Make resource operation outcomes explicit

[djgrant]

This PR removes the error-name and error-message matchers from resources.

Resource error API

Import both errors from @notation/resource:

import{ResourceNotFoundError,ResourceOperationPendingError,}from"@notation/resource";

newResourceNotFoundError(message: string,options?: {cause?: unknown });newResourceOperationPendingError(message: string,{retryAfterMs: number;callbackContext?: Readonly<Record<string,unknown>>;cause?: unknown;});

Handler result	Meaning	Reconciler action
Return normally	The operation finished.	Continue.
Throw ResourceNotFoundError from read	No resource exists for this key.	Treat it as absent during planning and refresh.
Throw ResourceOperationPendingError	The operation has not finished.	Wait, then call the same handler again.
Throw any other error	The operation failed.	Stop the deployment.

ResourceNotFoundError

thrownewResourceNotFoundError("Lambda function was not found",{cause: error,});

Only read throws this error. A delete handler must catch the provider's missing-resource error and return normally.

ResourceOperationPendingError

thrownewResourceOperationPendingError("Waiting for Lambda to become active",{retryAfterMs: 1_000,callbackContext: { requestId },cause: error,});

retryAfterMs is required. callbackContext is optional plain data. The reconciler passes it as the final argument on the next call:

read: async(key,callbackContext)=>{// ...};

The default limit is 30 calls. maxOperationAttempts changes the limit. A pending operation that reaches the limit fails.

The reconciler does not guess that a missing read after create or update is temporary. The resource must throw ResourceOperationPendingError if it knows the resource is still becoming visible.

Other changes

• Removes ResourceNotReadyError, the matcher declarations, the old retry settings, and the indeterminate plan result.
• Makes reading a missing zip side-effect free.
• Stops plan input errors from being hidden as values that will become known later.
• Keeps delete operations successful when the resource has already gone.

Tested with pnpm build, pnpm typecheck, and pnpm test:once.