Release v0.19.0: AGY 1.2.2 compatibility and native repair - #133

AGY 1.2.2 can finish with exit zero after a partial-output timeout, and native same-conversation repair needs additional narrowly scoped initialization and SQLite metadata access. This change activates the verified 1.2.2 compatibility records, classifies refused actions and the observed timeout without discarding valid candidates, and restores the exercised native normal/repair workflow.

The skill keeps automatic selection, authority boundaries, same-scope repair, and driver-owned assurance while moving advanced procedures into references. Package and client metadata advance to 0.19.0.

Validation:

• Stable implementation: all 44 local offline CI stages passed; 304 candidate files had no content, mode, or file-set drift. Independently accepted.
• Bounded live qualification: session/native normal tasks and two-turn same-conversation repairs passed driver-owned checks using the caller-selected model. Thirteen total starts under renewed authority; no new Boost qualification.
• Release metadata delta: focused packaging, documentation, and Codex-usage checks; exact PR-head hosted aggregate test passed for 5db9f2416c7bc53fbf4a7bf05e11d30c9fede01e (run 34757832501).

Limits: live denial/partial-output canaries established the provider signatures, while valid-candidate preservation is covered by synthetic controller tests. This does not establish general provider correctness, future authentication/quota, or hostile same-user isolation. SkillStore publication is tracked separately from the GitHub release.

Release CI correction: the native settings fixture now invokes the canonical Command Line Tools Python image directly, avoiding the host-selected developer-tool shim. All expected permission denials remain unchanged; no production sandbox rule was widened. The final exact-head hosted gate reruns every canonical stage.

One local locator-cleanup assertion failed once without per-condition diagnostics. The unchanged retry and targeted four-condition probe passed; the same checks also passed in both full gates. The initial observation is retained with an unconfirmed cause. No cleanup assertion or production boundary was weakened.