[#82] Input long-edge cap (max_input_edge_px, default 1280) before tracing

[realproject7]

Fixes #82.

What

Add ConvertOptions.max_input_edge_px: int = 1280: downscale the input so its longest edge ≤ N px before any classify/preprocess/trace (LANCZOS, aspect-preserved, only when larger, never upscales), so CLI/API/web share one behavior. 0 disables (full resolution). The Report records original->effective size for audit. CLI: --max-input-edge-px N.

Bump 0.7.0 → 0.8.0 (behavior change).

Why

On svgsmith's target (flat/illustration/logo art) excess resolution + JPEG/AI-upscale noise fragment the trace. Downscaling to ~1280 gives a cleaner, more economical trace, ~30% faster, with no visible detail loss vs 1536/full-res — validated on 5 Midjourney samples incl. high-zoom on hair/face + feather pattern (SSIM delta +0.002–0.004 = staircase noise; eye prefers 1280). Also bounds host cost (the x402 API hit a 120s timeout on large inputs).

High-detail→SVG is a weak niche, so this is default-on with a 0 opt-out. Naming/opt-out picked with the agent-user perspective (single numeric knob, _px suffix).

Distinct from a host's reject-cap (e.g. the API's max_image_px, which rejects oversized uploads): this resizes, it does not reject.

Verification

ruff clean; full pytest green (incl. a new test_max_input_edge_px_downscales_large_input covering downscale + aspect + Report note + 0 disables + below-cap passthrough + negative-rejected).

Follow-up (separate)

svgsmith-x402 API will drop its own _downscale_if_large pre-resize and pass max_input_edge_px once 0.8.0 is on PyPI (one place owns the cap).

🤖 Generated with Claude Code

[realproject7]

Sub-PO review OK. No findings. I checked placement before classify/preprocess/trace, report warning behavior, CLI option wiring, version source, and the transparent_background full-res mask note. Local validation: ruff clean; pytest 127 passed / 1 skipped; explicit 2048x1024 -> 1280x640 transparent-background smoke confirmed corners transparent and subject preserved. GitHub CI is green on 3.11 and 3.12, PR is mergeable.\n\nNote: GitHub would not allow a formal approval because this session is authenticated as the PR author account.