[v0.17 proof] Keep qualification controls on the server's active directory

[TheGreenCedar]

Context

The accepted 0.16.3 candidate exposed a qualification-harness race in proof run 30600248269. The Windows job ended with embedding_qualification_control_event_timeout:crash_server, after real Vulkan execution and the publication fault phase had already succeeded.

The retained evidence identifies the seam:

• publication-fault production left replacement server PID 75712 polling qualification-suite;
• the measurement producer rebound CODESTORY_EMBED_QUALIFICATION_DIR to qualification-suite/artifacts without replacing that server;
• reset_owner observed the live server, wrote crash_server under the child directory, and waited for an event from a process still polling the parent directory.

The exact released candidate later passed Windows Vulkan package and candidate-installed runtime proof in Auto Release run 30602757294. This is post-0.16.3 harness work, not a release defect.

Acceptance

• Keep a resident qualification server and every control/event producer on one authenticated directory, or replace the server before rebinding the directory.
• Add a sub-90-second Windows-native regression that reproduces the parent/child directory mismatch without rebuilding the package.
• Prove the measurement reset and the named inflight server_crash scenario in the same producer sequence.
• Mutate the implementation to restore the directory rebind with a surviving parent-directory server; the regression must fail with the named mismatch rather than a generic timeout.
• Use focused checks during implementation. Run broad proof only on a later accepted integration head.

Refs #1179

v0.17 issue-map role — 2026-08-01\n\nThis issue remains the exact dispatched Windows qualification tracker. The control-directory source repair is isolated in #1707; its implementation PR closes #1707 only. Do not use #1634 itself as a stack layer.