Summary

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 → C.1.2 mapping → NFKC → prohibition + bidi check) when computing the stored verifier. Passwords whose NFKC form differs from themselves authenticate against psql/libpq but fail against pg with 28P01.

This bug bites real users: a password like abcd123456789123456789¨‑¼### authenticates fine in psql but fails in any consumer using pg.

Fix

Inline a minimal SASLprep helper in packages/pg/lib/crypto/sasl.js and call it in continueSession immediately before crypto.deriveKey. The helper performs only the three RFC 4013 steps that change the byte content fed to PBKDF2:

1. RFC 3454 Table C.1.2 (non-ASCII space) → U+0020.
2. RFC 3454 Table B.1 (commonly mapped to nothing) → empty.
3. NFKC normalization.

Prohibition (RFC 4013 §2.3) and bidi (RFC 3454 §6) checks are intentionally omitted: libpq's pg_saslprep and Postgres's own saslprep.c are forgiving on those paths for legacy roles, so omitting them keeps existing accounts working without adding complexity. Because the three remaining steps cannot throw on a string input, no try/catch fallback is needed.

flowchart LR
  rawPwd["raw password"]
  saslprep["inline saslprep()<br/>C.1.2 → SPACE<br/>B.1 → empty<br/>NFKC"]
  derive["crypto.deriveKey<br/>PBKDF2-SHA256"]
  scram["SCRAM proof"]

  rawPwd --> saslprep
  saslprep --> derive
  derive --> scram

Why in-tree, no dependency

Updated in response to review feedback from @brianc and @charmander on this PR:

• Supply-chain surface: `@mongodb-js/saslprep` ships an unpinned transitive (`sparse-bitfield`); a runtime dep just for SASLprep was disproportionate to the problem.
• Smaller code than the dep: ~10 LOC of regex + `String.prototype.normalize('NFKC')`. The Unicode tables are inlined as character classes; no external code-points file (so it works under `workerd` without bundler tricks).
• Charmander's 3-replace shape (review comment) is the basis for the implementation. One small correction applied: U+200B (ZERO WIDTH SPACE) is in RFC 3454 Table B.1 (mapped to nothing), not C.1.2 (non-ASCII space), so it lives only in the second regex matching PostgreSQL's `saslprep.c` and `libpq`. (Cross-checked against the dep, which maps U+200B to SPACE and is therefore the divergent one here.)
• Prior art: no `brianc/node-postgres` issue/PR mentions `saslprep`, `stringprep`, `NFKC`, RFC 4013, or RFC 5802. `postgres.js` (porsager/postgres) doesn't normalize either and has the same latent bug.

Why no behavior change for existing users

• ASCII passwords: all three steps are the identity. The existing CI scram test (`'test4scram'`) is byte-for-byte unchanged.
• ASCII control characters (e.g. BEL, U+0007): unchanged by C.1.2, B.1, and NFKC alone — same bytes go to PBKDF2 as today (the renamed unit test pins this with the same SCRAM signature snapshot the previous fallback path produced).
• Cloudflare Workers: validated via the existing worker test harness — `SASLprep is engaged on the SCRAM path under workerd` still passes without any external dep.

What changed

Commits

1. `fix: apply SASLprep (RFC 4013) to passwords before SCRAM-SHA-256 PBKDF2` — original `@mongodb-js/saslprep` wiring and tests.
2. `fix: inline SASLprep, drop @mongodb-js/saslprep dependency` — review-feedback follow-up. Net diff vs commit 1: −1 dep, −1 transitive, −1 try/catch, +10 LOC inline helper, 1 test renamed (snapshot bytes unchanged).

Happy to squash to one commit on merge if you prefer; I kept them separate so the diff in response to your feedback is easy to read.

Test plan

• `make test-unit` — full pg unit suite (exit 0; all 4 SASLprep tests pass against the inline impl).
• `make test-worker` — Cloudflare Workers harness; `SASLprep is engaged on the SCRAM path under workerd ✓` passes without `@mongodb-js/saslprep` (the pre-existing `default` test fails for an unrelated reason — no local Postgres listening on the worker network).
• `yarn lint` — clean.
• `yarn build` — clean.
• Smoke test: every snapshot input fed to the inline impl produces a SCRAM signature byte-identical to the dep-based commit, including `'\u0007abc'` (BEL passthrough) and `'abcd123456789123456789¨‑¼###'` (production-bug case).
• Full integration suite against a Postgres with both `scram_test` and `scram_unicode_test` roles (runs automatically in CI on this branch across Node 16/18/20/22/24/25).