perf(gpu): share Canny classification and root flags

[rsasaki0109]

Depends on #184 (Trimmed ICP selection tree); retarget to main after that exact source is merged.

Canny previously allocated separate per-pixel classification and strong-neighbor root flags. Preserve classification in the low two bits of an atomic word and OR the root flag into bit 2, removing one 4*N buffer for every aperture. Full parent indices, complete hysteresis, exact edge bytes, five/six passes, metadata and explicit image transfers remain. Required storage-buffer count falls from five to four; classification destination footprints and receipts reflect the removed allocation.

At 4K, requested private scratch falls from 165,888,008 to 132,710,408 bytes (aperture 3) or from 199,065,600 to 165,888,000 (5/7): 33,177,600 bytes saved. Two isolated surveys show 1.183–1.574x completed resident improvements across the tested 4K fixtures. Full-call results are mixed and retain adverse cases. Timings include acquisition/encoding/completion under the same default 64 MiB retention policy; shader-only or general-device speedup is not claimed. All 20 paired rows and conditions are in notes/2026-10-13_gpu_canny_shared_state_flags.md.

Validation:

• 864 edge conditions match CPU and frozen PR Add resident GPU Canny edge extraction #180/Reduce resident GPU Canny gradient memory by half #181/Halve GPU Canny horizontal scratch without rounding #183 GPU implementations.
• 433,320 shared state words match old classification | (old root flag << 2), including deliberately poisoned scratch; twelve contended/long/disconnected masks match frozen GPU and independent BFS.
• Native Canny: 13 passed, zero failed, four measurements ignored. Image regression: 92 passed, zero failed, 17 measurements ignored. Historical gradient and horizontal surveys were explicitly re-run against actual frozen layouts.
• All-feature/all-target GPU Clippy, feature-free GPU, GPU-image WASM, minimal gpu-image facade and formatting checks pass. Browser execution was not measured.