Scrubbed production shell startup intermittently fails with 0xC0000142 (STATUS_DLL_INIT_FAILED) on hosted runners #838

Describe the bug
The hosted Windows shell-integration job can fail during Scrubbed production shell startup with exit -1073741502 (0xC0000142, STATUS_DLL_INIT_FAILED) after the preceding process-heavy checks pass. In PR #835 at head 46469d751975d6cf614acf1501ef5a47dd9af086, concurrent two-shell daemon handoff and shell smoke/stress passed, followed by WINDOWS_SHELL_DEFAULT_PROFILE_UNCHANGED=verified, then the startup failure reported by Tools\windows\validate.ps1:267.

The operator's read-only history inspection found the same exit signature on two unrelated earlier branches. This is a recurring hosted-environment signature, not evidence that #835's new command wrapper caused the failure. The failing DLL and underlying cause have not been established.

To reproduce

1. Exercise the existing automatic Windows shell-integration workflow, including its production smoke/stress checks followed by scrubbed startup. The concrete observed executions are linked below; this is not a deterministic local reproducer.
2. In the failing job, inspect the transition from successful smoke/stress and concurrent handoff to Scrubbed production shell startup.
3. Observe the -1073741502 exit and the resulting validate.ps1:267 failure. No new local native execution or reproduction has been performed for this issue.

Expected behavior
Scrubbed production-shell startup succeeds after the earlier checks, or emits enough narrowly scoped diagnostics to identify the DLL-initialization failure and distinguish environmental resource exhaustion from an incomplete scrubbed environment. Preserve the original failure; do not silently weaken the startup check.

Screenshots
None. This is process-start/CI evidence; no screenshot or visual-parity claim is involved.

Environment

• GraphCode version: Windows development head 46469d751975d6cf614acf1501ef5a47dd9af086 for the current observation.
• macOS version: Not applicable; these are hosted Windows jobs.
• Backend: pinned Windows shell validation environment; no particular backend defect has been established.

Additional context
Observed jobs:

• PR Add guarded Windows command execution wrapper #835, run 38832332464, integration job 116753712306.
• Earlier lane A/Windows parity: Dialog and sheet framework: in-window dark sheets that size to content and never clip #768 branch, run 38588158638, job 115936134617.
• Earlier deferred-loop-cleanup-v2 branch, run 38175936161, job 114578249675.

The operator found the prior two occurrences among 15 inspected failed shell-workflow runs. The current startup step follows roughly 460 earlier process-heavy checks (464 prior checks reported by the operator). Desktop-heap exhaustion or a missing environment variable required by DLL initialization are hypotheses, not proven causes. Do not present this as a #831 workspace-convergence assertion or as proof of the #826 daemon/EOF mechanism.

Current raw CI log retained locally at C:\gcs\coord2\wrapper835-original-integration-failed-raw.log; SHA-256 55B913B777F3C1F588F4784903F1D9F9BE116146DCB6F71329E2744D14E2349F. The local path is a coordinator receipt, not an uploaded attachment; the public job link is the hosted evidence. The operator independently inspected the historical occurrences; the coordinator directly retained and read the current #835 job log.

Disposition: no implementation worker yet; queue investigation after #834. Do not rerun #835's old head merely for this issue: its post-#836 rebase receives fresh CI. If this exact signature recurs on that new head, the operator permits one diagnostic failed-job rerun per head, citing this issue. A different signature must be reported for a new decision; this issue does not grant local native execution, a new fix owner, or unbounded reruns.