refactor(benchmarks): centralize workload policy

[thejustinwalsh]

Summary

• make one typed catalog authoritative for all nine live benchmark workloads, including Main and Presentation defaults, font policy, controls, interaction, preload, and surface kind
• normalize unknown workload URLs within the selected mode before font and control policy execute
• derive runtime resets and benchmark workload metadata from the catalog instead of parallel string switches
• move technique-invariant text styling and color motion below workloads/shared, with a boundary test rejecting static and dynamic renderer imports from workload modules
• classify loader/Worker parity as conformance and isolate direct baker/shaper Wasm imports behind one shared selected-ABI adapter
• record the benchmark-driven public API audit and defer grouped multi-Text publication until a delayed-peer probe proves a product need

Why

Workload identity and defaults were duplicated across URL parsing, runtime state, controls, and navigation. An unknown or newly added workload could therefore render a fallback scene while retaining unrelated font or control state. Example modules also depended back on renderer implementation helpers, and the low-level target hierarchy described a shared conformance/measurement ABI adapter as measurement-only.

This change establishes the policy and dependency boundary before the remaining workload scene and route extraction. Renderer, canvas, RAF, GPU timing, and telemetry ownership stay inside the benchmark's persistent host; no renderer-specific package API is added.

Verification

• mise exec -- pnpm benchmarks:check
• mise exec -- pnpm docs:check
• 49 benchmark test files / 278 tests passed
• benchmark lint, formatting, type checks, package-size gates, and production build passed
• focused workload boundary and shared-helper suites passed after the second commit
• both commits are SSH-signed with the configured maintainer identity

Stacked on #17.