Require measured WebGPU frame completion and reject missing performance data

[qqrm]

The performance workflows passed with no FPS samples; timing ignored render errors and counted submission without waiting for GPU completion. This batch requires real completed-frame measurements and fails CI on absent/invalid data or a per-workload median regression over 30%.

• Run deterministic 1k/50k/1m source-candle workloads with production LOD, 10 warmup frames and 5 x 30 measured frames per case. Await GPU readback completion and reject no-op renders, invalid inputs and GPU errors.
• Measure a pinned reviewed reference and the candidate on the same hosted runner, using identical browser, compiler and SwiftShader settings. Keep separate logs/sample sets and record both code SHAs plus environment provenance. Cross-host absolute calibration remains historical evidence only.
• Require the reusable read-only benchmark in ci-gate; remove the duplicate workflow, reject missing/invalid baselines, and refresh compatible cache/setup-node pins.

Validation: formatting, locked all-feature checks, strict Clippy, actionlint and 10 Python regression tests passed. Hosted CI 38682129385 passed on ed16cfc: 117 browser Rust, 1 Node Rust, 7 Playwright, 10 Python tests; reference and candidate each passed 2 GPU tests and produced all 15 samples. Paired median changes were -0.01%, +9.09%, +0.61%; all security and CodeQL checks passed. Local cargo test builds WASM but cannot execute it directly on native Windows, so runtime execution uses hosted browsers.

Scope: cached LOD chart throughput with readback overhead on software SwiftShader. Physical GPU/display FPS, initial loading and live-update cost are not established. PR #407 contains additional major dependency migrations and is not superseded.

Code: F:tests/performance_limit.rs; F:scripts/parse_perf_log.py; F:scripts/compare_fps.py; F:.github/workflows/benchmark.yml; F:.github/workflows/PIPELINES.md.