Safer exit policies and cache reuse

Commands now fail when they exit with 0 but their success_exit_codes list excludes it. Accepted nonzero exits continue to count as success, and failures retain their captured diagnostics. After upgrading, affected actions run again rather than reusing results recorded under the earlier behavior.

Cache validation no longer trusts mutable timestamps where reliable change markers are unavailable. Content-based action caches continue to work, while unsafe metadata-only analysis and build shortcuts are disabled on those platforms. Rebuild checks on Windows may take longer as a result.

Very deep build graphs no longer risk stack exhaustion while scheduling. Action plans validate dependencies in linear time, tolerate repeated dependency entries, and report worker panics instead of waiting indefinitely. SDK action keys keep their existing format while hashing inputs without buffering a second copy.

Garbage collection remains a best-effort filesystem snapshot. Concurrent publication can cause a cache miss, which execution safely repairs by running the action again. Files disappearing during a collection pass no longer make the scan fail.