Stream G: run the front end once per Sky.Spa build

Machine: 16 GB Mac. Release builds. Median of 3 warm runs. Parent = the sky build process (its own SKY_TIMINGS table). base = perf/build-speed HEAD c4d17f467.

Whole build, N=200

apppathfront half (parent)split stepbuild wallparent peak RSSfront-end note
genwhole 200, direct Spa entrybase0.65 s0.62 s4.70 s328 MB2 loads, 2 checks
genwhole 200, direct Spa entrynew0.66 s0.15 s4.20 s330 MB1 load, 1 check
bench-gen 200, direct Spa entrybase0.28 s0.19 s1.54 s155 MB2 loads, 2 checks
bench-gen 200, direct Spa entrynew0.28 s0.05 s1.45 s153 MB1 load, 1 check
genwhole 200, --target web:appbase0.67 s0.62 s4.78 s328 MB2 loads, 2 checks
genwhole 200, --target web:appnew0.66 s0.62 s4.79 s329 MB2 loads, 2 checks
bench-gen 200, --target web:appbase0.28 s0.20 s1.62 s151 MB2 loads, 2 checks
bench-gen 200, --target web:appnew0.28 s0.20 s1.63 s153 MB2 loads, 2 checks

"Direct Spa entry" = the staged .skyapp/web-app project (the synthesised Spa.app entry) built as its own project. Base counts are by reading the code (the base binary has no counter). Neither generated app keeps a continuation on the server, so the shared second check does not show in these rows; the fixture test shows it (2 checks -> 1).

An earlier cut held the front half's CheckOutput.def_types alive through the split. The web:app parent peak then rose to 385-396 MB in most runs (a control binary with no behaviour change measured 327-331 MB). The per-def types are now dropped before the split (no analysis reads them); peak is back to 329 MB.

Byte identity

Diagnostics

111 runs, base vs new, stdout + stderr + exit, durationMs masked: 0 differ.

Tests