perf(gfql): elide the interval shift in the dense-domain two-hop kernel (q8@100k, round 2)

[lmeyerov]

What

Round 2 on the q8@100k residual, stacked on #1844 (perf/gfql-q8-class-fix; this PR's base is that branch, so the diff here is only the round-2 commit plus its changelog entry — GitHub will retarget to master when #1844 lands).

Round 1's kernel report blamed the ~31.5 ms local floor on "memory-bound single-threaded bincount over 2.4M edges" — line-level profiling says otherwise. On the 100k graph-benchmark (2.42M FOLLOWS edges, Person ids 1..100000 so lo=1), the 26.3 ms local kernel split: 16.2 ms materializing the two shifted id arrays (src-lo, dst-lo: two fresh ~19 MB allocations per call, page-fault bound in-context — the same subtract standalone is 0.9 ms), 3.2 ms in the two per-Series bounds-proof chains, and only 6.2 ms in the bincounts. The bincount was never the floor; the shift around it was.

Three changes, all inside the proof-gated kernel, values identical by construction:

• Shift ELISION: when lo >= 0 and hi+1 fits the SAME table budget the domain guard already enforces (4E+1024), count the raw arrays with tables of size hi+1. The bounds proof guarantees nothing lands below lo, so both tables stay aligned and the [0, lo) prefix is provably all-zero — contributing nothing to the product-sum.
• Distant (lo large) or negative intervals still shift, but through ONE reused scratch buffer (xp.subtract(..., out=)) instead of two fresh E-length arrays; mixed src/dst dtypes keep the plain shift.
• The polars bounds proof fuses all six reductions (min/max/null_count per endpoint column) into one parallel select.

Measured rejections (2.42M edges, 24-core local, diagnosis-only): polars group_by degree counting 18.7 ms vs bincount 6.9 ms; np.add.at 7.9 ms; one-pass concat bincount 25.7 ms; chunked threaded bincount 4.8 ms best-case — real but marginal against pool overhead, dropped.

Honest transfer note

The local ~2x (counting body 23.6 → 6.9 ms; polars q8@100k 33.9 → 17.3 ms local) did NOT transfer to the DGX board host: the shift cost there was local page-fault behavior absent on DGX. On the receipted DGX candidate lane the q8@100k polars cell moves 15.94 ms (round-1 kernel alone) → 14.72 ms, and the comparator verdict stays LOSE 1.52x against warmed Kuzu 9.70 ms. The change is kept for the elision's correctness-neutral simplification plus the small real gain. Round-3 target: the Kuzu warmed goalpost is ~9.7 ms.

Measured cell (attributed: q8@100k, jointly with #1844)

Measurements come from the COMBINED candidate build 938f22851e68918799da2bee49c88db2e2c103ee (master f875724ce + perf/gfql-q8-class-fix + perf/gfql-singlehop-subsumed-semijoin + perf/gfql-q7-two-star-minimal-join + this branch), with per-fix attribution by disjoint cell sets: the q8 branches own q8 (this one the 100k residual), the single-hop PR owns q1/q3/q4, the two-star PR owns q5/q6/q7.

Cell	Baseline (master pin)	Round-1 kernel (DGX)	Candidate (this build)
q8 @ 100k	polars 33.25 ms — LOSE 2.44x	15.94 ms	polars 14.72 ms — LOSE 1.52x
q8 @ 20k	polars 8.24 ms — LOSE 2.94x	—	polars 2.22 ms — WIN 1.27x (WEAK: slot ranges overlap) (attributed to #1844; 20k was unchanged within noise by this round locally)

Untouched cells did not move: cells outside the q8 attribution changed only where attributed to the sibling branches; the unattributed cells (q2, q9) kept their WIN verdicts at both scales.

Receipts

• Lanes: graphistry/pyg-bench#170 — results/graphbench-board-20k-cand-20260803/ and results/graphbench-board-100k-cand-20260803/ (compare.txt, SOURCE_COMMITS, lock-receipt, 1 Hz load receipts, SHA256SUMS).
• Build SHA in both lanes' SOURCE_COMMITS: pygraphistry 938f22851e68918799da2bee49c88db2e2c103ee, pyg-bench fae1e975.
• VALIDITY: 20k "VALID: no external process in any over-limit capture (0 receipted benchmark-own spike seconds)"; 100k "VALID: … (106 receipted benchmark-own spike seconds)".
• Protocol: position-balanced K G I I G K K G I I G K slots, RUNS=51/WARMUP=5, per-slot medians, tie band 10%, weak-flag on overlapping slot ranges; verdicts published only where GFQL row sets match Kuzu.

Tests: lane pins via an array-namespace spy (shift-elision lane counts raw arrays with hi+1 tables and never allocates; distant-interval lane routes both columns through one scratch buffer; mixed-dtype lane keeps the plain shift), both sides of the table-budget boundary, non-integer edge-column declines on both engines, the helper's standalone empty-frame guard, and a CPU emulation of the cudf null-mask series shape. Existing round-1 T6 value-parity, offset/negative-interval, and decline pins all pass unchanged.

Do not merge without owner review.

🤖 Generated with Claude Code

https://claude.ai/code/session_011AB4RZpph3uSFUpzKnZJcr