query: positive first pushdown is unsound for allof, lossy-index eq, and geo functions

What

calculatePaginationParams pushes a positive first down to the worker for functions whose results are post-filtered or intersected after posting.(*List).Uids has already truncated the posting list. The truncation drops uids the later pass never sees, so the query can return fewer rows than it should, or none.

The exclusion list is incomplete

query/query.go, calculatePaginationParams:

switch sg.SrcFunc.Name {
case "regexp", "alloftext", "allofterms", "match", "ngram":
    shouldExclude = true

Those five are correctly kept off the pushdown. Three families are not:

The ordering is what makes it wrong. In helpProcessTask, handleUidPostings (which calls pl.Uids(opts)) runs first; handleCompareFunction and filterGeoFunction run afterwards. x.PageRange at the query layer cannot restore a uid the worker already discarded.

Union functions are fine: anyofterms merges rather than intersects, and head_N(A) ∪ head_N(B) contains head_N(A ∪ B), so taking the first N of the merged result is still correct.

Reproduction shape

name: string @index(term) .
0x1,0x2 name "great wall"    0x3,0x4 name "great"

q(func: eq(name, "great"), first: 2)

The bucket for token "great" is {1,2,3,4}. Pushed down, Uids returns {1,2}, handleCompareFunction drops both, and the query answers nothing instead of 0x3, 0x4.

Confirmed reachable, not hypothetical: TestPaginationPushdownExcludesIntersectingFunctions (added in #9809) asserts the five excluded names return math.MaxInt32, and the same harness shows allof, anyofterms, eq and has all receive a pushed-down count.

Why the negative case is already fixed and this one is not

#9809 removed the negative-first trim from Uids entirely, which fixes this for first: -N across all three families, because a negative first never had a paired early stop and so the trim bought nothing.

The positive trim cannot be removed the same way: it is paired with the early stop in the iterate loop, so it stops the read rather than just truncating the result, and that saving is the point. The fix belongs at the planner — adding these functions to the exclusion list, which gives up the pushdown for them — not in Uids.

Notes

Pre-existing, and independent of #9809; that PR only made the negative half correct. Discussed with the PR author, who confirmed the positive half touches enough query-planning and post-filtering paths to want its own change:

I took a look at fixing positive first pushdown for lossy eq, but it touches a broader set of query-planning and post-filtering paths than expected, so I agree that it should be handled separately.

There is no test anywhere for x.PageRange, and no test for allof or geo with a first argument, so nothing currently guards this.