Summary

Background

Every worker process statically imports the coverage module chain (7704.js → 4411.js), loading ~108KB of picomatch/glob code at startup — even when coverage is disabled (the common case).

Implementation

• Changed createCoverageProvider from a top-level static import to a dynamic import() guarded by coverage.enabled.
• This removes 7704.js (1.6KB) and its transitive dependency 4411.js (106KB) from every worker's static import chain. These chunks are now only loaded once in the main process when coverage is actually enabled.

Worker static JS asset reduction:

+------------------+----------+----------+---------+
| Chunk            | baseline | cov-lazy | Δ       |
+------------------+----------+----------+---------+
| worker.js chain  | 262KB    | 163KB    | -99KB   |
| 7704.js (cov)    | 52 PIDs  | 1 PID    | -98%    |
| 4411.js (glob)   | 54 PIDs  | 1 PID    | -98%    |
+------------------+----------+----------+---------+

Benchmark

Fixture: 50 test files × 20 tests (1000 tests), 14-core machine, isolateWorkers: true, NODE_COMPILE_CACHE disabled.

hyperfine --warmup 5 --runs 30

Benchmark 1: baseline
  Time (mean ± σ):     571.0 ms ±  15.9 ms    [User: 3478.4 ms, System: 709.6 ms]
  Range (min … max):   546.9 ms … 616.1 ms    30 runs

Benchmark 2: cov-lazy
  Time (mean ± σ):     558.9 ms ±  25.5 ms    [User: 3264.9 ms, System: 716.2 ms]
  Range (min … max):   533.6 ms … 637.5 ms    30 runs

Summary
  cov-lazy ran
    1.02 ± 0.05 times faster than baseline

Wall clock: ~2% (limited by 14-core parallelism). User CPU: -214ms (-6.1%), reflecting ~4ms less parse/compile per worker × 52 worker processes.

User Impact

Marginal wall-clock improvement for typical runs; more visible on lower-core machines or larger test suites.

Checklist

• Tests updated (or not required).
• Documentation updated (or not required).