docs/planner: scoped bm25 through a traversal is ~15x slower than global bm25 on a 2.26M-row type

[ragnorc]

Summary

On a local store with 2,264,574 Passage rows (9,288 revisions, 266 matters), after optimize:

Query	Time
bm25($p.text, $q) over all passages, limit 10	0.04 s
same, scoped to one matter via $p passageOfRevision $r $r: SourceRevision { matter_number: $m } (ranked variable first)	0.65 s
bm25 over 125,931 Chunk rows, limit 10	0.02 s

This matches the documented behavior (the traversal leaves the bounded top-k window unfilled, so the engine rescans without the bound), so it is not a defect. Two suggestions: (1) a sentence in docs/user/search/index.md telling catalog authors to scope on the ranked type's own indexed property where possible, or to rank a coarser type first; (2) if cheap, let the bounded scan use a selective traversal prefilter for bm25 the way rrf does, so scoped ranking does not degrade to a full-type rescan.

Version / environment

omnigraph 0.11.0, macOS arm64, file:// store, Homebrew binary.

Related: #750 (ranking on a traversal-introduced target; the query above declares the ranked variable first, so it is the supported shape), #563 (ranked read plus join materialization, closed).

[ragnorc]

Measured on main (80406fee) at three sizes, to show the scaling. The fixture has the issue's shape: Passage -passageOfRevision-> SourceRevision { matter_number @index }, 9,288 revisions over 266 matters, one matter selecting about 0.4% of revisions, optimize run after the loads. Times are CLI end to end, median of 5.

Passages	engine	global bm25	scoped bm25	scoped rrf(bm25, bm25)
200k	v1	21 ms	56 ms	104 ms
200k	v2	25 ms	49 ms	89 ms
600k	v1	24 ms	123 ms	258 ms
600k	v2	31 ms	107 ms	214 ms
1.2M	v1	26 ms	215 ms	457 ms
1.2M	v2	39 ms	200 ms	388 ms

Global stays flat. Scoped grows linearly with the size of the ranked type (about 0.16 ms per 1,000 passages), which extrapolates to the reported ~15× at 2.26M. The results are correct: spot-checked rows belong to the selected matter.

Neither suggestion has landed. docs/user/search/index.md has no guidance on scoping, and plain bm25 has no traversal prefilter. The rrf form does not avoid the cost at this selectivity either: it is about twice the scoped bm25, so the "few entities" gate did not help this shape.