Preserve owner resume and consistent execution recovery

[mikamikasuki]

An Agent can resume too early while its accepted submission is still waiting for actual execution receipts. Independently declared graph tasks with identical methods can also be skipped, and stale cancellation state can prevent a later owned process from being stopped.

This change:

• Waits for a settled submission receipt before continuing the parent Agent.
• Gives independently declared graph tasks their first execution while retaining duplicate-work suppression for later clones and reopened completed work.
• Allows an explicit owner Resume to use newly available project time without resetting usage, replaying work, or bypassing an exhausted task limit.
• Retires confirmed process cancellation state and reads process records as a stable snapshot across atomic publications.
• Uses the existing historical-control quarantine during Backup/Restore so restored records preserve evidence without retaining authority over original execution.

Validation: Linux regression suites passed 78 tests (9 PostgreSQL cases skipped) and 64 process, handoff, and budget tests. The owner-resume, Backup/Restore, and controller-parallelism suite passed all 18 cases with SQLite and PostgreSQL enabled on Linux; backup and migration tests also passed on macOS. Independent-task suppression and process-publication races were reproduced on main and passed with these fixes. Full GitHub CI runs separately.