docs(readme): ktor and Exposed index figures for the current fork
A full index of ktor is 26% faster and of JetBrains Exposed 27% faster at fork 5b5ac434 than at d161ef74, with 19% and 17% lower peak memory and identical graphs (codegraph init, Node 24, median of 3 alternating runs per build). The README's mid-size paragraph still quoted #515 alone (22% and 19%, measured before #517 and #518 landed), so this updates that sentence.
Change
One sentence in README.md "Measured results", in the paragraph on the 2026-10-15 mid-size comparison. It now gives the figures for the fork as it is after #515 (Kotlin companion check), #517 (shared compiled-pattern cache) and #518 (no checkpoints inside the resolution commits). It also says that both corpora still take longer than upstream e8ec7fd8 but now peak lower than it. No code changes.
Results
| Corpus, measure | d161ef74 median (range) | 5b5ac434 median (range) | Floor | Benefit |
|---|---|---|---|---|
| ktor, wall | 13.46 s (13.40-13.76) | 9.91 s (9.88-10.16) | 0.66 s | 26.4% faster |
| ktor, CPU | 52.47 s (52.18-53.09) | 36.29 s (36.24-37.29) | 1.71 s | 30.8% less |
| ktor, peak memory | 2,479 MiB (2,418-2,514) | 1,996 MiB (1,957-2,004) | 66 MiB | 19.5% lower |
| Exposed, wall | 9.55 s (9.48-9.75) | 6.95 s (6.92-7.07) | 0.16 s | 27.2% faster |
| Exposed, CPU | 36.08 s (35.95-36.53) | 26.43 s (26.38-26.63) | 0.29 s | 26.7% less |
| Exposed, peak memory | 2,427 MiB (2,402-2,502) | 2,022 MiB (1,992-2,059) | 91 MiB | 16.7% lower |
Upstream e8ec7fd8 in the same runs: ktor 7.62 s, 2,636 MiB; Exposed 5.71 s, 2,133 MiB. Upstream's graphs differ from the fork's, as in the mid-size comparison.
The floor is the spread (max - min) over six runs of two identical copies of the 5b5ac434 build. A change counts only when the medians differ by more than the floor and the two ranges do not overlap; every row above meets that. Graph hashes were identical across d161ef74 and 5b5ac434 on all 18 fork runs (ktor 57e9c5136b5e, Exposed b332758c901b). The run used 24 accepted runs, with none rejected, on a 16-vCPU WSL2 host. Each run was a systemd-run scope capped at 16 GiB, at normal priority. A run was admitted only with load under 2 and no swap.
Checks
npx vitest run __tests__/readme-sync.test.tsona493f393: 6 of 6 passed, exit 0. The chart of these rows is not attached; the table carries every figure.npm run eval:precision -- <corpus>on all 11 pinned corpora at5b5ac434(vite, vitest, svelte, flask, gin, petclinic, exposed, javalin, slim, jq, json): every case held, each exit 0. It had not run since #517, #518, #519 and the upstream merge in #520.
README rows checked
- Mid-size paragraph, ktor and Exposed sentence: moved; updated here.
- Full-index table, pretix and supabase rows (fork column 11.3 s / 3.35 GiB and 20.8 s / 3.81 GiB): not changed. Same method, d161ef74 against 5b5ac434. pretix wall showed no clear change (9.12 to 8.86 s, -2.9%, floor 0.34 s, ranges overlap). Peak memory also showed no clear change (-2.2%). supabase wall showed no clear change (-0.9%) and neither did peak memory (-0.9%). CPU fell 4.1% on pretix and 3.8% on supabase, but the table does not quote CPU. Graphs were identical.
- Exposed's 1,020 extension call links: unaffected, since the graphs are identical.
- Adaptive pool sentence (11-48% faster at 56-132% more memory against its pre-change baseline): historical, not affected.
- No site page or
docs/benchmarks/file quotes the ktor or Exposed figures.
Not covered: the full-index and sync tables compare fork ba972275 with upstream 3f714bcc (2026-10-14, Node and Bun). Refreshing them for the current fork needs that whole benchmark re-run, with an upstream build. Bun was not measured here.
Summary by cubic
Updates the README's mid-size benchmark paragraph with figures for the current fork. The previous sentence quoted only #515; after #517 and #518, the fork at 5b5ac434 indexes ktor 26% faster and Exposed 27% faster than d161ef74, with 19% and 17% lower peak memory and identical graphs. The paragraph also notes the fork still takes longer than upstream but now peaks below it.