Wavelet fixes (CDF 9/7, fair bit estimates), DCT edge fixes, and viewer UX

[thundermage117]

Summary

Fixes the DWT path, makes the DCT/DWT bitrate comparison fair, fixes DCT edge-case bugs, and adds several viewer features.

Codec (C++)

• CDF 9/7 wavelet (the JPEG 2000 lossy wavelet), implemented with lifting and symmetric extension, alongside Haar. transform_mode is now 0 = DCT, 1 = Haar, 2 = CDF 9/7, so existing callers that pass 1 still get Haar.
• Bit estimate. Bits are now estimated per subband as the zeroth-order entropy of the quantization indices. Previously the estimator measured dequantized values, charged 0.5 bit for every zero (so every curve bottomed out near 0.75 bpp), and added a 400-bit header to DWT only. At equal bitrate, CDF 9/7 now beats DCT by about 1.5 dB on the test images, in line with JPEG 2000 vs. JPEG.
• DWT quantization shrinks the step by 0.75 per coarser level instead of halving it. 0.75 was the best value measured on the bundled images. Chroma steps are 1.5× coarser, matching the ratio the DCT path already uses.
• DCT fixes:
  • Partial edge blocks are padded and coded instead of copied through unchanged, which inflated PSNR and left their bits uncounted.
  • Output is clamped to 0–255.
  • Quality is clamped to [1, 100]; quality ≤ 0 used to divide by zero.
• Chroma upsampling is bilinear instead of nearest-neighbour.

WASM bindings

• The reconstructed image is cached, so switching views no longer redoes colour conversion.
• New rd_encode for RD sweeps. It skips SSIM and leaves the displayed result untouched.
• New Subbands view (view mode 7).
• The Makefile builds EXPORTED_FUNCTIONS from a single WEB_EXPORTS list. _set_artifact_gain was missing from the old list.

Web

• Transform selector: DCT / Haar / CDF 9/7. transforms.ts is the single registry for labels, URL slugs and chart colours.
• RD chart draws one curve per transform.
• Subbands view shows the wavelet pyramid full-frame; coefficients quantized to zero are drawn dim.
• In wavelet mode, the block inspector entry is replaced by a "Wavelet Subbands" tool. Before, the inspector showed zeros for DWT.
• Settings are saved in the URL hash (#q=30&t=cdf97&cs=420&v=subbands&img=3) and restored on load.
• Download PNG button.
• Shortcuts: [ / ] change quality by 5, T cycles the transform.

Build / CI

• make sanitize builds in build-sanitize/. The old recipe tested -d on the cache file, so sanitizer flags stayed cached in build/ and the next make test hung.
• dev, test and coverage now set their CMake flags explicitly, so one target's settings can't carry over to another.
• Apple clang 17's ASan runtime deadlocks at startup on macOS 26. make sanitize detects this with a 10 s probe and falls back to UBSan only.
• New sanitize-native CI job runs ASan + UBSan on Ubuntu and gates build-web.

Behaviour changes

• File-size estimates and RD curves change everywhere, because the bit model changed.
• Loading an image keeps the current quality instead of resetting it to 50. Shared links need this.

Test plan

• make test: 52/52 GTest. New tests cover CDF 9/7 reconstruction and gain, subband indexing, coded edge blocks, pixel range, quality clamping, bits monotonic in quality, flat image ≈ 0 bits, 9/7 > Haar, subband map, and smooth chroma upsampling.
• make sanitize locally: 52/52 under UBSan (ASan is unavailable on this toolchain, see above).
• npm run check: 0 errors, 0 warnings.
• npm run test: 193 unit tests. New: url-state, transforms, rdEncode.
• npm run test:browser: 42 tests, including a new real-WASM suite covering all three transforms, rd_encode isolation and the Subbands output.
• Manual check in Chromium via Playwright: shared-link restore, Subbands view, shortcuts, RD chart, PNG download; no console errors.
• First run of the new sanitize-native CI job (ASan on Linux has not been run locally).

Note

feature/motion-estimation has this branch merged into it locally, with conflicts resolved, so it should merge cleanly after this PR lands.

🤖 Generated with Claude Code