docs: refresh the mid-size indexing comparison
The refreshed mid-size sweep records Ktor at 13.90 s for the fork versus 7.69 s upstream, with unequal graph outputs and incomplete host observations that keep the comparison exploratory.
The README now links a dated report with all 90 accepted measurements, four rejected attempts, source/corpus identities, CPU, RSS, output hashes and identical-build fork controls. Across ten corpora, four have no clear wall-time change and six have longer fork times outside the prospective range criterion. No equivalent-work speedup or regression is claimed. The historical 2026-10-12 table and its unattainable p-value threshold remain intact.
| Corpus | Upstream median (range), s | Fork median (range), s | Fork/control noise range, s | Observed wall change |
|---|---|---|---|---|
| Ktor | 7.69 (7.69-7.93) | 13.90 (13.75-14.05) | 0.56 | +80.8% longer; unequal output |
| Exposed | 6.07 (5.84-6.18) | 9.81 (9.58-9.93) | 0.42 | +61.6% longer; unequal output |
Three accepted runs per engine and three identical-build fork controls per corpus, Node 24.22.0; fork f405d86b, upstream c71a3d03. The survey fills the missing 111,000-150,000 interval with Ktor at 144,028 processed references. Missing Windows load observation, complete cgroup counters, artifact hashes frozen before measurement and observer calibration limit interpretation.
Validation repeated on 566e3893:
- All 94 attempts independently checked against raw time receipts and preset counter rejection limits; 90 accepted keys, stable hashes for all 30 corpus/engine groups, matching fork/control output on all ten corpora.
- Existing
readme-sync.test.ts: 6/6 passed in a focused documentation workspace. The standard invocation was interrupted when global setup began an unrelated native build. Full engine suite not run for documentation-only changes. npm run check:agent-docs, whitespace check, new-report prose check and SVG XML/90-point validation passed.
README "About this fork" rows checked: resolver-pool behavior and measured-results mid-size paragraph updated; the dated 2026-10-14 full-index/sync tables and feature inventory remain accurate for their stated revisions. No feature, command or configuration key changed; no corresponding site inventory change is needed.
Related: #490 documented the old gap; #506 added adaptive pool engagement. This PR records the follow-up comparison after #488, #489 and #506, without isolating any one fix's contribution.
Summary by CodeRabbit
- Documentation
- Updated benchmark guidance with resolver-pool start criteria and the available override.
- Added a 2026-10-15 comparison of fork and upstream builds across ten corpora, reporting timing, memory, CPU, and graph results. Four corpora showed no clear wall-time change; six had longer fork times under the stated criterion.
- Documented the measurement methodology and limitations, including differing graph outputs across all ten corpora. The comparison is exploratory and does not establish equivalent-work performance changes.