Measure software performance comprehensively

Measure ship-studios performance (import/startup cost, offline test-suite timing, the DSP benchmark + scaling + memory, perf-instrumentation overhead, CLI cold-start) on a SINGLE SERIAL measure pass so the numbers stay honest, then fan out one analysis agent per surface to read the code against those measured numbers, adversarially verify every finding, and synthesize a prioritized report. Re-runnable: it re-measures the repo each run rather than trusting baked-in numbers.

Perf fixes where file groups are disjoint and can edit in parallel without race conditions.

• · structured data

• · analysis results

• · Claude Code agent runtime (parallel/fan-out)

• · Artifact must conform to specified schema
• · Claude Code workflow harness

• · Blocking dependencies unresolved causes topo-sort failure
• · Agent returns data not matching schema → structured rejection

• · Adversarial cross-check phase
• · Independent verification lens