Imageflow3

.gitignore

@@ -351,3 +351,9 @@ docker/imageflow_bench_ubuntu20/results/vips_2000x2000.jpg
 docker/imageflow_bench_ubuntu20/results/vips_2000x2000.png
 docker/imageflow_bench_ubuntu20/results/vips_reference_2000x2000.png
 external/
+
+# Fuzz corpus and artifacts (auto-generated by cargo-fuzz)
+/fuzz/corpus/
+/fuzz/artifacts/
+/fuzz/coverage/
+/fuzz/Cargo.lock

CONTEXT-HANDOFF.md

@@ -0,0 +1,93 @@
+# Imageflow3 Context Handoff
+
+## State: 132/193 tests passing (57 failures, 4 ignored)
+
+Branch `imageflow3`. Both v2 and zen backends run against shared v2 golden checksums.
+
+## Test commands
+
+```bash
+just test                        # both backends, shared v2 golden
+just test-filter NAME            # filter by test name
+ZENPIPE_TRACE=1 just test-filter NAME  # with pipeline trace
+```
+
+## Root cause of remaining 57 failures
+
+### 1. JPEG decoder difference: delta=49 (affects ~30 tests)
+
+zenjpeg produces different pixels than mozjpeg despite LibjpegCompat chroma upsampling being set.
+
+**Proven facts:**
+- LibjpegCompat chroma config IS reaching zenjpeg (trace confirmed `StripProcessor::new chroma_upsampling=LibjpegCompat`)
+- CMS is NOT the cause (moxcms sRGB→sRGB is identity, delta=0, proven in `/home/lilith/work/moxcms/tests/srgb_roundtrip.rs`)
+- No-op resize is NOT the cause (removed, trace confirms Source→Output with no Resize)
+- The Canon JPEG may be progressive (buffered mode) where chroma config might not apply
+- The delta is purely from JPEG decode differences
+
+**Next step:** Run the Canon 5D JPEG through both mozjpeg and zenjpeg (LibjpegCompat mode) in isolation, compare raw pixels. If delta>0, the bug is in zenjpeg's LibjpegCompat implementation. Check if the JPEG is progressive and if buffered mode respects chroma config.
+
+### 2. CMS/ICC differences (affects ~15 ICC decode tests on top of JPEG delta)
+
+Wide-gamut profiles (Adobe RGB, P3, ProPhoto, Rec.2020) go through CMS on both sides. Both use moxcms. But JPEG decode differences get amplified by the CMS transform — different source pixels → different CMS output.
+
+**Proven:** ICC profile bytes are extracted identically on both sides. Both backends apply moxcms. The delta is from decode, not CMS.
+
+### 3. Trim detection (5 tests, score=0)
+
+Zen uses corner-color comparison, v2 uses Sobel-Scharr edge detection. Different algorithms → different crop bounds. Fix: implement Sobel-Scharr in zenpipe.
+
+### 4. Watermark compositing (6 tests, 5-8% differ)
+
+Watermark compositing is pixel-identical for synthetic inputs (proven with red-on-green and red-alpha-on-blue tests, delta=0). The integration test differences come from JPEG decode differences in the watermark source image.
+
+### 5. WebP alpha (2 tests)
+
+zenwebp vs libwebp decoder alpha differences.
+
+### 6. EXIF alpha normalization (1 test)
+
+`crop_exif` — RGB identical, alpha=255 on 24% pixels. The `alpha_meaningful` flag isn't propagating correctly for Crop+Within pipeline.
+
+## Zennode Consolidation (2026-03-29)
+
+Node schema ownership changed — affects imports in `imageflow_core/src/zen/`:
+- **zencodecs::zennode_defs** — all codec encode/decode, Quantize, QualityIntentNode (16 nodes)
+- **zenpipe::zennode_defs** — Constrain, Resize, CropWhitespace, FillRect, RemoveAlpha, RoundCorners (6 nodes)
+- **zenfilters::zennode_defs** — all 43 filter nodes (unchanged)
+- Individual codec crates (zenjpeg, zenpng, etc.) no longer export zennode_defs
+- Constrain node renamed fields: sharpen→unsharp_percent, lobe_ratio→kernel_lobe_ratio, added matte_color
+
+## Architecture
+
+### Zen module structure (`imageflow_core/src/zen/`)
+- `execute.rs` (1202 lines) — pipeline execution, decode, encode
+- `cms.rs` (521 lines) — ICC/gAMA/cICP transforms
+- `translate.rs` (600 lines) — v2 Node → zennode translation
+- `converter.rs` (420 lines) — NodeConverter implementations (white_balance, color_matrix, region, expand_canvas)
+- `watermark.rs` (600 lines) — watermark compositing with zenresize
+- `preset_map.rs` (354 lines) — EncoderPreset → CodecIntent
+- `nodes.rs` (90 lines) — custom NodeInstance types
+- `color.rs` (70 lines) — shared color parsing
+- `context_bridge.rs` (163 lines) — v2 JSON → zen pipeline
+- `riapi.rs` (153 lines) — RIAPI expansion
+
+### Key decisions made this session
+- Zen compares against v2 golden (no separate `_zen` baselines)
+- `NodeOp::Materialize` has labels for pipeline tracing
+- All `ColorFilterSrgb` variants use sRGB-space color matrices (not Oklab)
+- `ColorMatrixSrgb` operates in sRGB gamma space
+- LibjpegCompat chroma upsampling configured for JPEG decode
+- No-op Resample2D nodes stripped from expanded CommandStrings
+
+### Patches
+- zenpixels, zenpixels-convert (local)
+- zencodec (local — has SourceColorExt::is_srgb, icc_profile_is_srgb)
+- zenjpeg (local — has ICC extraction fallback fix)
+- moxcms (local — has PR #152 #153 fixes)
+
+### Tests
+- `/home/lilith/work/moxcms/tests/srgb_roundtrip.rs` — proves moxcms sRGB identity
+- `zen_watermark_red_on_green` — proves watermark compositing works
+- `zen_watermark_red_alpha_on_blue` — proves alpha compositing matches v2
+- `zen_watermark_fullframe_resized` — proves resize+compositing matches v2