test(playtest): add PRD-564 resolution measurement gate

[jonit-dev]

Adds a measurement-only harness for PRD-564, where resolution changes may recreate render targets and pipelines. It reuses the authored starter bootstrap, checks actual runtime/adapter identity, records first-frame and following-window census data, and computes a fail-closed gate. Runtime optimization is not implemented.

Validation: 53 nearest probe/startup-ready tests and playtest TypeScript previously passed; current whole-repository lint passes. Normal pre-push agents/docs/lint/drift checks passed; these do not prove runtime behavior. A prior hardware preflight passed functional checks only on named NVIDIA Turing at pinned scale 1. The current full measurement reached its 240-second deadline with 0/3 contexts and 0/9 steps completed, no gate verdict, and unqualified host timing. Descriptor byte sums are not physical memory measurements.

The separate lint repair preserves the donor's fixed public/math-schema property names through an exact-file naming override. Isolated containment checks passed, while nonallowlisted names and other paths remain checked. Donor source is unchanged.

Progress remains 0%. No optimization, native parity, cost improvement or visual verdict is claimed. Visual comparisons/judging remain open; draft/skipped CI is not a passing runtime qualification.

TL;DR — ResolutionScaler (packages/core/src/resolution-scaler.ts) picks one of 10 rungs, from 1.0 to 0.23. Its own comments say why the rungs are coarse: "every step reallocates render targets on WebGPU, and fine granularity buys smoothness with hitches" (:20-24), and "the resize reallocates every render target, which rebuilds pipelines" (:72-73).

0% by the PRD's own boxes: 0/3 phases (0/5 boxes), 0/2 acceptance criteria. PRD: docs/PRDs/unreal-source-borrowing/PRD-564-the-resolution-scale-changes-without-reallocating-a-target.md.

Phases

Phase 1: Price the resize — 0/1

• Three forced steps (1.0 → 0.85 → 0.61 → 1.0) on the starter record the step frame, the next 30 frames, pipelines built and bytes reallocated, over 3 runs. proof: node packages/playtest/dist/runner/cli.js perf on the scenario, results under ## Decisions with the continue or decline call.

Phase 2: Draw into the sub-rectangle — 0/2

• A scale change sets the scene pass viewport and scissor and allocates nothing; a canvas resize still reallocates once. proof: red-green pnpm exec vitest run packages/core/__tests__/renderer*.spec.ts with an allocation-counting renderer stub.
• Each built-in chain stage samples inside the sub-rectangle, and a stage without support reports refused and keeps the reallocation path. proof: red-green chain spec, one case per stage in RENDER_CHAIN_STAGE_ORDER.

Phase 3: Finer steps and native — 0/2

• The controller takes continuous fractions (quantized to 8 px, 2% threshold, at least 8 frames apart, amortized up-steps), and PRD-384's existing scaler suites still pass. proof: red-green pnpm exec vitest run packages/core/__tests__/resolution-scaler*.spec.ts.
• The native desktop host draws the same sub-rectangle frame as the browser. proof: pnpm parity --target desktop --only-tests resolution-subrect.

Acceptance criteria

• AC-1 [local]: On the starter in browser WebGPU (RTX 2080), a forced scale step allocates no texture and builds no pipeline, and the step frame stays within the window's p95. proof: node packages/playtest/dist/runner/cli.js perf with a forced-step scenario, TN_FRAME_BUDGET and the pipeline census. — Evidence: pending.
• AC-2 [local]: A frame at scale 0.72 through the sub-rectangle matches the same frame through today's reallocation path within the visual gate's threshold. proof: pnpm visuals:ab. — Evidence: pending.

Still open

• Phase 1: Price the resize: Three forced steps (1.0 → 0.85 → 0.61 → 1.0) on the starter record the step frame, the next 30 frames, pipelines built and bytes reallocated, over 3 runs. proof: node packages/playtest/dist/runner/cli.js perf on the scenario, results under ## Decisions with the continue or decline call.
• Phase 2: Draw into the sub-rectangle: A scale change sets the scene pass viewport and scissor and allocates nothing; a canvas resize still reallocates once. proof: red-green pnpm exec vitest run packages/core/__tests__/renderer*.spec.ts with an allocation-counting renderer stub.
• Phase 2: Draw into the sub-rectangle: Each built-in chain stage samples inside the sub-rectangle, and a stage without support reports refused and keeps the reallocation path. proof: red-green chain spec, one case per stage in RENDER_CHAIN_STAGE_ORDER.
• Phase 3: Finer steps and native: The controller takes continuous fractions (quantized to 8 px, 2% threshold, at least 8 frames apart, amortized up-steps), and PRD-384's existing scaler suites still pass. proof: red-green pnpm exec vitest run packages/core/__tests__/resolution-scaler*.spec.ts.
• Phase 3: Finer steps and native: The native desktop host draws the same sub-rectangle frame as the browser. proof: pnpm parity --target desktop --only-tests resolution-subrect.
• Acceptance Criteria: AC-1 [local]: On the starter in browser WebGPU (RTX 2080), a forced scale step allocates no texture and builds no pipeline, and the step frame stays within the window's p95. proof: node packages/playtest/dist/runner/cli.js perf with a forced-step scenario, TN_FRAME_BUDGET and the pipeline census. — Evidence: pending.
• Acceptance Criteria: AC-2 [local]: A frame at scale 0.72 through the sub-rectangle matches the same frame through today's reallocation path within the visual gate's threshold. proof: pnpm visuals:ab. — Evidence: pending.

_{Generated from the PRD's checkboxes. Regenerate on each push so the PR and the PRD never disagree.}