Zero-valued pixels at the edge of TESS cut-outs bias a border-ring background and can flip a pixel check in either direction: a scan of 1186 cut-outs, a positive control, and three edge rules for a hash-registered search

A pixel check that follows a light-curve search for short-period transit signals subtracts, for every cadence, the 30th percentile of the 40 border pixels of an 11×11 TESScut cut-out. TESScut cut-outs can contain border pixels that are exactly or nearly zero instead of sky (only 5 of the 12 affected cut-outs lie within 6 pixels of the detector edge), NaN rows, or no cut-out at all. In a scan of 1186 cut-outs of the stars that reached the pixel check, 12 cut-outs of 9 stars contain such zero-valued pixels in the border ring, 4 of them with a background subtraction that drops to zero, and two cut-outs consist of NaN rows in every cadence. In three of the four severe cases the sector depth is wrong by 5-28 errors (11 errors in the best-documented case), in the fourth by 0.5 errors, and, for one star with a 7-hour period and a search signal-to-noise ratio of 17.8 (25.6 in the trapezoid fit), a rejection (χ²/dof = 23.5) becomes a borderline pass (2.41) once the ring is cleaned (2.52, a borderline fail, with the sector omitted). Inserting zero-valued pixels into clean cut-outs of 15 stars that had passed the check reproduces the effect: the aperture flux grows by a factor 1.66 (the background remains in the aperture flux), but the sector depth shifts by a median +3.4σ with a long positive tail, which is red noise of the un-subtracted sky, not dilution. A single mildly damaged sector (background about 3 % low) pushes a passing star over the χ²/dof = 2.5 threshold in 6-16 % of cases, depending on the number of sectors. We describe three edge rules (cut-out not delivered, NaN rows, zero-valued pixels) that were added to the registered procedure after these cases appeared, and that, for stars with a known earlier verdict, can only move it to "not assessable" or downward; for other stars they can yield "passed, restricted". The rules were chosen after the results of the nine stars were known; this is stated openly. The first real run under the final rules has not yet been observed. Star identifiers are pseudonyms. Nothing here is a detection or a validated planet candidate. The analysis was AI-assisted (Claude, Anthropic); the author has not yet repeated the calculations. The cut-outs are public TESScut products of TESS full-frame images; the light curves are public QLP products (MAST). The work was triggered by the corrections of two earlier notes: TIC 38699429 (Version 4, 10.5281/zenodo.23854096) and TIC 309780779 (Version 5, 10.5281/zenodo.23854128).