Cruxible

Added

• cruxible gate: declared merge gates enforced from state: a gates:
  config element declares named, kind-based gates — {kind, entity_type, match_property, condition}, where kind selects a source adapter that
  supplies the candidate values to check. cruxible gate check <name>
  evaluates a gate; the only v1 kind, git-pre-push, reads git's pre-push
  protocol and requires every parent of every pushed merge commit to be
  pinned by a matching entity in state, refusing the push otherwise (fail
  closed on any error). The agent-operation kit ships a merge-review gate
  (ReviewRequest / change_head / approved) so a repo can gate merges on
  approved reviews with a one-line pre-push hook. Doctrine: a guard blocks
  a write into state; a gate lets the world act only if state agrees.
• Approval actor separation (distinct_from_creation_actor): mutation
  guards can now require that the acting actor differ from the actor that
  created the target entity — anchored on the creation receipt's
  server-derived actor identity, never on writable properties. Fail-closed:
  entities with no committed creation receipt or no recorded creation actor
  refuse the guarded transition, and create-with-guarded-value is always
  refused. The agent-operation kit's review-approval guard now combines its
  allow-list with separation, so the actor that files a ReviewRequest can
  no longer approve it. Consequence: importing records in an
  already-approved state is refused — land reviews as requested and
  approve under a second credential.

Security

• Feedback channel now honors write-tier boundaries
  (wi-feedback-write-tier-bypass): a governed_write feedback correct
  could apply arbitrary edge property corrections to relationship types
  whose direct-write surface requires graph_write, and reject/flag
  could move an edge out of live review state with no actor identity under
  server auth. Corrections are now gated at the corrected relationship
  type's config-declared write_tier (default graph_write) across
  feedback, feedback_batch (strictest corrected type in a mixed batch),
  and feedback_from_query (target resolved from the receipt before the
  check), refusing with the same PermissionDeniedError as the direct-write
  facades. Under server auth, every feedback action (approve /
  correct / reject / flag) now requires a resolved actor identity —
  anonymous retraction ends alongside anonymous promotion. Auth-off local
  behavior is unchanged, as are governed corrections on types that declare
  write_tier: governed_write.