KB5101650 / KB5121003: Thunderbolt 4 NVMe causes WHEA PCIe errors, bugcheck or hang, rollback fixes

Thanks both — and thank you Jamony for the careful reply.

First, a housekeeping note. This thread is a shorter, link-free duplicate. My original post, with the full timeline, event tables, and evidence links, was auto-flagged as spam and has since been restored by the Tech Community team. Please treat that one as the canonical thread:

https://techcommunity.microsoft.com/discussions/windows11/kb5101650kb5121003-hang-thunderbolt-4-nvme-%E2%86%92-whea-17%E2%86%92-bugcheck-0xa--z790-26200-8/4550187

The Feedback Hub item, which carries the attachments (System/Setup event logs, the 233-event WHEA/disk/bugcheck export as CSV and XML, the July minidumps, and the sysinfo/msinfo export), is here — it is public and upvotes/"I have the same issue" help:

https://aka.ms/AA139bx1

On the technical point — agreed. WHEA-Logger 17 is a corrected PCIe error record, so the drip on root port 0000:00:1C.4 (8086:7A3C) is evidence of link instability, not proof of any particular cause. I named the USB-C Connection Manager change in KB5095093 as the best lead because it is the only documented change in the regression window (26200.8655 → .8875) that touches this subsystem and because Microsoft cited it for the acknowledged Dell/Intel-IPF issue; I do not claim it is the mechanism. What the A/B does establish is that the only variable between clean and failing is the cumulative update — same BIOS, drivers, enclosure, cables, and boot cycle.

Versions, for the record (all already in the FH item): ASUS ProArt Z790-CREATOR WIFI, BIOS 2801; Intel JHL8540 TB4 controller (8086:1137) behind root port #5 (8086:7A3C, driver 10.1.46.3); Intel Thunderbolt software 1.41.1412 (ASUS's current WHQL release for this board); OWC Thunderbolt Express 4M2 with 2× Crucial P3 4 TB + 2× WD Blue SN5000 4 TB, all SMART-clean; Windows 11 25H2 — fails on 26200.8875 (KB5101650) and 26200.9168 (KB5121003), clean on 26200.8655.

On the recommendations. I'll open a Microsoft support case referencing the Feedback Hub item and the good-build/bad-build comparison — that is the one channel I had not used. The rollback is containment, not a destination: updates are paused to a fixed date, and my plan for the September cumulative is an attended install with the enclosure detached, then a repeat of the same hot-attach reproduction — with PCIe Link State Power Management disabled first as a cheap discriminator, since a similar USB4/PCIe-tunnel freeze on another platform was worked around that way. Backups are on separate media and were re-verified after the rollback. The reproduction is repeatable on request if anyone at Microsoft or ASUS wants a specific trace captured.