[actuallyitsnathaniel]

What

• On macOS, RT64's present thread was reading a stale Cocoa window size from plume (getWindowAttributes returned a cache and only scheduled async refresh). Metal swapchain / ImGui targets could stick after drag-resize or fullscreen. New registered plume patch always reads live attributes.
• HUD Fit-to-window gutters only republished viewport extent while the ImGui overlay (or other UI) was drawing. With the menu closed, Expand world + Fit HUD could disagree after resize. ui::draw now calls set_viewport_extent every present, before the overlay-closed early return.

Pattern checklist

• First-party changes are in-place edits under runtime-recomp/, modding/ or scripts/ (no new patches/<name>/)
• Third-party changes are *.patch files registered in patches/manifest.json (patches/plume/0004-live-cocoa-window-attributes.patch)
• Guest hooks only via guest-hooks/hooks.toml; symbols only via symbols/*.toml (n/a)
• New settings, launcher pages and owned services are registered, and their contract docs are updated (n/a)
• netplay/legacy_identity.hpp updated (reviewed by both owners) if simulation, wire or owned-service behaviour changed (n/a)
• Tests added or updated; python3 scripts/check.py --suite guard passes locally (present-thread Metal overlay resize check; HUD extent publish-before-early-return pipeline guard)
• Platforms considered: rebuilt and ran on macOS (build/dkr-runtime-macos). Host preset (RT64 off) 141/141. Windows / Linux / iOS / Android not rebuilt for this PR.

Risk

Low. Plume change is Apple window-size reads only; HUD change only republishes Fit viewport aspect when size changes. Capsule base_sha256 for runtime_ui.cpp updated. Revert the commit to roll back.