chore: sync upstream/master + bump @supabase/pg to 8.21.1 (#34)

* test: Ensure failure to throw at all doesn’t pass (brianc#3671)

* Assorted test fixes and cleanup (brianc#3672)

* cleanup: Remove duplicate test

* cleanup: Remove nonsense test

* cleanup: Simplify promise rejection test

* test: Fix and tighten assertion that would always pass

because of the `SELECTR` typo.

* cleanup: Add missing `await`s when using `assert.rejects` in tests; remove unneeded function wrappers

* ci: Node 26 followup (brianc#3670)

* Revert unneeded pg-native→libpq dependency range adjustment

This reverts part of commit 1025d12.

* dev: Upgrade libpq/nan in lockfile for Node 26 compatibility

* fix: apply SASLprep (RFC 4013) to passwords before SCRAM-SHA-256 PBKDF2 (brianc#3669)

* fix: apply SASLprep (RFC 4013) to passwords before SCRAM-SHA-256 PBKDF2

`pg`'s SCRAM-SHA-256 client passes the raw password into PBKDF2 with no
normalization, while PostgreSQL's server (and libpq) apply SASLprep
(B.1 mapping -> NFKC -> prohibition + bidi check) when computing the
stored verifier. Passwords whose NFKC form differs from themselves
(e.g. containing U+00A8 dieresis, U+2011 non-breaking hyphen, U+00BC
vulgar one quarter, NBSP, soft hyphen) authenticate with psql/libpq
but fail against pg with `28P01`.

Wire `@mongodb-js/saslprep` (the maintained fork used by mongodb's
official Node driver) into `continueSession` before `crypto.deriveKey`,
with a try/catch fallback to the raw password on prohibited / bidi
violations to match `libpq`'s `pg_saslprep` behavior.

Also adds:

- Unit tests covering the soft-hyphen B.1 mapping equivalence, the
  Roman-numeral-IX NFKC asymmetry, the prohibited-char fallback, and a
  deterministic snapshot for the original bug-report password.
- A gated integration test block (SCRAM_TEST_PGUSER_UNICODE /
  SCRAM_TEST_PGPASSWORD_UNICODE) covering raw + NFKC-equivalent + wrong
  password.
- A `scram_unicode_test` role (password `U&'IX-\2168'`) provisioned in
  CI plus matching env vars so the new integration tests run on every
  Node version.
- A Cloudflare Workers regression guard that exercises
  `sasl.continueSession` to ensure `@mongodb-js/saslprep` resolves
  cleanly under workerd.
- A `pg@8.21.0` CHANGELOG entry.

* fix: inline SASLprep, drop @mongodb-js/saslprep dependency

Per review feedback on brianc#3669: ship the SASLprep step as a small in-tree
function instead of pulling a runtime dep with an unpinned transitive.
The function performs only the three byte-changing steps from RFC 4013
(Table C.1.2 -> SPACE, Table B.1 -> empty, NFKC) and skips the
prohibition (RFC 4013 section 2.3) and bidi (RFC 3454 section 6) checks,
since libpq is forgiving on those paths and Postgres's own SASLprep is
similarly lenient. Removes the try/catch fallback (no code path throws).

The deterministic snapshot tests stay byte-for-byte valid because none
of them touch U+200B, the only edge case where the inline impl diverges
from `@mongodb-js/saslprep`. RFC 3454 places U+200B in Table B.1
(mapped to nothing); the dep maps it to SPACE. PostgreSQL's
saslprep.c follows the RFC, so the inline impl matches libpq more
closely on that codepoint. The B.1 unit-test rename ("passes ASCII
control characters through normalization unchanged") keeps the same
snapshot bytes since BEL is unchanged by all three steps.

Co-authored-by: charmander <charmander@noreply.github.com>

* Revert unrelated no-op changes to yarn.lock

now that the associated dependency isn’t being added.

* cleanup: Allow Prettier to format some lines

* cleanup: Remove changelog entry for unreleased pg version

normally added as part of the release process

* refactor: Simplify comments in sasl.js and remove unused test cases

Updated comments in sasl.js to clarify the password normalization process and removed redundant test cases from vitest-cf.test.ts, streamlining the codebase.

* Remove redundant NFKC-only SASLprep test

Confirmed in pull request comments that the “macOS/iOS” thing was an AI inventing an unneeded justification, and NFKC is already covered by another test.

* fix: SASLprep zero-width space the same way PostgreSQL does

As mentioned in the test comment, RFC 3454 defines appendix B for mapping tables and appendix C for prohibition tables. RFC 4013 SASLprep is probably misusing that list of non-ASCII spaces, and says nothing about the overlap. (At least it’s obsoleted.)

* cleanup: Simplify regex character classes with ranges

---------

Co-authored-by: charmander <charmander@noreply.github.com>
Co-authored-by: Charmander <~@charmander.me>

---------

Co-authored-by: Charmander <~@charmander.me>
Co-authored-by: charmander <charmander@noreply.github.com>

Author: 3 people

.gitignore

@@ -10,3 +10,4 @@ dist
 /.eslintcache
 .vscode/
 manually-test-on-heroku.js
+*.tsbuildinfo

packages/pg-native/package.json

@@ -34,7 +34,7 @@
   },
   "homepage": "https://github.com/supabase/node-postgres/tree/master/packages/pg-native",
   "dependencies": {
-    "libpq": "^1.11.0",
+    "libpq": "^1.8.15",
     "pg-types": "2.2.0"
   },
   "devDependencies": {

packages/pg/lib/crypto/sasl.js

@@ -2,6 +2,34 @@
 const crypto = require('./utils')
 const { signatureAlgorithmHashFromCertificate } = require('./cert-signatures')
 
+// SASLprep (RFC 4013) — minimal in-tree implementation.
+//
+// Per RFC 5802 §2.2, the SCRAM-SHA-256 client must normalize the password via
+// SASLprep before feeding it into PBKDF2. PostgreSQL's server applies the same
+// SASLprep when computing the stored verifier, and libpq does the same client
+// side, so passwords whose NFKC form differs from the raw form
+// would otherwise authenticate against psql/libpq but fail against pg with `28P01`.
+//
+// We deliberately implement only the three steps that change the byte content:
+//   1. RFC 3454 Table C.1.2 (non-ASCII space) → U+0020 SPACE.
+//   2. RFC 3454 Table B.1 (commonly mapped to nothing) → empty.
+//   3. NFKC normalization.
+// We skip the prohibition (RFC 4013 §2.3) and bidi (RFC 3454 §6) checks.
+// libpq is forgiving on those paths and Postgres's own SASLprep matches that
+// leniency for legacy roles, so omitting the rejection logic keeps existing
+// roles working without adding complexity.
+function saslprep(password) {
+  // RFC 3454 Table C.1.2 — non-ASCII space characters, mapped to U+0020.
+  const nonAsciiSpace = /[\u00A0\u1680\u2000-\u200B\u202F\u205F\u3000]/g
+  // RFC 3454 Table B.1 — "commonly mapped to nothing". The set intentionally
+  // contains zero-width joiners and variation selectors — the very characters
+  // ESLint's no-misleading-character-class warns about — because they combine
+  // with their neighbors and the RFC strips them for that reason.
+  // eslint-disable-next-line no-misleading-character-class
+  const mappedToNothing = /[\u00AD\u034F\u1806\u180B\u180C\u180D\u200C\u200D\u2060\uFE00-\uFE0F\uFEFF]/g
+  return password.replace(nonAsciiSpace, ' ').replace(mappedToNothing, '').normalize('NFKC')
+}
+
 function startSession(mechanisms, stream) {
   const candidates = ['SCRAM-SHA-256']
   if (stream) candidates.unshift('SCRAM-SHA-256-PLUS') // higher-priority, so placed first
@@ -70,7 +98,7 @@ async function continueSession(session, password, serverData, stream) {
   const authMessage = clientFirstMessageBare + ',' + serverFirstMessage + ',' + clientFinalMessageWithoutProof
 
   const saltBytes = Buffer.from(sv.salt, 'base64')
-  const saltedPassword = await crypto.deriveKey(password, saltBytes, sv.iteration)
+  const saltedPassword = await crypto.deriveKey(saslprep(password), saltBytes, sv.iteration)
   const clientKey = await crypto.hmacSha256(saltedPassword, 'Client Key')
   const storedKey = await crypto.sha256(clientKey)
   const clientSignature = await crypto.hmacSha256(storedKey, authMessage)

packages/pg/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@supabase/pg",
-  "version": "8.21.0",
+  "version": "8.21.1",
   "description": "PostgreSQL client - pure javascript & libpq with the same API",
   "keywords": [
     "database",

packages/pg/test/integration/client/promise-api-tests.js

@@ -13,13 +13,6 @@ suite.test('valid connection completes promise', () => {
   })
 })
 
-suite.test('valid connection completes promise', () => {
-  const client = new pg.Client()
-  return client.connect().then(() => {
-    return client.end().then(() => {})
-  })
-})
-
 suite.test('valid connection returns the client in a promise', () => {
   const client = new pg.Client()
   return client.connect().then((clientInside) => {
@@ -28,25 +21,7 @@ suite.test('valid connection returns the client in a promise', () => {
   })
 })
 
-suite.test('invalid connection rejects promise', (done) => {
+suite.test('invalid connection rejects promise', async () => {
   const client = new pg.Client({ host: 'alksdjflaskdfj', port: 1234 })
-  return client.connect().catch((e) => {
-    assert(e instanceof Error)
-    done()
-  })
-})
-
-suite.test('connected client does not reject promise after connection', (done) => {
-  const client = new pg.Client()
-  return client.connect().then(() => {
-    setTimeout(() => {
-      client.on('error', (e) => {
-        assert(e instanceof Error)
-        client.end()
-        done()
-      })
-      // manually kill the connection
-      client.emit('error', new Error('something bad happened...but not really'))
-    }, 50)
-  })
+  await assert.rejects(client.connect(), Error)
 })

packages/pg/test/integration/client/sasl-scram-tests.js

@@ -108,3 +108,82 @@ suite.test('sasl/scram fails when password is empty', async () => {
   )
   assert.ok(usingSasl, 'Should be using SASL for authentication')
 })
+
+/**
+ * SASLprep regression coverage. RFC 5802 / RFC 4013 require the SCRAM client
+ * to normalize the password (B.1 mapping → NFKC → prohibition + bidi check)
+ * before feeding it into PBKDF2. PostgreSQL's server applies the same
+ * SASLprep when computing the verifier, so any password whose NFKC form
+ * differs from the raw form would otherwise authenticate against psql/libpq
+ * but fail against pg with `28P01`.
+ *
+ * To exercise these tests, provision a role whose password contains an
+ * NFKC-asymmetric character. For example, in psql:
+ *
+ *     SET password_encryption = 'scram-sha-256';
+ *     CREATE ROLE scram_unicode_test LOGIN PASSWORD U&'IX-\2168';
+ *
+ * `\2168` is ROMAN NUMERAL IX; the server SASLprep-normalizes this to
+ * `IX-IX` when computing the verifier. Then export:
+ *
+ *     SCRAM_TEST_PGUSER_UNICODE=scram_unicode_test
+ *     SCRAM_TEST_PGPASSWORD_UNICODE='IX-\u2168'   (i.e. the raw form)
+ *
+ * If either env var is unset the suite is skipped, matching the convention
+ * of the ASCII SCRAM block above.
+ */
+const unicodeConfig = {
+  user: process.env.SCRAM_TEST_PGUSER_UNICODE,
+  password: process.env.SCRAM_TEST_PGPASSWORD_UNICODE,
+  host: process.env.SCRAM_TEST_PGHOST,
+  port: process.env.SCRAM_TEST_PGPORT,
+  database: process.env.SCRAM_TEST_PGDATABASE,
+}
+
+if (!unicodeConfig.user || !unicodeConfig.password) {
+  suite.test('skipping SCRAM unicode tests (missing env)', () => {})
+} else {
+  suite.test('sasl/scram authenticates a password requiring SASLprep (raw form)', async () => {
+    const client = new pg.Client(unicodeConfig)
+    let usingSasl = false
+    client.connection.once('authenticationSASL', () => {
+      usingSasl = true
+    })
+    await client.connect()
+    assert.ok(usingSasl, 'Should be using SASL for authentication')
+    await client.end()
+  })
+
+  suite.test('sasl/scram authenticates the NFKC-equivalent ASCII form of the same password', async () => {
+    // The unicode password contains a codepoint that NFKC-decomposes to ASCII
+    // (e.g. U+2168 → "IX"). The server stored the verifier from the
+    // SASLprep'd ASCII form, so feeding the client the ASCII form directly
+    // must also authenticate. This proves that the prep step is symmetric:
+    // any NFKC-equivalent representation reaches the same PBKDF2 input.
+    const client = new pg.Client({
+      ...unicodeConfig,
+      password: unicodeConfig.password.normalize('NFKC'),
+    })
+    await client.connect()
+    await client.end()
+  })
+
+  suite.test('sasl/scram fails when unicode password is wrong', async () => {
+    const client = new pg.Client({
+      ...unicodeConfig,
+      password: unicodeConfig.password + 'append-something-to-make-it-bad',
+    })
+    let usingSasl = false
+    client.connection.once('authenticationSASL', () => {
+      usingSasl = true
+    })
+    await assert.rejects(
+      () => client.connect(),
+      {
+        code: '28P01',
+      },
+      'Error code should be for a password error'
+    )
+    assert.ok(usingSasl, 'Should be using SASL for authentication')
+  })
+}

packages/pg/test/integration/gh-issues/3174-tests.js

@@ -104,7 +104,9 @@ const testErrorBuffer = (bufferName, errorBuffer) => {
     if (!cli.native) {
       assert(errorHit)
       // further queries on the client should fail since its in an invalid state
-      await assert.rejects(() => client.query('SELECTR NOW()'), 'Further queries on the client should reject')
+      await assert.rejects(client.query('SELECT NOW()'), {
+        message: 'Client has encountered a connection error and is not queryable',
+      })
     }
 
     await closeServer()
@@ -129,7 +131,9 @@ const testErrorBuffer = (bufferName, errorBuffer) => {
     if (!cli.native) {
       assert(errorHit)
       // further queries on the client should fail since its in an invalid state
-      await assert.rejects(() => client.query('SELECTR NOW()'), 'Further queries on the client should reject')
+      await assert.rejects(client.query('SELECT NOW()'), {
+        message: 'Client has encountered a connection error and is not queryable',
+      })
     }
 
     await client.end()