fix(ai): prove retained resume readiness before dispatch

[first-assist]

Summary

• expose an exact retained-resume readiness predicate requiring the correlated response and matching started notification in either order
• add a closed typed pre-dispatch rejection path that releases unstarted reservations and preserves uncertainty for every failure during or after business dispatch
• make proven pre-dispatch failures retryable while fencing retained provider state for cleanup
• add deterministic response-first, notification-first, and three-consecutive-turn coverage

Release

• graphql-orm-ai 0.95.12
• graphql-orm-ai-tool-profiles 0.10.3
• persistent AI schema module remains 0.64.0

Verification

• full AI all-target/provider-codex test matrix

• full tool-profile all-target test matrix

• focused retained-session, dispatch-classification, cleanup, coordinator, and retry tests

• warnings-denied Rustdoc

• warnings-denied Clippy with only the unchanged existing large-enum lint allowed

• bounded SemVer checks for both changed packages

• documentation, inventory, release-state, dependency-boundary, package-release, release-manifest, formatting, lock, and diff checks

• Documentation updated

• No documentation impact

Ordinary provider failures and all failures during or after business turn dispatch remain uncertain and keep the recovery boundary.