fix(store): prevent steering and runner status deadlocks

[dstkwll]

Summary

Steering could deadlock with a concurrent runner status event because the two transactions acquired agent and run locks in opposite orders. Lock the exact observed run first, retain the existing agent/lease protection, and revalidate the destination before inserting a command. Changed destinations fail with a conflict and rollback; committed replay retains its original destination.

Closes #223.

Provenance and scope

• Base: main at b252396.
• Candidate: 35b76f1, branch codex/issue-223-steering-deadlock.
• Production scope: queue_message in crates/crony-store/src/lib.rs.
• Tests: crates/crony-store/src/steering_lock_tests.rs; opt-in status stream in scripts/fake-codex-app-server.mjs; Tokio test dependency in crates/crony-store/Cargo.toml and Cargo.lock.
• Documentation: docs/EVALS.md and docs/evidence/2026-09-15-steering-lock-order.md.
• Native capability: SQLx 0.8.6 and existing PostgreSQL row/advisory locks. No migration, lock weakening, provider retry engine or new permission mechanism.

Validation

• Exact baseline deadlock reproduced through native store operations with observed PostgreSQL blocking relationships.
• Ten real-migration concurrency, authority, replay and rollback regressions passed in disposable PostgreSQL.
• All six repository gates passed. Workspace tests: 517 passed, 0 failed, 333 ignored; opt-in tests were run separately.
• Actual local server and runner with a synthetic Codex protocol fixture: forced status/steering contention, HTTP replay, browser steering, two durable acknowledgments, output, usage and completion passed. No paid provider execution.
• GitHub Copilot Astra/high design and independent review completed. Test-strengthening findings addressed and retested.

Evidence and reproduction.

No hosted CI pass, real-vendor acceptance or global deadlock freedom is claimed. PRs #283 and #255 touch the store file but their inspected patches did not change these functions; refresh before integration. Merge, auto-merge, deployment and Factory intake remain disabled and unauthorized.

[Copilot]

🟢 Approval recommended

The lock inversion is corrected with consistent revalidation and comprehensive focused evidence.

Pull request overview

Prevents steering/run-status deadlocks by enforcing run-before-agent lock ordering and revalidating message destinations.

Changes:

• Locks the observed run before agent grant checks.
• Adds PostgreSQL concurrency and rollback regressions.
• Adds a bounded status-stream fixture and validation evidence.

File summaries

File	Description
crates/crony-store/src/lib.rs	Implements lock ordering and destination revalidation.
crates/crony-store/src/steering_lock_tests.rs	Adds concurrency and authority regressions.
crates/crony-store/Cargo.toml	Adds Tokio test dependency.
Cargo.lock	Records dependency metadata.
scripts/fake-codex-app-server.mjs	Adds opt-in status contention fixture.
docs/EVALS.md	Documents validation procedure.
docs/evidence/2026-09-15-steering-lock-order.md	Records scope and runtime evidence.

Review details

• Files reviewed: 6/7 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[shyamsridhar123]

Reviewed commit 35b76f162e3fbf7f80ae2819ac639fab906de8de against ECorp's product, architecture, security, and eval contract.

Recommend approval.

Validation: Complete production/test/evidence diff reviewed; migration checker and cargo fmt passed; all 10 ignored issue223 PostgreSQL tests explicitly executed on owned PostgreSQL 17 fixture and passed (19.30s); six exact-head hosted checks green. Authored full-stack synthetic-fixture evidence inspected, not independently replayed.