fix: stop deriving the WebAuthn user handle from the typed name

[NiKrause]

Closes #45.

user.id was TextEncoder().encode(userId). An authenticator keeps one discoverable credential per (rp.id, user.id) and replaces the previous one when both match — silently, with no prompt and nothing to undo. So two people registering as "anna" on a shared device destroyed each other's passkey, and with it the DID and every entry signed under it. One person re-entering their usual name after clearing storage did the same to themselves.

WebAuthn L2 §5.4.3 is explicit on both counts: the handle MUST NOT carry personally identifying information such as a username or e-mail address, and SHOULD be 64 random bytes. It is now exactly that. userId keeps its rightful place as user.name, the label the credential picker shows — a label only: it identifies nothing, two credentials may carry the same one, and nothing here looks a credential up by it.

Nothing needs migrating

No path in this package resolves a credential by handle. Recovery goes through discoverable credentials (readLargeBlobMetadata calls get() with no allowCredentials), or through an explicit credential ID where those are switched off. Credentials registered under the old scheme keep working untouched, and the derived did:key comes from the credential's public key, so it does not change.

What consumers will notice

Re-registering under a name that was used before now adds a passkey instead of replacing one. That is the point — a replaced passkey is data loss, a second entry is a choice — but the picker can now show several, so user.name and displayName should be distinguishing enough to choose from. The new handle is returned as credential.userHandle (base64url) for callers that will one day pass allowCredentials; the authenticator keeps its own copy, so storing it is optional.

Two consumer-side follow-ups, out of scope here: the two input fields in yogasuci's StudioGate.svelte can become one — a code comment there already anticipates exactly this change — and its hint "If somebody on this device has already used the same name, pick another: theirs would be replaced" is now false and should go. simple-todo's passkey01/acl01 chapters carry the same two fields plus a user id (e.g. alice@example.com) placeholder that should stop inviting an e-mail address.

The mock now models the rule it was silent about

tests/helpers/mock-authenticator.js returned one fixed credential ID per instance, so replacement could not be observed at all. It now files credentials by handle and drops whatever was in the slot, which is what makes the regression test meaningful rather than decorative.

Verification

• All 67 node tests pass; prettier and eslint clean.
• Negative check: with the old line restored, three of the four new tests fail. The fourth asserts the label, which was always correct.
• The load-bearing assertion reads the user dictionary handed to create(), so it holds regardless of how faithfully the mock behaves.
• Added to playwright.node.config.js, so it gates npm publish rather than only the browser run.