feature: add deterministic editor performance lab

[SimonSiefke]

Summary

• measure Enter-to-DOM-ready and paint confirmation inside the renderer
• compare fresh-process samples in randomized same-runner ABBA/BAAB blocks
• disable core AI work, record process manifests, and reject bundled Copilot
• fingerprint optimized executables, bundles, source maps, commits, and dirty state
• collect source-mapped deterministic work separately and aggregate hosted replicas hierarchically
• fail closed on invalid scoring while permitting non-claiming mechanism exploration on underpowered hosts

Fresh-agent trials

The first fresh agent retained no patch. Identical builds appeared 4.69% different with a CI of -13.56% to +3.60%, and bundled Copilot was present despite --disable-extensions.

After refinement, a 24-sample-per-arm A/A correctly became invalid: -2.44%, CI -11.52% to +5.82%. Its 8.67% interval half-width cannot detect a 5% target. Copilot was absent and semantic code/* phases were populated.

A second fresh agent then used deterministic mechanism exploration without delaying behavior. Across three exact runs per arm it found:

• breadcrumb false focus resets: 3 to 0
• breadcrumb reveal calls: 4 to 1
• no-item focus-event allocations: 3 to 0
• editor option update/clone/migration/registry passes: 3 to 2
• migration reads: 117/90/51 to 78/60/34

Every claimed counter had 0% variation. The combined action ceiling was below 0.4%, far below local timing resolution, so the patch was saved as a mechanism candidate and reverted rather than described as a speedup.

The trial refined the loop further:

• deterministic candidates can be queued on noisy hosts, but cannot become proxy-win or UX claims
• timing is deferred until a compatible batch approaches calibrated detectability
• tracker arms are grouped to avoid repeated 2.5-3 minute transformations
• unchanged noisy counters are ignored; unstable counters claiming a delta invalidate the mechanism
• the scoped optimized desktop build and source-map audit procedure are documented
• hosted work attribution now includes breadcrumbs, editor configuration/migration, text-file readonly setup, and inline completions

Verdict and validation

A proxy win requires less stable deterministic work and a latency CI upper bound at or below +2%. A UX-confirmed result additionally requires a renderer-latency CI excluding zero. Profiles and instrumented samples never enter scoring.

• performance-lab: 30 tests passed
• workflow/build package: 94 tests passed
• launch-worker: 31 passed, 3 skipped
• test-worker: 10 passed
• affected TypeScript checks and repository formatting passed

This PR changes the benchmark and performance lab only. It claims no VS Code speed improvement and retains no VS Code patch.