Guest editor first compilation — 2026-10-11

The checked-in guest edit.kt program reproduces the first-build pattern observed in PID: preparation removes most cold work, but the first project compilation still takes roughly a second. A synthetic Hello World moves more work before that compilation without demonstrating a reduction in total work.

Method

Use the same real subprocess probe, timings, alternating trial order, JVM settings, and frozen compiler payload as the PID first-build report. Three fresh workers per variant compile the editor five times. The source snapshot is system/programs/edit.kt last changed at a4a9997122116d118b51a20f3d106da801984d60: 614 lines, 19,260 bytes, SHA-256 430233e599ace8858740b1f015bb125b0a171a34a8fdd5d07a475ff90e741348. The probe submits it at virtual path src/edit.kt; this path participates in diagnostic metadata.

The compiler payload matches sources at aa748215, before the concurrent reference-array changes, while the editor source includes its newer cursor/paste changes. All runs use exactly that combination, which compiled successfully. This is not a measurement of the latest complete checkout or a comparison to historical editor size measurements with different source and platform identities.

The editor selects stdlib:core and compukter:core, including their dependency closure, without Sable or Propulsion addon bundles. PID selected those addons, so its preparation time should not be interpreted as a same-library comparison with this editor. Compiler startup includes the real handshake; compilation includes IPC and result decoding, excluding source collection, dependency resolution, output writes, and Minecraft UI.

Variants are cold compilation, library preparation followed by compilation, and library preparation plus a synthetic Hello World compilation followed by compilation. Environment: Linux x86-64, Ryzen 9 9950X3D, powersave CPU governor, JDK 21, 2 GiB worker heap, 512 MiB metaspace. Frequency, affinity, and background activity are uncontrolled.

Results

Milliseconds, with separate phase medians over three trials. The serial-total median is calculated from each trial’s actual sum, not by adding column medians. Repeated compilation covers samples 4–5, six observations per variant. Three first compilations per variant are insufficient for tail-latency claims.

After the synthetic compile, the first editor compilation is approximately 390 ms faster than after library preparation alone, a 41% reduction. The synthetic compile itself takes approximately 649 ms. The measured serial total is higher, so this is latency shifted into potential background time, not a demonstrated total-work optimization. Existing IDE preparation already runs asynchronously. The experiment does not test the CPU interference or perceived responsiveness of adding another background compilation.

Verification and reproduction

All 45 editor outputs are byte-identical: 151,112 bytes, SHA-256 0874242b0a24e76ed3d386c9570f2bacb69a876c4c9a6692af1b031b6fd6e23b. All 45 editor outputs and all three Hello World outputs pass the native ABI-22 stateless artifact verifier. Terminal interaction, file editing, and in-game behavior were not exercised. No production code or protocols changed, and no production test suite was rerun for this documentation-only stage.

Samples, summary, provenance, and verification accompany nine raw logs. Use FirstBuildProbe.kt from the PID first-build report with this frozen editor source copied to src/edit.kt in a scratch project. Pass project root, fresh output directory, and frozen tooling payload root; omit addon arguments. Set BENCH_MODE to cold, prepared, or hello, and start a fresh JVM per trial.