Cold usage history: lower elapsed time, higher CPU and memory

Issue #36036. Product base 74573fc29b, candidate beabb1bf3e. One clean Linux VM, 8 CPUs and 16 GB RAM; six paired rounds after base/base calibration. A generated 15.52 GB corpus preserves representative transcript sizes and line shapes. No original transcript content was used.

Cold requested-history read

The whole command includes cold reading, restart, append, eviction and test-runner startup. Its paired median elapsed change was −13.5%, with the middle half −15.9% to −8.9%. The helper calls this a real change: median exceeds its 10% floor and the interval does not cross zero. Two of six pairs were inside 10%, so this is a modest improvement, not a guarantee for every run.

Whole-command CPU cycles rose 33.2% and instructions rose 12.6%. Parallel parsing trades more aggregate work and about 1.8 GB more peak memory for lower elapsed time. Completed cold medians were 14.710 s and 9.9055 s. Completed restart was 0.7845 s and 0.816 s; completed append 0.9625 s and 0.932 s. Unchanged warm reads remained near 0.2 s.

What was verified

Base/base elapsed calibration stayed within noise: +4.7%, middle half −4.6% to +9.7%. All 32 cold-read checks passed with fixture-page eviction and substantial physical input. All eight fragmented-line files, from 256 KiB through 8 MiB for Claude and Codex, were opened. Cold, restart and warm summaries matched at 116 buckets; append summaries also matched and restored byte-for-byte.

One-core, two-core and a 2 GB cgroup stayed inline and retained identical totals. Capacity sizing is a startup heuristic, not a hard RSS ceiling. The pool leaves one core, caps at seven workers and allocates at most a quarter of reported available memory using the measured per-worker allowance. Workers have a 2 GiB heap ceiling, and unusually large reducer state can exceed the usual resident footprint.

Both usage parsing and document extraction passed on the same VM through development source, npm dist, a real Node SEA, Electron as Node and Electron ASAR. This was Linux x86_64. Native macOS, Windows, WAN and mobile were not tested.

The parent’s separate real-client pass observed seven workers without fallback, nonempty usage, refresh/reopen and all workers reclaimed after idle. Its timings were uncontrolled functional observations and are not used for the speed claim. A real Node-watch failure was reproduced and fixed by removing watch-reporting instrumentation from child environments only.

Review and remaining limits

Council scores by round, in Opus/Sol/Grok order: 7/8/6, 8/8/8, 9/9/9, 9/9/9. Final check:changed passed 131 tests and affected format, lint, knip, GraphQL and typechecks. The targeted proof audit supports filing with these qualifications. The real PR #36124 is open and unmerged; hosted checks and review are pending.

Permanent inline fallback after a worker failure, no job deadline for stuck filesystems, sampled capacity and compressed/tiny-append routing remain explicit limits. This change does not reduce the total CPU work or promise a fixed multiplier. The owned benchmark VM was deleted and verified absent.

Created with GPT 6.1 Sol, Claude Opus 5.5 and Grok 4.7 in T3 Code. This PostPlan copy uses black and system-font fallbacks because the active host theme could not be read.