FR-26219 Namespace every SpiceDB read to the resolved instance - #81

InvalidObjectTypeException extends Error, not ConfigurationInputIsInvalidException,
so the rethrow guards did not cover it and a caller passing another instance's
prefix — v_other/document — was answered with the fallback boolean instead of an
error. That is the behaviour d05162d exists to prevent.

Single calls throw, which is what isEntitledTo should do for a caller-input
error. Batches fail per item instead: rethrowing inside the Promise.all made one
bad entityType discard every other result, so the offending item now carries
`error` while its neighbours answer normally. `result` is left undefined rather
than false, so nothing is granted and a caller can tell a refusal from a denial.

EntitlementsResult gains an optional `error`, which is additive for consumers.

Verified against an unreachable endpoint, so SpiceDB is never consulted: before,
isEntitledTo resolved to {"result":false}; now it throws. A three-item batch with
one bad entityType returns true / error / true.

The block splitter counted every brace in a line, including braces inside
string literals and comments. A caveat body containing x == "{" left the block
open, so everything after it was returned as part of that instance:

  definition v_aaa/user {}
  caveat v_aaa/c(x string) { x == "{" }
  definition v_bbb/secret {}

  readSchemaFor('a') -> "... definition v_bbb/secret {}"

SpiceDB v1.53.0 accepts that schema, so it was reachable, and it defeats the
reason readSchemaFor exists. Braces are now counted only outside single and
double quoted strings, line comments and block comments, with escapes handled.
A block-comment line is also no longer mistaken for a block header.

Six cases pinned in spicedb-entitlements.client.read-schema.spec.ts, including
that the instance's own caveat body still comes back intact.

BREAKING CHANGE: spiceClient is no longer public. Direct access bypasses
instance namespacing and can read another instance's data. Use isEntitledTo,
the lookup methods, or readSchemaFor with an instanceId instead.

The field carried @deprecated in an earlier revision of this branch. A private
field cannot be deprecated for callers who can no longer reach it, so the note
goes with it.

… document them

- ClientConfiguration extends InstancesConfiguration now that the client reads it
- specs follow the registry API from the base branch (instanceId on every namespace, 'legacy' id)
- README documents instances, defaultInstanceId, schemaPrefix and the resolution exceptions

…nstance, and fail closed on bad input

- the schema block parser lives in its own module, carries string and comment state across lines, and throws SchemaParseException on unbalanced input
- readSchemaFor takes options and returns the instance's blocks with its prefix stripped
- a failing isEntitledToMany item is { result: false, error } and is logged with its instanceId
- LoggingClient.log and error accept an optional { instanceId }
- only InvalidObjectTypeException is treated as caller input; prefix-escape specs run through the real query client
- route checks log the unprefixed type and key the cache by schema prefix
- the client builds its registry once; the injectable registry parameter is gone

…ranch

- InstanceOptions lives with the other instance types
- the schema block parser is a utils module with its constants in instance.constants.ts

A caveat body using a normal CEL field access broke readSchemaFor for every
instance, not just the one whose caveat it was:

  caveat v_aaa/c(attrs map<any>) {
    attrs.definition == 1
  }

SchemaParseException: 'definition' starts before the previous block is closed

The header scan already refused a keyword glued to an identifier or preceded by
a slash, but a dot passed, so the `definition` in `attrs.definition` looked like
a new block. It now refuses a dot too.

Failing closed rather than leaking was already the safer half of this; what was
left is that the read failed at all.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

… canonical

The rewritten base made SchemaNamespace's constructor private behind
prefixed() and legacy(); three spec call sites replayed from this branch still
used new. They follow the base now rather than reopening the constructor.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>