benchmarks: cover retained renderer workloads

[tritao]

Builds on #74 and completes the workload/contract coverage for the current
retained-performance stack.

Purpose

The earlier benchmark foundation provides measurement tools, but performance
contracts need scenes that resemble the renderer's important scaling cases:
large incremental edits, repeated geometry, staged assemblies, selection, and
picking. Without shared workload definitions, two renderer paths can appear
comparable while doing different semantic work.

This PR expands the shared workloads and turns the architectural expectations
of the optimization stack into deterministic, non-timing assertions. It adds
coverage and diagnostics rather than another renderer optimization.

What changed

• Add focused incremental mutation workloads and batch-size curves.
• Model expanded, shared-source, recipe-based, and staged assemblies.
• Add closest-pick, depth-stack, hover-style refresh, and selection scaling
  workloads.
• Report visual submission phases and ordered-submission opportunities.
• Share the same scenes between benchmarks, parity tests, and the workload
  viewer.
• Document how to run, pin, compare, and interpret benchmark results.

Performance contracts

Wall-clock results remain informational. Tests instead assert properties such
as whether an eligible edit avoided a full draw-list rebuild, whether plan or
resource revisions changed, whether application-owned framebuffers were
preserved, and whether all renderer paths produced equivalent output.

These contracts make later performance changes reviewable without depending
on noisy CI timing thresholds.

Validation

• Mutation and interaction curves across compatibility and core profiles.
• Workload parity and interactive-viewer smoke coverage.
• Cache-invalidation and application-owned framebuffer invariants.
• Workload suite exercised by the full top-stack platform CI and CodeQL run.