NortheBridge/luminalshine 26.05.0-rc1

on GitHub

LuminalShine 26.05.0 Update Release Candidate 1 (RC1)

This release lands the full GPU TDR (Timeout Detection and Recovery) remediation plan as four cooperating layers: detect the failure unambiguously, bound the session-shutdown window, recover the SudoVDA virtual display in place without ever falling back to a different driver, and warn proactively when the at-risk workload pattern is present. The driver-level interaction between sustained AV1 NVENC streaming, DLSS / DLAA / Frame Generation, and 4K HDR — which has been the dominant cause of mid-session black screens on RTX 40 / 50 series — is now addressed at every layer LuminalShine controls.

General

• Streams no longer freeze for minutes after a GPU crash. When the encoder hits the post-TDR D3D11CreateDevice retry exhaustion, the session unwinds within ~10 seconds and the Moonlight client receives an immediate, intelligible failure instead of staring at a frozen image. Previously the recovery loop could leave the session wedged for ~2.5 minutes before the 10-second hang watchdog fired.
• New "GPU / Display Stack Health" card in the Troubleshooting view. Shows the count of TDR-class events recorded since the process started, when the most recent one fired, which subsystem detected it (encoder D3D11, Desktop Duplication probe, virtual-display enumeration, or QueryDisplayConfig), and the underlying HRESULT. Refreshes on login and on demand.
• Recovery-aware session gate. New streaming sessions are refused while LuminalShine is mid-recovery from a TDR, so reconnects don't pile up against a stuck display stack. The client sees a clean disconnect and retries cleanly once the recovery completes.
• Single high-signal log line for every TDR escalation. Each incident emits one GPU TDR escalated to WDDM stack failure line tagged with the source subsystem so support readers can grep one symptom instead of correlating five.
• New /api/health/tdr and /api/health/render-stack endpoints. Power users and dashboards can poll the same data the Troubleshooting card surfaces.

Windows 11 General Availability

• SudoVDA auto-recovery ladder. When the display stack is wedged, LuminalShine now escalates through a two-level in-place recovery sequence keyed on SudoVDA's hardware ID — the user-mode driver handle is recycled first; if that's not enough, the SudoVDA root device is PnP-disabled and re-enabled (the equivalent of right-clicking the device in Device Manager, but run from the SYSTEM service). Cooldown-gated: 15 seconds for the cheap recycle, 5 minutes for the disable/enable. MTT VDD and unrelated UMDF drivers are never touched.
• "Restart Virtual Display Driver" recovery button in Troubleshooting. A one-click escape hatch that runs the PnP disable+enable on demand, ignoring the auto-recovery cooldowns. Useful when you suspect SudoVDA is wedged but the auto-recovery hasn't fired yet. Includes a confirmation dialog because it tears down any active streaming session bound to the virtual display.
• "Show Diagnostic" support button in Troubleshooting. Opens a modal with the full SudoVDA PnP status — device-present flag, instance ID, hardware IDs, parsed CM_PROB status, last recovery attempt and outcome — with a Copy-to-clipboard action. Replaces "open Device Manager and screenshot it" for support tickets.
• "Virtual Display Adapter Failure detected" banner. Persistent warning across the top of the Troubleshooting view whenever the GPU / WDDM stack is mid-recovery, so the user sees what's happening before scrolling to the cards.
• Pre-flight D3D11 health check at session start. Before launching the encoder thread, LuminalShine runs a single D3D11CreateDevice probe (~5 ms) to confirm the GPU stack is healthy. If it returns DXGI_ERROR_UNSUPPORTED or DXGI_ERROR_DEVICE_REMOVED, the session is refused with a clear log line — no more sessions hanging in display init after sleep/wake or after a TDR the recovery-recent flag has already aged out.
• Streaming environment tip when DLSS / DLAA / Frame Generation is detected alongside 4K HDR NVENC. LuminalShine walks foreground processes for nvngx_dlss.dll, nvngx_dlssg.dll, nvngx_dlaa.dll and, when found alongside a 4K HDR 10-bit NVENC session, emits a single soft tip per (codec, resolution, HDR) tuple: a friendly heads-up that the combined Tensor + NVENC workload is a known driver-level interaction (not a LuminalShine defect), with a suggestion that switching to HEVC sometimes rides out concurrent workloads more smoothly. Surfaces as a system-tray toast plus a compact "Streaming environment" card in Troubleshooting. No action is required from the user — it's an FYI, not a warning.
• System-tray notification on TDR detection. One toast per 2-minute cooldown so a single incident doesn't fire repeatedly.

Windows 11 Insider Preview

• WDDM reset (Ctrl + Win + Shift + B) as recovery Level 3. When the cheaper recovery levels can't bring SudoVDA back — the canonical fingerprint of the WUDFHostProblem2 / HostTimeout pattern on Insider Preview channels — LuminalShine now synthesises the Windows built-in graphics-driver-reset keystroke. The request is routed through the existing display helper process (which runs in the user's interactive session and has the desktop integrity level required for SendInput) so SYSTEM-context callers can trigger it safely. After the keystroke is dispatched, the SudoVDA user handle is recycled and the next AddVirtualDisplay opens a fresh kernel binding against the recovered display port.
• Rate-limited at 15 minutes per process. A permanently broken adapter is never hammered in a tight loop — but a single in-session wedge that survives Levels 1 and 2 of the ladder will be cleared without the user reaching for their keyboard.
• Manual recovery button never triggers WDDM reset unprompted. The "Restart Virtual Display Driver" button caps at the PnP-cycle level (Level 2) because Level 3 momentarily blanks the entire desktop. Auto-recovery still escalates all the way to Level 3 when needed.
• The previously-stuck install pattern (TDR followed by a 2.5-minute frozen reconnect) is now handled end to end. Detect → bound shutdown → ladder recovery → pre-flight gate on the next attempt. The Moonlight client should never see a multi-minute black screen again after a single TDR.

Automated release for 26.05.0-rc1.

Full Changelog: 26.05.0-beta.7...26.05.0-rc1