[coneilen]

Summary

Fixes paired-review finding F2 by applying the macOS terminal chrome measurements as logical dimensions throughout the Windows physical-pixel workspace. The DPI192 live gate now distinguishes the bounded visible tab roster from the mounted tab topology instead of treating a hidden overflow tab as absent.

Changes

• Scale the 37-logical strip, 25-logical tab pills, 22-logical pane headers, 24-by-22-logical split/New Tab controls, close targets, fonts, glyphs, radii, and hairlines exactly once at the workspace DPI boundary.
• Preserve terminal surface allocation below 59 logical pixels of tab/pane chrome, DPI-aware UIA and pointer bounds, close guards, stable selection/focus roles, and overflow windows.
• Keep 11 visible tabs at a 1440-logical client and admit all twelve minimum-width tabs at the Mac-observed 1508-logical width.
• Record physical and DPI-normalized logical tab dimensions in matched capture observations.
• Prove a DPI192 New Tab mounts both Goal and Shell even when the expanded detail panel permits only the selected Shell in the visible roster; collapse the panel before verifying both stable tab identities and both selection round trips.

Test plan

RED: & $env:GRAPHCODE_ZIG0152 test src\TerminalSurface.zig -target x86_64-windows-msvc -lc -luser32 -lkernel32 "-I$env:GRAPHCODE_WINGHOSTTY_ROOT\include" "$env:GRAPHCODE_WINGHOSTTY_ROOT\zig-out\lib\winghostty-win32-host.lib" zig-out\lib\ghostty-vt-static.lib -lgdi32 -lopengl32 -limm32 -lole32 -loleaut32 -luiautomationcore -lshell32 -ladvapi32 -lwinhttp --test-filter "workspace terminal chrome keeps mac logical density" -> failed: expected 50 physical pixels at DPI192, found 22
GREEN: same focused pinned-Zig command after the change -> 1/1 passed at 96, 144, and 192 DPI with 1440/1508 width contracts
REGRESSION: pwsh -NoProfile -File Tools\windows\validate.ps1 -Task windows-shell -ShellPart unit -ShellValidationRoot C:\v36 -> 60/60 registered sections passed; native accessibility provider 224 checks, 0 failures

• TerminalSurface.zig full command -> 298/298 passed.
• App.zig full command -> 1583 passed, 17 intentional skips, 0 failures.
• Tools\windows\Tests\ValidationRunner.Tests.ps1 -> PASS; UiCaptureDriver 182/182.
• pwsh -NoProfile -File Tools\windows\validate.ps1 -Task hardening -> deterministic fixtures passed on 3/3 repeated runs.
• pwsh -NoProfile -File Tools\windows\validate.ps1 -Task windows-shell -ShellPart integration -ShellValidationRoot C:\v37 -SkipTrayLive -> F2 tab phase passed with overflowObserved=true, stable Goal/Shell IDs, and 50-physical/25-logical tab pills. Exact-head CI then reproduced a later worktree SelectionItem observer race. The follow-up binds the exact automation/runtime identity and rejects zero, duplicate, wrong-element, and stale-generation events; local live GREEN recorded UIA_SELECTION_EVENT_DECISION with count 1 and exact identities plus passing UIA_WORKTREE_SELECTION_EVIDENCE.
• The larger chrome also exposed an early zmx endpoint-publication race: resize returned ConnectionRefused/FILE_NOT_FOUND for a newly attached 50x10 session. The final change retries that exact owned surface/session/rows/cols up to five times with a 100 ms delay, cancels owned controls on teardown, logs session/geometry/exit/stderr, and still surfaces exhaustion. Pinned zmx resize passed independently; the live gate no longer reported terminal PTY resize command failed and advanced through the Move-status phase to later unrelated UIA fixture work.
• Authorized F2_TERMINAL_CHROME_FOCUSED_VALIDATION attempt on the exclusive PMv2 DPI192 desktop launched once from fresh C:\f6 and stopped before the first PNG because the focused wrapper did not include split/New Tab controls in its row query (Control roster differs for s-tabs-agent-unknown). The product, peer, and zmx processes cleaned to zero survivors. C:\f6\outputs\f2-focused-manifest.json SHA256 66EB0EB2614CA8608EBAB053E3214BA51FC6C724B3180E32727D5B8163D10B80; source ce92fe6970fa27f8a4da071e87fa21faeb84d904; binary SHA256 367E77D99F7CB8890D6D81EBAFC251D6C53B1313E2EC4B82D0DED3189FF105C5; images 0/11; privacy review had no images to inspect. No retry was authorized after launch, so the full focused scene matrix remains unproved.

Checklist

• I have read the Contributing Guidelines
• I have signed off my commits (git commit -s) per the DCO
• Tests pass locally (make test)
• Code follows the existing style (make check)
• I added the test/contract before the implementation and observed the intended RED failure