Windows: stop counting already-open recent projects as rendered sidebar rows

[coneilen]

Fixes #428.

Three distinct changes. Only the first is a product change; the other two are test corrections with justification, called out separately so they are reviewable.

1. Product fix: layout counted rows that are never rendered

appendRows() skips recent projects that are already open:

Sidebar.zig:507   if (project.isRemote() or isProjectOpen(model, project.path)) continue;
Sidebar.zig:518   if (!project.isRemote() or isProjectOpen(model, project.path)) continue;

but layoutFor() counted every entry:

Sidebar.zig:588   .project_count = model.recent_projects.items.len,

and that count drives geometry at :485 ((project_count + project_heading_count) * 24). So the project section reserved space for rows that were never drawn whenever an open project also appeared in recents, inflating section height, every offset derived from it, and scroll extent - and causing hit testing to resolve a click to the wrong row kind.

The sibling helpers projectHeadingCount() at :647/:652 already applied !isProjectOpen(...), so heading visibility respected the skip and only the row count diverged. That asymmetry is what identified this as an oversight rather than an intentional model.

Fixed so the count is derived from the same predicate appendRows() uses, rather than duplicating the condition in a second place - duplication is what let producer and consumer drift apart originally.

2. Test correction: stale oracle in the scroll test (not a product bug)

sidebar scroll clamps overflow, shrink, and resize expected 334 and found 410. That 76px delta is not project overcount. It is exactly the Activity block:

Sidebar.zig:817   if (model.activity.items.len != 0) {
Sidebar.zig:818       bottom += 18 + 24 + 34;      // = 76

paint() at :282 renders that block - header, "Attention only" filter, prev/next controls, cards - whenever activity is non-empty, so contentBottom is correct to reserve the space. The test appends 3 activity events but its oracle omitted the Activity term.

410 - 334 = 76. Changing the product to satisfy the old expectation would have excluded visibly-rendered content from the scroll extent and clipped the Activity block - introducing a bug rather than removing one. contentBottom is unchanged; the expectation now includes the missing term, written as literals rather than as a call to contentBottom() so the oracle stays independent of the code it checks.

3. Test correction: clamp loop no longer coupled to the expected value

The overflow loop was for (0..10), i.e. exactly 10 * 40 = 400. Against the corrected maximum of 410 it never overflowed, so expectEqual(short_max, scroll) was no longer exercising clamping at all - the bound had been tuned to the old 334. Widened to overshoot decisively so it cannot silently stop testing clamping the next time this geometry moves.

Context

Sidebar.zig's 22 tests had never been executed by any harness (issue #424): WindowsShell.Tests.ps1 runs Zig tests from a hand-maintained list, this file was not on it, and Zig only runs test blocks from the root file passed to zig test. These failures surfaced on their first ever execution.

RED: three tests failed on unmodified main under pinned Zig 0.15.2 -> expected 0 found 1; expected 334 found 410; expected .open_project found .quick_chat_overview
GREEN: after the layout fix and the two justified test corrections -> all 22 Sidebar tests pass, 93/93 in the Sidebar root run
REGRESSION: the row-count and hit-test assertions re-fail if layout again counts unrendered project rows, and the scroll oracle re-fails if the Activity block reserved height changes -> drift between what is counted and what is drawn is caught

Verification

Pinned Zig 0.15.2. Sidebar root: 93/93 pass. The App root additionally verified at 231/231 with PR #426 pending App test-fixture fields applied temporarily, then reverted - App.zig does not compile its test build on current main without them (missing sidebar_state, declared_entry_ids, kept_worktree_paths at App.zig:6005), and that fix belongs to #426.

Tools/windows/Tests/WindowsShell.Tests.ps1 is deliberately untouched here. #426 quarantines these three tests by name; lifting that quarantine should follow after #426 lands.