test: unify testing strategy, CI parity, high-value coverage + gl-deprecation hardening

[DennisSmolek]

Summary

Unifies and standardizes how R3F is tested, closes the biggest coverage gaps on the new v10 surfaces, and hardens one deprecation bug found along the way. Builds on the recent alpha.3 work (scheduler extraction, useTextures, the multi-canvas/HMR/background fixes).

What's here

Canonical testing guide — docs/development/TESTING.md rewritten as the single source of truth: a three-tier model (Tier 1 unit+mock in CI, Tier 2 browser/GPU pre-release gate, Tier 3 headless spike), local↔CI parity rules, test-organization conventions, coverage policy, and an open roadmap.

CI / local parity

• verify-bundles + verify-types now run in test.yml (they were local-only via pnpm ci — CI could drift)
• text-summary coverage reporter + coverage/ uploaded as a build artifact

High-value tests (~120 new; overall lines coverage ~78%)

• Scheduler integration — render-phase takeover + fps end-to-end (drop: true/false), as integration with the external @pmndrs/scheduler (not its internals)
• Canvas size control / textureColorSpace / gl deprecation warning path
• Multi-canvas pure-JS — canvasRegistry, renderer config-bag parser, ScopedStore Proxy semantics
• Events — interactivePriority, XR register/unregister, frame-timed edges (+ resolved a pre-existing it.todo)
• WebGPU hook lifecycle via the WebGPUContext mock (useUniforms/useNodes/useBuffers/useGPUStorage/useRenderPipeline, now 62–90%) + useRenderTarget. GPU-only paths are left as it.todo tagged for Tier 2.
• state.clock removal contract (v10 regression guard)

Refactor (behavior-preserving, for testability)

• Extract the Canvas renderer config-bag parser into utils/parseRendererConfig.ts (mirrors parseBackground.ts; not re-exported)

Fix

• Harden the state.gl deprecation heuristic (utils/isInternalRendererAccess.ts): inspect only the immediate caller frame with path-independent markers, replacing a brittle hard-coded repo-path substring that mis-fired for installed packages and any non-react-three-fiber-named checkout (R3F's own <Environment> read spuriously warned; in the canonical repo it suppressed all access).

Notes for reviewers

• Two real source behaviors are captured as passing regression tests, not changed: NoColorSpace is unselectable via the config bag ('' || SRGBColorSpace), and the state.gl getter only lives on the initial store object.
• 6 it.todos are all genuinely GPU-dependent (Tier 2), not skipped work.

Test plan

• pnpm test → 43 files, 532 passed / 6 todo / 0 failed
• pnpm typecheck, pnpm eslint, pnpm format all pass
• pnpm build && pnpm verify-bundles && pnpm verify-types pass

Not included / follow-ups

• Tier-2 Playwright harness (pnpm test:gpu) — strategy documented, harness not built
• Coverage floor — deferred to stable (per roadmap)
• Tier-3 headless-WebGPU CI spike

🤖 Generated with Claude Code