Recover incompatible retained provider sessions

[first-assist]

Summary

• fence incompatible active retained-provider bindings into the existing cleanup lifecycle before provider dispatch
• permit a deleted generation to rebind within the same pre-dispatch run only after exact absence and committed no-dispatch evidence
• allow recovery-required runs to retry only when committed rows prove no assistant output and no possible provider dispatch
• preserve cleanup origin across transient deletion backoff and audit each retry reason separately

Safety

• live incompatible cursors are never resumed or silently replaced
• reserved, committed, uncertain, and unknown reservation states remain non-retryable
• same-run replacement remains limited to server-authored descriptor, transcript, watermark, or expiry cleanup
• provider cleanup and replacement remain exact binding-generation operations

Verification

• full graphql-orm-ai all-target test suite
• focused retained-session cleanup, backoff, exact-absence, rebind, and retry-admission tests
• warnings-denied Clippy and Rustdoc
• PostgreSQL retained app-server provider build and MSSQL alternate-feature build
• documentation, inventory, release-state, release-manifest, dependency-boundary, package-release-policy, SemVer, formatting, lockfile, and diff checks

Documentation

• Documentation updated
• No documentation impact

Versioning

graphql-orm-ai is 0.96.0 because AiRunRetryEvidence gains one required public evidence field. The persistent schema module remains 0.64.0; no database or GraphQL schema migration is introduced.

[first-assist]

Refreshing CI event metadata at the unchanged reviewed head so documentation governance evaluates the corrected PR body.