feat(aws-ssm): Run Command lifecycle, parameter checks, list and cancel ops (SSM-2)

[NitinKumar004]

Summary

Run Command now behaves like real SSM instead of recording an instant success. This is SSM-2 from the build-out plan, and it also covers tracker row SSM-X3 plus the SendCommand version and parameter checks deferred from #1367.

SendCommand

• DocumentVersion accepts $DEFAULT, $LATEST or a number. A bad pattern returns ValidationException and a missing version returns InvalidDocumentVersion. The response echoes the requested version ($DEFAULT when none is given).
• Parameters are checked against the declared parameters of the selected document version: undeclared names, required parameters with no default, type (String/Integer/Boolean/StringList/StringMap/MapList), allowedValues, allowedPattern (full match), minItems/maxItems and minChars/maxChars. Failures return InvalidParameters. AWS-owned names that are not in the catalog are still accepted without checks.
• Request constraints from the botocore model: MaxConcurrency/MaxErrors patterns, TimeoutSeconds 30..2592000, Comment up to 100 characters, at most 50 instance ids and 5 targets, and the instance id pattern.
• Explicit instance ids must be running. An unknown id is InvalidInstanceId, and so is a stopped one ("not in a valid state"). Tag targets skip terminated instances. A stopped instance picked by a tag stays Delayed and ends as DeliveryTimedOut after TimeoutSeconds.
• The response reports Pending, echoes every request field and includes TargetCount, RequestedDateTime and ExpiresAfter.
• With S3 wired, OutputS3BucketName gets the output object under <prefix>/<command>/<instance>/<plugin>/<step>/stdout once the invocation finishes. Without S3 the bucket is only echoed back. The output URLs are reported on the plugins, and on the invocation when the document has a single plugin.

Lifecycle

• Status is worked out from the clock at read time, following the FIS pattern. With --async-settle, invocations go Pending (1s), then InProgress (5s), then Success. They run in MaxConcurrency batches. An executionTimeout shorter than the run ends as TimedOut/ExecutionTimedOut.
• Command status is aggregated with the real rules: Success, Cancelled, Failed once the errors go over MaxErrors, TimedOut, Incomplete and NoInstancesInTag. CompletedCount, ErrorCount and DeliveryTimedOutCount are filled in.
• With settle off, commands finish at send time, as before.

New ops

• ListCommands: CommandId, InstanceId, the five filter keys (InvokedAfter/InvokedBefore/Status/ExecutionStage/DocumentName), newest first, paging. An unknown key is InvalidFilterKey and a bad token is InvalidNextToken.
• ListCommandInvocations: the same filters (except ExecutionStage, which is InvalidFilterKey here) and --details, which returns CommandPlugins (one per runtimeConfig plugin or mainSteps step).
• CancelCommand: cancels unfinished invocations on all targets or on the listed instances. An unknown command is InvalidCommandId and an instance that is not a target is InvalidInstanceId. Finished invocations keep their result.
• DescribeInstanceInformation: every EC2 instance that is not terminated is listed as a managed node. Running instances are Online and the rest ConnectionLost. It supports both filter lists and paging (5..50).
• GetCommandInvocation reads from the same state. It takes PluginName (an unknown one is InvalidPluginName) and reports ResponseCode -1 until the run finishes, along with ExecutionStart/End/Elapsed.

Persistence

Command history is saved in the snapshot under commandHistory, which includes cancels and output-written flags. A restored command keeps settling from its send time. Snapshots in the older per-invocation commands form are converted on restore.

Testing

• New SDK tests (server/aws/ssm/run_command_lifecycle_test.go) all failed on the old code. The provider tests cover MaxConcurrency staggering, delivery timeout on a stopped tag target, the snapshot round trip and legacy restore.
• Existing tests were updated where real SSM rejects what they sent: a missing required commands parameter, and instance ids that don't match the id pattern.
• go build ./.... go vet and go test -race on internal/settle, providers/aws, providers/aws/ssm, server/aws/ssm and persist. golangci-lint --new-from-rev=origin/development is clean. coveragegen regenerated.
• E2e against cloudemu serve, with and without --async-settle:
  • aws CLI: send-command with AWS-RunShellScript, with a missing required parameter, with a bad pattern and with a bad version; a user document with typed parameters (Integer, allowedValues, required); list-commands with filters and paging; list-command-invocations --details; get-command-invocation with --plugin-name; cancel-command while the command was InProgress (it ended Cancelled) and with an unknown id; the S3 output objects listed.
  • Terraform aws_ssm_document: apply, no-drift plan, update to version 2, no-drift plan, send-command against the TF document, destroy.
  • aws_ssm_association was skipped because Associations are not implemented yet (SSM-3).