Sparse (content-defined) gram selector.

Selects only a content-determined subset of trigrams, shrinking the index and the candidate sets. A gram at byte position i (the trigram over Bytes[i..i+2]) is kept iff a hash of its local window passes a sampling test: phash2(Window, SampleRate) =:= 0. The window has radius r, spanning Bytes[i-r .. i+2+r] (width 3 + 2r), so the choice depends only on local bytes and identical substrings select identical grams.

The boundary rule

When a literal L occurs in a document D at position p, a gram at literal-position i whose window lies entirely inside L (i >= r and i+2+r =< len(L)-1) has, in D, the identical window (that range sits inside the matched region), so it is selected in D iff selected in L. Grams whose window spills past L's edge see different bytes in D and are NOT reliable.

So select_grams_positional/2 (index side) selects over every position, padding document-edge windows with a sentinel, while reliable_grams_positional/2 (query side) selects only over INTERIOR positions whose window is fully inside the literal, and falls back to brute_force when there is no such gram. Relying on a boundary gram would be a silent false negative, so the query never does. select_grams/2 and reliable_grams/2 (the plain, non-positional behaviour callbacks) are exactly the gram sets of these with the offsets dropped -- implemented as a thin wrapper, not a separate computation, so the two can never drift apart.