TEXTURE_MAX_LOW_TILE_RATIO

Maximum fraction of low-texture image tiles (0..1) a captured frame may contain and still be sent to inference. Frames above this — blank walls, floors, ceilings — are skipped before the expensive BoQ step, since they rarely localize. 1.0 disables the gate.

0.40, after 0.20 was tried and reverted (Run 24). The research bench's app runs 0.20 (its BALANCED preset; its TextureFilter.MAX_LOW_TILE_RATIO library constant of 0.40 is never passed), so matching it looked obviously right. Measured, it was not: on a gates-off calibration walk, tightening to 0.20 would have skipped 5 frames that produced good fixes and prevented zero bad ones, and on a lanyard leg it skipped 41 of 88 scans for no measurable gain. An earlier reading of a 60% ambiguity rate in the 0.20–0.40 band came from n=5 and did not survive a larger sample.

The deeper reason it does not help: the bad fixes this was meant to prevent do not come from poor frames. They come from clean, sharp frames with a strong cluster margin — a wrong place matched confidently. Frame gating cannot see that, and no threshold on this metric will.