The presenter and game host use the same mapping between a DXGI present identifier
and the refresh where that image reached the display. PresentRefreshCount already
includes composed presentation. Adding another refresh for DWM delayed the test
presenter’s shutter command; subtracting one in the game host selected the previous
eye. Composition mode now describes the output path without modifying that mapping.
Slips (2026-09-13): a frame that lands one refresh late, a command with too little lead, or a late USB transfer is now ridden through. The eye sequence is anchored to the display refresh counter and the emitter free-runs one period without a command (its own eye bit toggles at every boundary), so only the slipped refresh shows the wrong eye and the next frame is correct again. Earlier builds blanked 16 frames and re-locked the emitter on every slip; at 144 Hz fullscreen on the Hisense this happened every few seconds (session.log: 4 to 7 misses per short run, maxIntervalMs about two refresh periods) and was visible as the image jumping with no input. Ordinary phase, shutter, depth and convergence edits apply live.
No resynchronization on slips (2026-09-14): the evening build still forced a full resynchronization (16 black frames, emitter suspended and re-locked) after four slips within one second, and on every composition-mode change. With a 4K game (Psychonauts 2) captured as the source, the game’s command buffers delayed the presenter’s draw behind them and it slipped 14 to 23 refreshes a second (session.log: lead bin below zero rising, present interval jitter above 1 ms, mode changes flat). The rule then chained resynchronizations three to six times a second, and the viewer saw the glasses go black for seconds at a time. An emulated N64 game had never loaded the GPU enough to hit the rule. A resync repairs nothing in that state: the eye for each present is derived from the observed present-to-refresh mapping on every frame, and the emitter’s own eye bit toggles every period, so neither drifts. Slips are now counted only (misses, plus a per-second figure in the status line and session.log as slipsPerSec); a composition-mode change is the same case, since the queued presents land one refresh off and the mapping has caught up by the next frame. Full resynchronization remains for disjoint statistics, an occluded or resized output window, an unavailable emitter, and cadence changes in the settings. To reduce the slips themselves the process asks the graphics kernel for the HIGH scheduling class (ABOVE_NORMAL when refused; the class obtained is logged as gpuPriority) and the presenter device requests GPU thread priority 7; the present thread also opens each shared source texture once instead of on every pair.
Refresh clock (2026-09-13, night): the presenter and the game host fit a least-squares line through the last 240 presentation timestamps (refresh counter against QPC time) and predict every vblank from that line. A single timestamp’s jitter, measured at 1 to 6 us rms on this PC, no longer moves the next command by its full amount, and the residuals are reported as vblank jitter. A sample that lands more than a quarter period from the line, or a period that differs by more than 8 %, restarts the fit. The blanking interval and the position of the timestamp inside it are measured separately with the kernel scan-line counter (docs/TIMING-CALIBRATION.md).
Regression checks cover queued presents, both refresh parities, counter rollover, and a refresh anchor updated after a delayed frame. These tests establish software mapping behavior; they do not measure optical leakage through the glasses.
Reference: Microsoft’s DXGI flip model and glitch recovery and DXGI frame statistics.