[softmarshmallow]

Summary

Adds a repeatable way to measure native and browser viewport work before changing performance-sensitive code. The harness builds the production hosts, warms up a fixed camera workload, preserves raw frame evidence, and produces CLI/JSON comparisons across independent runs.

• Define three measured scopes: the full editor, the viewport with editor projection/feedback but no egui work, and the renderer without editor projection/feedback. Keep the same physical scene rectangle across modes and verify execution with counters and exact compositor tests.
• Organize optional instrumentation in src/measurement/, host adapters beside each host, and automation/reporting/tests in tools/benchmark/. Ordinary builds exclude the measurement feature and browser measurement page.
• Add the bounded CC0 Chess Set fixture with pinned sources, reproducible GLB packing, attribution, and automated integrity checks. TESTING.md indexes the workflows, methodology, upstream suites, and fixture policy.
• Add View → Developer → Show FPS Meter: a passive, click-through readout at the app window's bottom-right. It reports application submission cadence without adding a timer, repaint loop, or GPU wait.

Related to #2. This delivers the testing and measurement foundation; profiling attribution and optimization remain deferred in that issue. CPU API timings do not establish GPU completion, displayed FPS, or physical-input latency. Performance captures remain opt-in and run in isolated browser processes.

Validation

• just verify: formatting, Clippy, tooling tests, native Rust tests, exact guide replay, measurement contracts, and scene-only compositor parity.
• just web-verify and just benchmark-web-check: ordinary WASM build plus measurement-enabled WASM Clippy. The latter is also added to hosted CI.
• just ci-docs check: 28 guide features and 136 exact Linux artifacts; updated native/Linux meter images reviewed.
• Real native/browser captures cover all three execution modes. Final FPS placement was checked in isolated headed Chromium; ten additional meter-off/on stationary Chess runs recorded the same 8.3 ms median cadence. This is a bounded check, not a zero-overhead or speedup claim.

All four hosted CI lanes passed on the PR commit: Ubuntu checks, exact Linux guide replay, macOS bundle/native integration, and WASM builds including the optional measurement adapter.