[mindfn]

Restore the unified message send and queue drain after the #222/#223 rollback, with both normal-send regressions corrected. The comparison is against the accepted #221 tree restored in c29dba70; it includes the reviewed unified-delivery changes, rather than only the preparation fix.

• Preserve empty implicit target arrays through Redis storage and transitions, so ordinary messages can drain and HTTP Steer does not raise an internal error.
• Keep guidance attached to its exact preparing parent; dispatcher readiness signals the existing drain instead of leaving the message waiting until Stop.
• Add formal composed tests using real owned Redis, the common send/drain, the real HTTP Steer route and delayed dispatcher readiness. FIFO, privacy, ownership, cancellation and target/policy replay constraints remain covered.

Published head: 247a8cf93893453968ee22bbc9f34c6c2b95c090. Its file tree normalizes exactly to reviewed source fd09784d3f4aea51b365258251639badbab90a9e, retaining the fork-only #221 plugin changes and baseline files. No process evidence packages are added.

Validation: author focused API regressions 247/247 and native-owner contract fixtures 50/50 passed. Astra independently ran the composed Redis/Steer and conversion checks, 7/7 passed; earlier exact preparation review remains recorded separately. Native dispatch is controlled by test fixtures, not a live provider. The repository CI workflow only targets PRs into main; this develop_base PR has no automatic checks. Remote CI is being collected on the same source tree in upstream draft zts212653#1398, head fdce1bc25ee18ebbc9511e46c43517345652626e, and has completed with 16 successful checks plus one planned serial-bootstrap skip (CI run38064976487; Windows run38064976484).

Keep Draft pending exact-candidate acceptance and operator experience/soak. The running checkout remains the accepted #221 rollback; it has not been updated or restarted. The delayed native session eventually connected, but the prior writer release cause is not established. This is not a declaration that the runtime incident is fully resolved.

Additional isolated provider check: real Redis/send/drain plus the production CodexAppServerClient and installed CLI passed 2/2 (automatic guidance and real HTTP Steer). Both fresh native sessions consumed the appended instruction and completed one turn; no transport stub was used. The routing/lifecycle adapter remains controlled, so this is not full index/browser or operator/soak acceptance. Existing runtime sessions and data were untouched.

Exact-candidate acceptance update (2026-10-10): Astra independently passed the complete product index → HTTP → isolated Redis → real Codex path, 2/2. Console interactions on exact fork HEAD 247a8cf then passed automatic guidance and visible Queue Steer, 2/2: each retained one native turn and one response carrying both inputs, guidance read receipts, and final empty Queue/active execution. Reload preserved the receipts. These are separate runs, not a combined test count. One Steer follow-up returned ENTRY_NOT_FOUND after ordinary drain won delivery; the existing client convergence handled it with no failure toast or duplicate execution. This is not a claim of zero network errors. Only one configured Codex member was exercised. Isolated services are stopped, all data retained, runtime stays on the #221 rollback. Keep Draft until operator experience and fork soak; earlier native writer release remains unexplained.