Cut core and CLI over to the Yieldstar runtime

[djgrant]

Stacked on #42.

#42 adds the durable reconciler without changing the application runtime. This PR switches core and the CLI to it, then removes the old Reconciler.

What changes

• deploy, destroy, and plan use the same SQLite workflow database.
• A pending resource operation can sleep without keeping the CLI busy. The runtime wakes it and resumes the same execution.
• --execution-id can resume a crashed deploy or destroy from its recorded steps.
• Each execution ID is bound to one deployment and one workflow, so it cannot be reused for different work.
• Existing .notation/state.json data is imported once and then renamed to .migrated.
• maxOperationAttempts is available on deploy, plan, and destroy options.
• The CLI reports one error from one boundary instead of calling process.exit inside each command.
• The old reconciler, old operation modules, refresh workflow, state leases, and their tests are removed.

Provider operations still use the #44 contract: return when finished, throw ResourceNotFoundError when a read confirms absence, throw ResourceOperationPendingError when another attempt is required, and let every other error fail.

Verification

pnpm install --frozen-lockfile, pnpm build, pnpm typecheck, and pnpm test:once pass. The cutover suite contains 96 passing tests and 4 skipped tests.

Yieldstar 0.5.0 limitation

The SQLite runtime records a claimed task time in milliseconds but compares it with seconds. A task abandoned by a dead process does not become visible again at the intended time. Rerunning the same execution ID still resumes the workflow. Fixing automatic recovery requires an upstream Yieldstar release.

[djgrant]

Superseded by #45, which combines this cutover with the final state of #42. This branch predated the last eleven commits on the foundation branch (the shared operations library refactor), so its reconciler internals were stale; #45 carries the cutover's core/CLI/state/docs changes merged onto the current foundation.