detect: no cliff past ten thousand open correlations; .not() on a named pattern

[cpoder]

What changes

The engine is pinned at varpulis 1cb201b, which brings in two engine PRs.

• perf: the watermark sweep visits only the partitions with a deadline due varpulis/varpulis#289: no cliff past ten thousand open correlations. The engine swept every partition's open runs on each event, to expire them and confirm absences. It now keeps their deadlines in order and visits only the partitions with one due.
• fix: .not() cancels a run of a named pattern varpulis/varpulis#288: .not(C) cancels on a named pattern. On a stream that reads a named pattern, .not(C) did nothing, so Unacked.not(Cancel) still raised cancelled orders.

Measured: bench/detect.sh, PROGRAM=sequence

Runs alternated between the v0.3.4 runtime and this build, in the same session on the same box, two runs each:

Open at once	v0.3.4	this build
1	28 400 – 28 500/s	28 600/s
100	27 100 – 27 200/s	28 200 – 28 600/s
1 000	19 500 – 19 600/s	25 900 – 26 200/s
10 000	11 900/s	21 600 – 22 600/s
30 000	2 075 – 2 085/s, 145 MB	14 200 – 14 900/s, 88 – 109 MB
100 000 (400 000 events)	not run	11 100 – 11 200/s, 448 MB
1 000, unpartitioned	4 280 – 4 290/s	4 250 – 4 260/s

Every run was complete, with every pair matched (32 000 alerts, or 200 000 at a hundred thousand open).

The multi-unit table (1 000 open per unit) is re-measured too: 27 100/s on one unit, 37 200 – 39 400/s on two, 48 300 – 48 900/s on four. The stateless ceiling is unchanged, at 41 000 – 43 000/s with 2 or 4 units.

Docs

bench/README.md has the new table, and the before/after comparison replaces the cliff paragraph. The sizing is updated. concepts/detects.md now says what it costs, and that memory, about 4 KB an open sequence, is what bounds a unit.

Checks

• e2e/detect: every invariant holds (D1 to D9).
• e2e/detect-demo: every beat holds.

🤖 Generated with Claude Code

https://claude.ai/code/session_01N3K1TGnWTvwYzt9rKuJXES