fix: type a foreign key by the model it points at, and raise on a cross-type id

[Azerothian]

Fixes #65.

The bug

belongsToMany drops a join model's own id and makes its two foreign keys the composite primary key, so RoleUser.roleId / RoleUser.userId are a primary key and a foreign key at once. Both the encoder (create-basic-fields) and the decoder (globalKeyTargets) tested primaryKey first:

targets[key] = field.primaryKey ? defName : (field.foreignTarget || defName);
//                                ^ RoleUser, for a column holding a User key

So RoleUser.userId was minted and demanded as a RoleUser id — an id no client can produce from anywhere else in the schema. foreignTarget now wins wherever it exists. A shared-primary-key 1:1 table has the same shape and is fixed by the same rule.

Adapters that could not name the target at all

• valkey — ensureJoinModel marked the auto-created join model's two keys foreignKey: true with no foreignTarget, so they fell through the || defName fallback. It now takes the two ends.
• sequelize — a through model declared as its own definition but without explicit belongsTos threw at startup: sequelize builds a join model's two BelongsTos with new BelongsTo(...), which never registers them on through.model.associations. The attribute still carries references, so the model owning that table is used before falling back to the throw.

Cross-type ids now raise

Previously an id minted for the wrong type failed to decode; the opaque string was then compared literally and matched nothing — indistinguishable from a filter that legitimately found nothing.

gqlize: "RoleUser.userId" expects a "User" id, but the id given is a "RoleUser" id

GraphQLError, extensions.code = "GLOBAL_ID_TYPE_MISMATCH". The check moved out of the codecs into decodeGlobalId: a codec's null already means "not one of mine, pass it through" (which is what lets a raw primary key survive a filter untouched), so folding a forged id into the same answer made the two the same answer — and only the caller knows which field to name. An out-of-tree codec gets the check for free; a codec declaring carriesType: false is exempt.

node(id:) is unaffected — it decodes with no expected type, since the id is the type declaration there. It now filters on the already-decoded key rather than handing the global id back for a second decode, which a pk-and-fk column would have failed.

Breaking changes (both noted in docs/migration-6-to-7.md)

• A join-model / shared-pk foreign key now carries the target's global ID. Ids persisted client-side across the upgrade need re-fetching.
• A cross-type global ID is an error rather than an empty result set. Custom IdCodecs should stop returning null for a recognised id whose type does not match.

Verification

pnpm test — 9/9 packages, 1593 tests green (incl. the valkey suite against redis-memory-server). pnpm typecheck and pnpm lint clean.

New coverage: end-to-end Role -> RoleUser <- User round-trip in gqlize/__tests__/codecs/id-regressions.test.ts; both join-model shapes (with and without belongsTo on the join) in the sequelize adapter's replace-id-codec.test.ts; a cross-adapter parity assertion in the valkey relations.test.ts. Every test that asserted the old "silently matches nothing" behaviour now asserts the error message.

🤖 Generated with Claude Code

https://claude.ai/code/session_013aahHVCcqUpyw8TfWM1gSs