migration 20260821130000_drop_tech_step_mapping — plus aucune table
n'est interrogée à l'exécution, les données de matching vivent en code.
TECH_STEPS (reference-seed-data.ts) simplifié en simple liste de uid,
les mappings ayant disparu.
Deux pièges trouvés en construisant ce pipeline, corrigés à la source :
- db/prisma.ts construisait PrismaClient sans importer config/env.ts —
un run de test isolé pouvait faire gagner la course au .env interne
de Prisma (dev) contre .env.test. Fixé en important config/env.js en
tout premier, pour effet de bord.
- NlpManager a autoSave/autoLoad: true par défaut — persiste le modèle
entraîné dans model.nlp et le recharge au lieu de ré-entraîner au
prochain démarrage. Les deux désactivés explicitement (sinon un
modèle obsolète masquerait silencieusement toute mise à jour du
corpus/seuil) ; model.nlp ajouté au .gitignore en garde-fou.
apps/api/src/db/prisma.ts, recipe.service.ts, sources.service.ts et
recipe-translation.ts adaptés à la matching async (le classifieur
entraîné remplace le couple loadTechStepMappingRules+matchTechStepSpans
synchrone) ; server.ts appelle techStepClassifier.warmUp() avant
d'accepter du trafic (le tout premier appel réel à
NlpManager.process() charge les ressources par langue de node-nlp,
plusieurs secondes).
Vérifié : tsc --noEmit, biome check (0 erreur), build complet des 6
packages, 308 tests API (dont un test-support/reset-db.ts corrigé —
référençait encore tech_step_mapping dans son TRUNCATE).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* feat(api): ajoute la délimitation de contexte aux tech steps et étoffe le vocabulaire du classifieur
Deux évolutions du pipeline NLP de détection des tech steps (PR #63) :
1. Délimitation de contexte — en plus du mot-clé qui déclenche un match
(start/end), chaque TechStepMatch porte maintenant contextStart/
contextEnd : la clause complète autour du mot-clé (ex : "poêle chaude"
comme mot-clé, "Dans une poêle chaude" comme contexte). Persisté sur
StepTechStep (colonnes nullables, migration dédiée), exposé via
StepTechStepView, et rendu côté web avec un style plus discret que le
mot-clé (StepDescription.tsx, .step-tech-step-context). splitIntoClauses
coupe désormais sur l'espace le plus proche du milieu de l'écart entre
deux candidats plutôt que sur le milieu brut, pour ne jamais couper un
mot en deux (findGapSplitPoint).
2. Vocabulaire du classifieur — synonymes et locutions supplémentaires par
technique (FR/EN) pour fiabiliser la détection sur des formulations que
le corpus initial ne couvrait pas. Plusieurs bugs de fond trouvés et
corrigés en cours de route, tous confirmés par la suite de tests
complète (309 tests) :
- un synonyme multi-mots qui est un préfixe-mot d'un synonyme plus court
déjà enregistré pour la même technique fait matcher les deux comme
candidats NER distincts et chevauchants, corrompant le découpage en
clauses (parfois jusqu'à une mauvaise classification) — retiré
partout où ce motif a été repéré (cook, fry, deglaze, simmer, boil,
roast, chop, mince, marinate, preheat, bake, plate, coat) ;
- "poêlé"/"poêlée" comme synonymes de panFry sont réduits à la même
racine que le nom "poêle" par le stemmer français de node-nlp,
provoquant un faux positif sur toute mention nue de "poêle" (dont
celle de preheat) — retiré ;
- "Fouetter les blancs en neige" était mal classé en foldIn (la phrase
d'entraînement de foldIn partage la même locution) — corrigé en
ajoutant des phrases d'entraînement dédiées à whisk ;
- "Émincer les tomates" est passé sous le seuil de confiance vers melt
après l'ajout du nouveau vocabulaire ailleurs dans le corpus — corrigé
en élargissant les phrases d'entraînement de mince à un autre légume.
Le test unitaire de splitIntoClauses avec un point de coupure obsolète
(pré-datant findGapSplitPoint) est aussi corrigé.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(api): corrige la découpe des clauses et le seuil de confiance du classifieur de tech steps
Trouvé en examinant des vraies recettes déjà en base après le dernier
étoffement du vocabulaire : plusieurs étapes bien réelles se faisaient
classer sur la mauvaise technique, sans lien avec un mot-clé manquant.
- splitIntoClauses coupe désormais sur la limite de phrase (juste après
un ".", "!" ou "?") la plus proche du milieu de l'écart entre deux
candidats quand il y en a une, plutôt que sur l'espace brut le plus
proche du milieu. Une description à deux techniques dans deux phrases
distinctes ("Préchauffer le four à 180°C. Dans un saladier, mettre le
beurre... et mélanger.") ne coupait qu'au milieu brut, ce qui pouvait
trancher en pleine deuxième phrase et envoyer au classifieur une
clause tronquée ("...(thermostat 6). Dans un saladier, mettre" sans
complément) — assez éloignée des phrases d'entraînement courtes et
complètes pour se faire mal classer avec confiance (préchauffer prédit
"mix", mélanger prédit "melt").
- CONFIDENCE_THRESHOLD passe de 0.65 à 0.75 : du texte anglais passé
dans le classifieur français (qui doit ne rien trouver, garanti par
le test d'isolation des locales) scorait 0.69 sur "boil" — du bruit
de petit corpus, pas un vrai verdict. Les cas réels que ce seuil sert
à faire confiance scorent 0.91 à 1.0 en pratique ; 0.75 sépare
proprement le bruit du signal sans rien casser (309 tests toujours
verts).
- Deux phrases d'entraînement ajoutées à `cook` pour deux clauses
réelles mal classées (feu doux + remuant, découvert + laisser cuire)
qui n'avaient pourtant pas de mot-clé manquant.
Ajoute aussi src/scripts/backfill-tech-steps.ts : la détection ne
tourne qu'à la création/modification d'une recette, jamais
rétroactivement — ce script recalcule le start/end/contextStart/
contextEnd de chaque étape existante contre le classifieur actuel,
pour ne pas avoir à rouvrir et resauvegarder chaque recette à la main
après un changement de corpus.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(web): rend le contexte des tech steps réellement visible
Le style existant (fond teinté à 6% d'opacité, sans autre indice
visuel) était structurellement correct — vérifié en base, l'API et le
DOM contenaient bien les spans de contexte — mais imperceptible à
l'œil sur ce thème sombre : --color-primary n'est pas assez saturé
pour qu'une teinte de quelques % se distingue du fond de la carte.
Vérifié en créant une recette test dans le navigateur et en zoomant le
texte rendu : littéralement aucune différence visible avant, un
rectangle net après.
Passe à 10% de fond + une bordure basse pleine à 45% d'opacité comme
second indice visuel indépendant, tout en gardant le mot-clé
(soulignement pointillé + fond à 14% + curseur + tooltip) nettement
plus marqué que son contexte.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(web): retire l'affichage visuel du contexte des tech steps
Ne touche que le rendu — le backend continue de calculer et de
persister contextStart/contextEnd (tech-step-matcher.ts, StepTechStep),
et splitDescriptionByTechSteps continue de découper la description
autour du contexte (segments isKeyword: false). StepDescription.tsx
rend désormais ces segments comme du texte brut, comme un segment sans
technique — plus d'encadré/bordure autour de la clause, seul le
mot-clé reste surligné avec sa tooltip.
.step-tech-step-context (CSS) retirée, devenue inutilisée.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* Centralize error handling (shared codes + API/client services), code quality pass
## Error handling
Requested: a centralized error-handling service on the API, custom error
codes shared across apps, and a client-side error service for i18n labels.
- packages/shared/src/errors/error-codes.ts — ErrorCode enum + ApiErrorResponse
contract. Single source of truth: neither side hardcodes a raw error string
the other has to guess at.
- apps/api: HttpError now carries an ErrorCode (not just a message).
ErrorHandlerService (new) centralizes every "how do we turn a thrown error
into an HTTP response" decision — app.ts's error middleware is now a thin
adapter calling into it. API messages reverted to English/dev-facing (they
were French from an earlier pass) since user-facing text is now generated
client-side from the code.
- apps/web: ApiClient (class, singleton instance) throws ApiError carrying
the code. ErrorMessageService (new) maps every ErrorCode to a localized
label, structured with a Locale type from the start (only "fr" exists, but
adding a language later is "add a locale to the map", not "hunt down every
hardcoded string"). LoginPage/SignupPage now display
errorMessageService.getLabel(err.code), never err.message directly.
- Tests strengthened to assert on `code`, not just HTTP status (Mocha +
Cucumber, new "the response error code should be" step). Cypress mocks
updated to the new {code, message} response shape.
## Code quality pass
Per explicit feedback: heavy JSDoc on every interface/type/class/function/
method/member touched in this PR, explicit public/private visibility on
every class member (ApiClient, ErrorMessageService, ErrorHandlerService,
HttpError), no HTML/logic mixing (styling extracted out of components
entirely, never inline).
ApiClient/ErrorMessageService were initially written as static-only classes;
switched to instance-based singletons (matching ErrorHandlerService's
existing pattern) after Biome's noStaticOnlyClass rule flagged the
static-only shape as an anti-pattern — same "class with visibility
modifiers" outcome, without fighting the linter.
## SCSS + theming
- apps/web/src/styles/_theme.scss — design tokens as CSS custom properties
on :root (colors, spacing, typography), not plain Sass variables — makes
them available at runtime, not just compile time, so a future theme
switch (e.g. dark mode) is "redefine these variables" rather than
rebuilding stylesheets.
- apps/web/src/styles/global.scss replaces the old single index.css:
reset + theme import only, loaded once from main.tsx.
- Per-page/component styles colocated (HomePage.tsx + HomePage.scss);
styles shared by multiple pages within one feature live in that feature's
folder (features/auth/auth-form.scss, used by both Login/SignupPage) —
not duplicated per page, not dumped in the global stylesheet either.
- Component-level .scss files intentionally don't `@use` the theme
partial: they only consume CSS custom properties (global at runtime via
global.scss), not Sass-level symbols, so importing it would do nothing —
documented inline rather than left as a silently-redundant import.
- vite.config.ts opts into Sass's modern compiler API to silence a
legacy-js-api deprecation warning on every build.
## specs/ updates
- New specs/error-handling.md — the ErrorCode/ApiErrorResponse contract,
both services, with a flow diagram.
- New specs/frontend-architecture.md — apps/web folder structure, routing/
auth-guard flow, SCSS/theming conventions.
- specs/batch-cooking-architecture.md links to both (original doc content
otherwise untouched — it's the user's own hand-authored source doc).
## Verification
Full lint/mocha/cucumber/build green. Manually re-verified the whole auth
flow in a real browser against native dev servers (not just the automated
suites): signup, the EMAIL_ALREADY_IN_USE → "Cet email est déjà utilisé"
translation end-to-end (confirmed the raw API response carries the English
dev message + code, and the UI shows the French label), wrong-password
INVALID_CREDENTIALS → its label, and confirmed the theme tokens actually
apply (computed button background-color matches --color-primary, card
max-width matches the token value) rather than trusting the build succeeding.
* Address review: no .d.ts, express-tools package, faker fixtures, numeric codes, real i18n lib
Five explicit review points, addressed on this same PR branch (not a new
PR) per updated preference.
## No .d.ts files in the codebase
- apps/web: vite-env.d.ts removed — its /// <reference types="vite/client" />
is replaced by "types": ["vite/client"] in tsconfig.app.json, same effect.
- apps/api: src/types/express.d.ts renamed to express-request.augment.ts —
`declare global` module augmentation works identically in a plain .ts
file as long as it has a top-level import (making it a module); the
.d.ts extension wasn't doing anything for us here.
## packages/express-tools — separate package for Express tooling
Moved HttpError and ErrorHandlerService out of apps/api into a new
workspace package, plus a new createErrorMiddleware() factory (the actual
Express 4-arg error-handling middleware, previously inlined in app.ts).
apps/api now just consumes @batch-cooking/express-tools. Has a real build
(tsc -> dist/, same pattern as packages/shared) — required for the same
reason shared needed one: apps/api's Docker image runs plain `node
dist/server.js`, no tsx. apps/api/Dockerfile updated to COPY the new
package's dist alongside shared's.
## faker.js for test fixtures
apps/api/test/auth.test.ts: replaced the hardcoded "Nicolas
Lefevre"/nicolas@example.com fixture (looked like real user data) with
@faker-js/faker, generated fresh per test via buildSignupPayload().
features/step-definitions/auth.steps.ts: fakerized the filler
firstName/lastName/password used for background state the scenarios
don't actually read.
Deliberately did NOT fakerize the literal example values inside
auth.feature itself (alice@example.com etc.) — those are the readable,
illustrative Gherkin examples that are the whole point of BDD scenarios,
not real PII, and randomizing them would make the scenarios harder to
read for no real gain. Flagged this reasoning in the README in case that
call should go the other way.
Caught a real bug while wiring this up: faker.internet.email() sometimes
capitalizes parts of the address, but signupSchema/loginSchema normalize
emails to lowercase — the test fixture needs to match what's actually
stored, so buildSignupPayload() lowercases the generated email too.
Found by actually running the suite repeatedly, not just once.
## ErrorCode: numeric enum, zero hardcoded values
packages/shared/src/errors/error-codes.ts: ErrorCode is now a numeric
enum (4000 VALIDATION_ERROR, 4001 EMAIL_ALREADY_IN_USE, 4010
INVALID_CREDENTIALS, 4011 NOT_AUTHENTICATED, 4040 NOT_FOUND, 5000
INTERNAL_ERROR — grouped by family like HTTP status codes).
Audited and fixed every place that hardcoded a raw code value instead of
referencing the enum: ApiClient's fallback (`"INTERNAL_ERROR" as
ErrorCode` — would no longer even type-check once the enum went numeric,
which is exactly the point), and the Cypress mock bodies (now import
ErrorCode from @batch-cooking/shared instead of typing the string).
Cucumber's "the response error code should be {string}" step still takes
the *name* in the .feature file (readable: "EMAIL_ALREADY_IN_USE") and
resolves it to the real numeric value via ErrorCode[name] — TypeScript's
reverse enum mapping — before comparing, so the Gherkin stays readable
without the step hardcoding a number either.
## Real i18n library (i18next), not a hand-rolled label map
apps/web: added i18next + react-i18next. New locales/fr/translation.json
holds every user-facing string — not just error labels (errors.*), but
the login/signup/home pages' labels, buttons and headings too
(auth.login.*, auth.signup.*, home.*) — via useTranslation()/t() in each
page. ErrorMessageService no longer owns its own label map; it converts
the numeric ErrorCode to its enum member name and delegates the actual
lookup to i18next (errors.<MEMBER_NAME>). Adding a language is now
"add a locale file", not a code change anywhere.
## specs/ and README updated
specs/error-handling.md and specs/frontend-architecture.md rewritten for
the new package, numeric codes, and i18next. New "i18n" and "no .d.ts"
sections. README covers the same, plus a note on the faker.js scope
decision (feature-file literals excluded, on purpose).
## Verification
Full lint/mocha (x3 runs)/cucumber/build green. Re-verified
express-tools' extraction against a real risk (not just tsc passing):
ran `node dist/server.js` standalone (mirrors the Docker runtime, no
tsx) and hit /health, a 404 (confirmed numeric code 4040 over the wire),
and a real signup + duplicate-email 409 (confirmed numeric 4001). Then
re-verified the full pipeline in a real browser against native dev
servers: signup, EMAIL_ALREADY_IN_USE -> i18next -> "Cet email est déjà
utilisé" end-to-end, home page i18next interpolation
({{firstName}}/{{lastName}}) rendering correctly.
* Address second review round: interface comments, res.locals, ExpressServer, assertIsNever
Five more explicit review points, on the same PR branch.
## Every interface key commented
Audited all 6 interfaces in the codebase. Two had partially-commented
members (violates the "every key gets /** */" rule): AuthResult
(apps/api/auth.service.ts) and SafeUserProfile (packages/shared) — both
now fully commented. The other four (AuthTokenPayload,
AuthContextValue, ErrorHandlingResult, ApiErrorResponse) were already
compliant.
## Removed the Express namespace augmentation
apps/api/src/types/express.d.ts (renamed to express-request.augment.ts
in the last round) is gone entirely. requireAuth now attaches the
authenticated profile to `res.locals.userProfile` — Express's own
built-in per-request mechanism for exactly this — typed via a new
AuthLocals interface and `Response<unknown, AuthLocals>`, instead of a
project-wide `declare global` silently changing every Request's type
whether or not it went through the middleware.
## ErrorHandlerService confirmed framework-agnostic
It already had zero Express import. Documented this explicitly (in the
package's index.ts and the new backend-architecture.md spec) as a
deliberate split: ErrorHandlerService is framework-agnostic (would work
behind Fastify too), ExpressServer/createErrorMiddleware are the actual
Express integration layer.
## packages/express-tools: server init + route/middleware utilities
New ExpressServer class, modeled on the pattern shared as a reference
(adapted, not copied 1:1 — deliberately left out the reference's custom
runtime param-type-validation system, since zod already does that job
in this codebase and running two parallel validation mechanisms would
be redundant, not "propre"):
- setupCore() — the common cors/json/cookie-parser stack
- addRoute() — registers a route, warns+skips instead of silently
double-registering the same method+path
- addMiddleware() / mountRouter() / setErrorHandler()
- listen()
- .instance — the raw Express app, for supertest
Also added wrapAsyncHandler() — forwards a thrown/rejected error from an
async handler to next(err) automatically, removing the manual
try/catch/next(err) every route needed.
apps/api/src/app.ts now builds via ExpressServer (createServer(),
consumed by both server.ts's .listen() and createApp()'s .instance for
tests). auth.routes.ts's signup/login handlers use wrapAsyncHandler
instead of manual try/catch. cookie-parser/cors moved out of apps/api's
own dependencies entirely — they're express-tools' concern now.
## assertIsNever (packages/shared/src/tools/)
Exhaustiveness-check helper for switch/if-chains over a union: takes a
`never`-typed value and throws, so a forgotten case in a later-added
union member becomes a compile error instead of a silent runtime
fallthrough. Verified for real (not just written and assumed correct):
wrote a throwaway switch missing a case and confirmed `tsc` rejects it
with the exact expected error, then deleted the scratch file. No
existing switch/if-chain over a union in the codebase yet to retrofit
it into — noted as ready for when one appears (e.g. the not-yet-built
batch-cooking calculation module or recipe-import pipeline).
## specs/ updated
New specs/backend-architecture.md — ExpressServer, wrapAsyncHandler,
the res.locals decision (with the "why not declare global" reasoning
spelled out), assertIsNever. error-handling.md and
frontend-architecture.md cross-link to it instead of duplicating.
README covers the same, briefly.
## Verification
Full lint/mocha/cucumber/build green. Re-ran `node dist/server.js`
standalone (mirrors Docker, no tsx) after the ExpressServer refactor:
/health, a 404 (numeric 4040), and a real signup + GET /me round trip
confirming res.locals-based auth actually works at runtime, not just
that tsc accepts the types.
* refactor: move ErrorHandlerService/HttpError out of express-tools
ErrorHandlerService has zero dependency on Express — it's a plain
"map an error to {status, body}" service that works identically
behind any HTTP framework. It had no business living in a package
named express-tools.
Extracted HttpError, ErrorHandlerService, and ErrorHandlingResult
into a new packages/error-tools package (same tsc-build-to-dist
pattern as shared/express-tools). express-tools now only keeps the
actual Express-specific layer: ExpressServer, wrapAsyncHandler, and
createErrorMiddleware (which adapts ErrorHandlerService, imported
from error-tools, onto Express).
- packages/error-tools: new package, depends on shared + zod
- packages/express-tools: drops zod dependency, adds error-tools
dependency for error-middleware.ts's type import
- apps/api: adds error-tools dependency; app.ts, auth.service.ts,
require-auth.ts now import HttpError/errorHandlerService from
error-tools instead of express-tools
- apps/api/Dockerfile: adds COPY for packages/error-tools in the
runtime stage
- specs/error-handling.md, specs/backend-architecture.md, README.md
updated to reflect the new package split
Verified: pnpm lint, pnpm build (all packages, correct dependency
order), pnpm test (9/9 Mocha), pnpm test:bdd (5/5 Cucumber), full
Docker rebuild + compose up (no crash-loop), curl + browser checks
of /health, unknown-route 404, signup (201), duplicate-email 409
(code 4001 EMAIL_ALREADY_IN_USE) — all going through the moved
ErrorHandlerService/HttpError correctly.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* Add signup/login (profile creation + JWT auth)
API:
- POST /auth/signup — creates a house + user_profile (transactional),
hashes the password with argon2, sets a JWT in an httpOnly cookie
- POST /auth/login — verifies credentials (generic 401 for both wrong
email and wrong password, doesn't leak which), sets the cookie
- POST /auth/logout — clears the cookie
- GET /auth/me — current profile, behind requireAuth middleware
- requireAuth verifies the JWT and re-checks tokenVersion against the
DB, so a stateless JWT can still be invalidated (password change /
logout-everywhere, not built yet but the field is in place)
Schema: user_profiles gets password_hash + token_version (not in the
original spec doc — required for auth). New migration, with
COMMENT ON for the new columns per the established pattern.
Decisions from the auth planning discussion: JWT in httpOnly cookie
(not server-side sessions), first profile created also creates its
house, argon2 for hashing.
argon2 pinned to 0.31.2 (not ^, deliberately): 0.45.1 segfaults at
runtime on this Windows machine — reproduced consistently across bash
(sandboxed and unsandboxed) and PowerShell, while 0.31.2 works fine
with the same API. Documented in the README as a trap for future
upgrades, since `tsc`/`prisma generate` succeeding doesn't catch a
runtime native-binding crash.
Tests: Mocha (unit-style, apps/api/test/auth.test.ts) and a Cucumber
feature (apps/api/features/auth.feature) covering the full signup →
authenticated flow, duplicate email, wrong password. Both share
test-support/reset-db.ts (TRUNCATE ... CASCADE) to start each
test/scenario from a clean slate. Test-only argon2 cost parameters
(NODE_ENV=test) keep the suite fast — argon2's real cost is
deliberately expensive, which made hashing dozens of times per run
slow and occasionally timeout-flaky at default cost.
CI: added a Postgres service container to lint-and-test (previously
none — tests didn't touch a real DB), runs `prisma migrate deploy`
before the test steps.
Verified end-to-end manually against the dev server (curl): signup,
duplicate email (409), wrong password (401), valid login (200),
validation errors (400), /me with and without cookie, logout (204) —
all behave as intended. Full suite (lint, mocha, cucumber, build) run
multiple times locally with no flakiness after the timeout/cost fixes.
* Fix CI: generate Prisma Client via postinstall
CI failed with "@prisma/client did not initialize yet" — pnpm install
never ran `prisma generate`, and `prisma migrate deploy` (unlike
`migrate dev`) doesn't do it either. Worked locally only because prior
`prisma migrate dev` runs had already generated the client as a side
effect.
Adding a postinstall script fixes it for CI and for anyone cloning the
repo fresh and running plain `pnpm install`.
kyuno053