v0.89 window admission artifact

admission.json is the v0.89 release decision for the window candidates, and samples.json contains the 60 measured samples used to derive it. No window family is runtime-enabled. Every function uses partition_recompute with a WINDOW_* reason.

The benchmark evaluated the narrow built-in ROW_NUMBER candidate over one direct keyed scan. The query used exact same-name projections and included the complete non-null identity in ORDER BY. The candidate filtered changed output rows but still recomputed the affected partition.

Rows Change State candidate median ms [range] Partition recompute median ms [range] Recompute/state ratio
1,000 Tail insert 29.941625 [28.490875, 31.011292] 12.065875 [11.525541, 13.216917] 0.4030
1,000 Front insert 92.454083 [44.982209, 152.869667] 20.670208 [18.595209, 38.816750] 0.2236
10,000 Tail insert 553.613125 [143.272875, 594.418208] 24.684708 [23.715125, 25.846125] 0.0446
10,000 Front insert 691.848292 [685.656292, 762.309333] 112.537125 [99.531458, 123.249584] 0.1627
100,000 Tail insert 1556.859583 [1311.842167, 1831.429541] 164.637041 [158.997375, 192.374791] 0.1057
100,000 Front insert 4190.417875 [4085.014458, 4430.121000] 1299.224083 [1229.935416, 1420.041042] 0.3100

The ratio is the partition-recompute median divided by the state-candidate median. Every ratio is below 1.0, so the candidate failed the 20 percent admission gate. v0.89 makes no incremental speedup claim.

Validate the repository contract offline:

python3 scripts/v0_89_release_gate.py

Reproduce the timing, WAL, and output-row samples with:

bash scripts/run_light_e2e_tests.sh \
  --package --test e2e_window_incremental_tests --ignored