fix: detect resampled official logos

[Minorli]

Summary

• recognize official logos that were shifted by up to three pixels during upstream down/upscaling
• retain the existing strict per-channel color tolerance and 90% match threshold
• preserve standard logo placement and avoid adding a second logo to edited images
• release PicGen 0.1.64

Production finding

A deployed customer-style edit preserved the source logo after upstream resampling, but exact-position matching scored about 0.68 and the client added a second logo. The bounded neighborhood matcher scores the unobscured real logo pixels at 0.9279 while an identical color pattern outside the radius remains rejected.

Test plan

• ./scripts/check.sh (410 passed; Ruff and Mypy clean)
• uv run --with pytest-cov pytest -q --cov=picgen --cov-report=term (87% total)
• uv run --with pip-audit pip-audit (no known vulnerabilities)
• real failed-image Chrome pixel replay (0.9279 match score)
• logo policy suite (15 passed, including local-shift acceptance and out-of-radius rejection)
• Docker build and non-root healthy-container smoke test