@@ -15,20 +15,15 @@ Reference: https://github.com/cntools/libsurvive
1515 ` rocketmark/libsurvive:upstream-pr/variance-nan-guard ` .
1616- ** Process noise dt cap** — PR open against collabora/libsurvive from
1717 ` rocketmark/libsurvive:upstream-pr/kalman-process-noise-dt-cap ` . References collabora/libsurvive #346 .
18+ - ** ` mpfit_nan_guard ` ** — PR open against cntools/libsurvive from
19+ ` rocketmark/libsurvive:upstream-pr/mpfit-nan-guard ` .
1820
1921---
2022
2123## High priority — bug fixes, no controversy
2224
2325These are small correctness fixes with no API surface. High acceptance probability.
2426
25- ### ` mpfit_nan_guard `
26- - ** File** : ` src/poser_mpfit.c `
27- - ** What** : Guards NaN optical angles in ` construct_input_from_scene ` . Same root cause as
28- above — bad FPGA timestamps from USB disturbances produce non-finite angles that crash
29- the MPFIT solver.
30- - ** PR pitch** : "poser_mpfit: skip non-finite optical angles instead of passing to solver"
31-
3227### ` gss_maxiter_guard `
3328- ** File** : ` src/poser_mpfit.c `
3429- ** What** : Treats ` MP_MAXITER ` (solver hit iteration limit without converging) as a
@@ -53,14 +48,6 @@ These are small correctness fixes with no API surface. High acceptance probabili
5348
5449## Medium priority — improvements worth submitting
5550
56- ### Process noise dt cap (hunk from ` reflection_rejection ` )
57- - ** File** : ` src/survive_kalman_tracker.c `
58- - ** What** : Caps ` dt ` at 0.05s in ` survive_kalman_tracker_process_noise ` . Without this,
59- a 1s IMU gap produces ` t^7 = 1.0 ` , causing NaN/Inf in the Kalman filter
60- (observed: quatrotateabout assertion on cold start).
61- - ** Note** : Extract as a standalone patch — don't submit with the full reflection_rejection.
62- - ** PR pitch** : "kalman: cap process noise dt to prevent t^7 explosion on cold start"
63-
6451### ` sync-cluster-window ` config (hunk from ` reflection_rejection ` )
6552- ** File** : ` src/survive_sensor_activations.c `
6653- ** What** : Makes the sync cluster time window configurable (was hardcoded to
0 commit comments