Match Windows worktree notice accounting to macOS semantics

[coneilen]

Summary

Windows computed the worktree notice from different semantics than macOS. The
previous parity-ledger wording was also inaccurate: macOS excludes the canonical
checkout by path, not by branch. It counts every remaining assessment,
including prunable rows, and excludes only prunable rows from the size loop.
This aligns the Windows computation and states those exact semantics in the
ledger.

Changes

• Identify and exclude the opened canonical checkout by path rather than by
  branch, so the first row is no longer misclassified.
• Count all remaining assessed rows, including prunable ones, toward the
  worktree count; sum only non-prunable rows toward size.
• Round each complete worktree to whole KiB before summing, approximating
  macOS du -sk allocated-usage reporting.
• Clarify the ledger row to state count/size semantics, thresholds, and the
  residual limits. The row remains Partial.

macOS reference evidence

macOS runtime evidence is added by PR #500 at commit 10173c9f (that PR is
still open, so these paths are not present at this PR's merge-base):
GraphCode 0.1.76 (299), built from 717240cd, macOS 26.7. In that report R
means direct observation of the running app; the seven/eight boundary and both
screenshots are R evidence.

Those macOS captures establish the target behavior only. They are not
Windows validation.

Test plan

RED: & $env:GRAPHCODE_ZIG0152 test src\WorktreeStatus.zig -> behavioral failures: 4 of 27 tests failed, covering first-row/opened-checkout identity, prunable count/size, threshold boundaries, and logical-versus-allocated size; output preserved in worktree-notice-red-test.log.
GREEN: & $env:GRAPHCODE_ZIG0152 test src\WorktreeStatus.zig -> 29 tests passed.
REGRESSION: pwsh -NoProfile -File Tools\windows\Tests\WindowsShell.Tests.ps1 -ZigExecutable $zig -> 654 tests passed; scaffold contract passed. & $env:GRAPHCODE_ZIG0152 build -Doptimize=ReleaseSafe "-Dwinghostty-dir=$env:GRAPHCODE_WINGHOSTTY_ROOT" -> passed.

Limits of this test plan

The GREEN result is unit-level only. No live Windows toolbar/chip boundary
walkthrough was performed, so the ledger row stays Partial.

The initial shell-suite attempt stopped early because the pinned Winghostty test
library was missing. That is not a green result and is not counted as one;
the successful retry above is the run that stands, and it has a positive test
count. Filtered Zig tests are deliberately not used as suite evidence, because a
zero-match --test-filter still exits successfully.

Windows reads FILE_STANDARD_INFORMATION.AllocationSize and rounds each
complete worktree to KiB. This is a deliberate approximation of macOS
du -sk allocated-usage reporting; identical byte accounting across filesystems
has not been demonstrated.

Checklist

• I have read the Contributing Guidelines
• I have signed off my commits (git commit -s) per the DCO
• Tests pass locally (make test): macOS/Xcode is unavailable on this Windows host; the Windows shell suite and build pass.
• Code follows the existing style (make check): macOS/Xcode lint is unavailable on this Windows host.
• I added the test before the implementation and observed the intended RED failure