diff --git a/docs/plans/2026-04-04-image-pipeline-gpu-wasm-webgl-design.md b/docs/plans/2026-04-04-image-pipeline-gpu-wasm-webgl-design.md
new file mode 100644
index 0000000..53baef7
--- /dev/null
+++ b/docs/plans/2026-04-04-image-pipeline-gpu-wasm-webgl-design.md
@@ -0,0 +1,144 @@
+# Image Pipeline GPU/WASM/WebGL Design
+
+## Goal
+
+Migrate the current CPU pixel-loop image pipeline to an incremental hybrid backend (WebGL first, optional WASM kernels, guarded WebGPU experiments) without breaking the existing worker fallback and cancellation semantics.
+
+## Current Branch Context
+
+- This branch already reduced duplicate preview work (shared in-flight preview requests, preview coalescing, opt-in histogram behavior).
+- `lib/image-pipeline/worker-client.ts` is already the stable orchestration seam for worker-first execution with main-thread fallback.
+- `lib/image-pipeline/preview-renderer.ts` and `lib/image-pipeline/bridge.ts` currently depend on CPU `ImageData` mutation, so backend switching must preserve existing return contracts (`ImageData` for preview, encoded `Blob` for full render).
+
+## Codebase Mapping Findings
+
+### `lib/image-pipeline/render-core.ts`
+
+- Contains all effect kernels as CPU loops over `Uint8ClampedArray` (`curves`, `color-adjust`, `light-adjust`, `detail-adjust`).
+- `applyPipelineStep` is a strong integration seam: per-step dispatch already exists and can route by backend capability.
+- Abort checks are periodic and lightweight; migration must preserve `AbortError` behavior.
+
+### `lib/image-pipeline/preview-renderer.ts`
+
+- Flow today: decode source bitmap -> draw 2D canvas -> `getImageData` -> apply step loops -> optional histogram.
+- This file owns preview sizing and output shape; it should remain the public API while delegating kernel execution to a backend adapter.
+- The yield loop (`requestAnimationFrame`/`setTimeout`) is CPU-friendly but should become backend-aware (GPU path should avoid unnecessary per-step yielding).
+
+### `lib/image-pipeline/worker-client.ts`
+
+- Central execution coordinator with worker transport, request dedupe, cancel fan-out, and main-thread fallback.
+- Best place to inject backend telemetry and backend selection hints (for rollout and safety).
+- Existing fallback logic (worker unavailable -> main thread) maps directly to backend fallback ladder and kill-switch controls.
+
+## Architecture Choice
+
+Adopt a **Backend Adapter + Capability Router** architecture:
+
+1. Keep existing preview/full public APIs and message formats stable.
+2. Introduce an internal backend interface for step execution and pixel transport.
+3. Select backend at runtime by capability + feature flag + step support matrix.
+4. Preserve CPU backend as correctness source and terminal fallback.
+
+### Integration Seam
+
+- New seam: `executePipelineSteps()` in a backend-router module called by both `preview-renderer.ts` and `bridge.ts`.
+- `render-core.ts` becomes the CPU backend implementation instead of the global default path.
+- Worker protocol receives optional diagnostics metadata (selected backend, fallback reason), not required for correctness.
+
+### Fallback Ladder (authoritative order)
+
+1. **WebGL2 worker backend** (preferred for preview/full when supported and flag-enabled)
+2. **WASM SIMD kernel backend** (for unsupported GPU steps or devices with poor GPU availability)
+3. **CPU worker backend** (`render-core.ts` current path)
+4. **CPU main-thread fallback** (existing `worker-client.ts` behavior)
+
+Each downgrade records a structured reason (`unsupported_api`, `shader_compile_failed`, `capability_mismatch`, `flag_disabled`, `runtime_error`) so rollout decisions are data-driven.
+
+## Context7 Guidance Buckets
+
+### Safe now
+
+- **WebGL2 shader pipeline** for color/curve/light/detail kernels with runtime feature detection.
+- **OffscreenCanvas in worker** where available, with same-origin + secure-context guardrails.
+- **WASM SIMD (single-threaded)** behind feature detection (`WebAssembly.validate` on SIMD module or runtime probe).
+- **Worker-first execution with transferable buffers** (already used in worker client).
+
+### Experimental
+
+- **WebGPU compute path** (`navigator.gpu` / `WorkerNavigator.gpu`): gate behind an explicit experiment flag and per-browser allowlist.
+- **Step fusion and mixed GPU+WASM scheduling**: only after baseline correctness/perf metrics are stable.
+- **Advanced WebGL packing optimizations** (multi-pass fusion, half-float intermediates) after parity suite exists.
+
+### Not now
+
+- **WASM threads + SharedArrayBuffer requirement**: requires cross-origin isolation (COOP/COEP) and can impact app embedding/3rd-party integrations.
+- **WebGPU-first default backend**: browser/feature variability still too high for this repository's reliability target.
+- **Full rewrite of decode/encode stack around custom codecs**: exceeds scope; current `drawImage` + canvas/blob path remains.
+
+## Phased Rollout
+
+### Phase 0: Foundation + Instrumentation (no behavior change)
+
+- Add backend router with CPU-only implementation.
+- Add telemetry points for preview/full latency, fallback reason, and error class.
+- Add golden-image parity harness for all 4 current step types.
+
+### Phase 1: WebGL PoC (preview only)
+
+- Implement WebGL2 backend for `curves` and `color-adjust` first.
+- Enable by dev flag on preview pipeline only.
+- Validate output parity against CPU baseline under tolerance thresholds.
+
+### Phase 2: WebGL expansion + guarded production rollout
+
+- Add `light-adjust` and `detail-adjust` kernels.
+- Enable progressive rollout (1% -> 10% -> 25% -> 50% -> 100%) with kill switch.
+- Keep full render on CPU backend until preview parity and crash rate are stable.
+
+### Phase 3: WASM SIMD fallback kernels
+
+- Add WASM SIMD implementation for hot CPU kernels as middle rung between WebGL and CPU.
+- Use when WebGL unavailable or disabled, still inside worker.
+
+### Phase 4: Full-render backend adoption
+
+- Route full render through same backend router.
+- Keep output encode path unchanged.
+- Roll out in smaller increments than preview due to export-critical path.
+
+### Phase 5: WebGPU experiment track
+
+- Optional experiment branch with strict allowlist and observability.
+- No default enablement in this plan.
+
+## Risk Controls
+
+- Feature flags:
+  - `imagePipeline.backend.webgl.enabled`
+  - `imagePipeline.backend.wasm.enabled`
+  - `imagePipeline.backend.webgpu.experiment`
+  - `imagePipeline.backend.forceCpu`
+- Auto-disable circuit breaker when fallback/error thresholds exceed SLO windows.
+- Golden-image regression suite enforced in CI before any rollout increase.
+- Keep worker protocol backward compatible until migration fully lands.
+
+## Metrics and Success Criteria
+
+- Preview latency (`p50`, `p95`) for representative step stacks and image sizes.
+- Full render latency (`p50`, `p95`) for default export sizes.
+- Main-thread long task count during rapid slider edits.
+- Backend selection distribution (% webgl/wasm/cpu/main-thread).
+- Fallback and runtime error rate per backend.
+- Pixel parity drift (max absolute channel delta and histogram similarity).
+
+Success target for initial migration:
+
+- >=30% preview `p95` improvement on supported devices.
+- No increase in user-visible render failures.
+- <=1% forced downgrade due to runtime backend errors after rollout stabilization.
+
+## Out of Scope
+
+- New visual adjustment features.
+- Server-side/offline rendering architecture changes.
+- COOP/COEP rollout for SAB-threaded WASM.