fix: close the [288,512) auto-enhance dead zone

[ozansozuozgit]

Problem

Real-world small web rasters (300–511px icons/logos) fell into an auto-enhance dead zone: below the 512px upscale floor, above the legacy <64px superres path, and usually not blurry enough to trip restore. A 500×500 JPEG app icon reproduced it: the tracer ate raw JPEG noise + AA, producing splotchy gradient regions and wobbly rounded-square edges (visibly worse than Vectorizer.AI on the same input).

Fix

Lower DEFAULT_MIN_UPSCALE_EDGE 512 → 288 so enhance=auto runs the proven SR-before-trace recipe on the [288, 512) band.

Why size-only gating: content-signal detectors were measured across all fixtures first (soft-edge fraction, unique-color ratio, 8×8 blockiness) — none separates clean synthetic art from degraded rasters (e.g. clean thin-lines scores blockier than a q80 JPEG). 288 separates by construction: every synthetic golden/corpus fixture is ≤256px, and real ≥288px rasters essentially always carry AA/JPEG degradation that SR cleans.

Known-inert side effect: the restore blur-threshold split rides min_upscale_edge, so [288,512) now uses the loose 1.0 threshold — behaviorally inert there since needs_upscale already fires and the restore downscale-ladder no-ops below its own 512 floor (documented in the docstring, asserted in the new test).

Deliberately out of scope: [64, 288) stays unenhanced — SR from <288px to 1600 is a >5.5× factor with real hallucination risk; needs its own R&D + fixtures.

Verification

• Full suite: 438 passed, goldens byte-identical (all ≤256px, below the floor by construction).
• New regression tests: band boundaries (287/288/500) in test_enhance_assess.py; end-to-end auto-fires-with-no-ladder-shrink in test_enhance.py.
• E2E on a 500px icon-proxy JPEG (auto → illustration): enhance fires (working res 1600), JPEG speckle patches gone, donut arcs smooth and continuous, rounded corners clean vs ragged/staircased before. Cost: 3.3s → 13.9s, 105KB → 449KB (gradient fits at higher working res — the known file-size axis).

🤖 Generated with Claude Code