Deliver mouse events in order so one handler's capture reaches the next

evidence/pointer-capture-order-20261006/README.md

@@ -0,0 +1,57 @@
+# Ordered native pointer admission
+
+App originally selected every pointer target before the queued MouseDown handler
+could establish capture. A queued move and release could therefore miss the press
+owner, then its late handler left capture active. This is a source defect; it is
+not proven to be the interleaving in closed physical02.
+
+Message owns one pending dispatch completion. MessagePump._post_message_and_wait posts the original
+message to its existing FIFO and releases completion only from the receiver which
+still owns it. Bubbling transfers that custody through the original DOM queues;
+the originating pump retains its own bubbled copy for later dispatch. A retiring
+receiver releases its original delivery without cancelling children or replay.
+App awaits this admission before resolving the next pointer packet. Screen and
+Widget return the existing AwaitComplete resource; Pilot awaits the same contract.
+Capture, release, selection, hit geometry, broker actions and disabled input keep
+their existing owners. No second input queue, capture state or timed delay exists.
+
+Native Input, TextArea and ScrollBar capture from their queued handlers. Toad
+SidebarResizeHandle, SidebarSlider, GoalBar and Mandelbrot do the same. Ancestor
+handlers retain ordinary bubbling. The awaited chain stops before the originating
+App pump, which cannot await itself; its forwarded application-level event remains
+on its original FIFO. This is target/DOM admission, not synchronous application
+notification delivery.
+
+Existing refactor-audit Package parsed 249 native production / 461 test modules
+and 288 Toad production / 400 test modules, with zero omissions. before.json
+contains 118 lexical owner/consumer sites. Dynamic overrides/factories remain a
+semantic resolution limitation, not zero-by-omission proof.
+
+Changed native checks passed: 33 in 10.45s through actual App/HeadlessDriver
+packet ingress, direct and ancestor capture, delayed press, release-before-wheel,
+Input/TextArea selection, filtered delivery, self-removal, original input messages
+and queued selection/copy behavior. The first selection batch caught 80 redundant
+selection rebuilds: original call_next flushed every newly awaited DOM delivery.
+That early fallback is deleted; original copy/release/paint/refresh-callback owners
+still flush the single pending projection. Final source retains all 80 mouse events
+and exact mid-burst/final copy values. No new timer or selection state exists.
+
+Two initial authored editor assertions incorrectly assumed an unpanned/unwrapped
+substring; their expectation is now movement selection, preserved original text
+and ended capture. Initial negatives stay in the tool transcript; final-checks.log
+retains the coalescing refusal and qualified-checks.log the completed result. Parent owns recorder motion
+acceptance; no physical run, installed pin, package or public runtime is changed.
+Frozen e15 / physical02 evidence remains unchanged.
+
+The first checkpoint broke the drivers' coroutine scheduling contract by changing
+_post_message to a synchronous return. That negative is retained in Parent's
+mounted.stderr. The original async _post_message coroutine and bool admission
+contract are restored for all Headless/Linux/inline/Web/Win32 driver callers,
+MessageTarget/EventTarget protocols, Markdown self-publication and input tests.
+Awaited target completion is a distinct MessagePump operation on the same FIFO;
+it does not change raw driver admission or add a queue. These callers are present
+in before.json; the first checkpoint failed to read their scheduling requirements.
+
+No physical02 causality or repaired physical motion is claimed. Parent owns the
+matched application/recorder acceptance. All local checks are terminal, with no
+provider, installed package, pin or public runtime operation.