Correctness fixes for shipped pooling/caching optimizations (if.84)

[oysteinkrog]

What this does

Fixes six correctness defects introduced by the pooling and caching optimizations that shipped in if.84 (base ref bbc8bfe), and repairs the smoke/perf CI gate that let those defects ship without a real test ever running. Each source fix ships with a regression test written to fail on if.84 and pass after the fix.

Why

An adversarial correctness review of the fork's shipping optimizations found one process-killing bug and five behavioral hazards, all in the recently added object-pool/instance-cache paths. The common thread: every one was analyzed only on its synchronous happy path, and the release gate that should have caught them was hollow (21 of 23 smoke tests were Assert.That(true) stubs, and the build fabricated an empty test report the aggregate gate then passed).

What's in here

Source fixes (WindowsBase, PresentationCore, PresentationFramework):

• DispatcherOperationEvent pool guard (blocker). The pooled cross-thread wait wrapper could run a captured completion callback after it had parked or recycled: a null dereference on a threadpool timer thread (process kill), or a stale signal that woke a waiter early. A wrapper-owned state lock plus a sender-identity check makes the callback safe in any state; the pooling is preserved.
• Revert DispatcherSynchronizationContext instance caching. Sharing one DSC instance per dispatcher/priority made an await continuation completed from a sibling operation run inline inside the completer instead of being posted, reordering async continuations. Restores fresh-per-operation semantics at all five sites (byte-identical to the stock base).
• Revert DispatcherOperation task-mapping skip. The mapping decision was baked into the task at construction, but dispatcher hooks attach later, so an unmapped task threw on DispatcherOperationWait. Marginal cold-path win, reverted.
• InputHitTest pooled-params re-entrancy. A re-entrant hit test shared one pooled parameters object and clobbered the outer walk's hit point. The slot is now taken and nulled for the walk, restored in a finally.
• HwndStyleManager stale-Dirty on pooled reuse. The activation reset ran on only one branch, so a reused manager could emit a spurious native style write. The reset is now unconditional.
• StreamGeometry pooled-context double-dispose. A pooled context handed to a public Open() caller could be stale-disposed after a later Open() revived the same instance, corrupting the second geometry. Public Open() now returns fresh contexts; pooling is confined to the strictly-scoped internal parse path via OpenPooled().

Plus a comment correction (removed a false "~17x" claim on AdornerLayer; the early return it credits exists in the stock base) and the CI repair: real dotnet test against the packed nupkg with a non-empty-report floor and an anti-stub lint, and perf-on-pack wired to BenchmarkDotNet JSON gating hard on allocations.

Six new regression fixtures live under test/InitialForce.WpfSmoke/Smoke/.

How to review

Start with the DispatcherOperationEvent guard (WindowsBase/.../DispatcherOperation.cs, nested DispatcherOperationEvent) and the DSC revert; those carry the load-bearing reasoning. The other four source fixes are small and local. Then check the CI changes in .github/workflows/build.yml and release.yml do what the fixes assume.

Verification status

Not built or runtime-verified locally (no arcade toolchain in the working checkout). The fixes and tests were authored by source inspection; the regression tests are designed to be red on if.84 and green after, but that is a design claim until this PR's CI builds the assemblies and runs them. Draft until CI is green and a human has reviewed. Do not merge on CI alone.