fix(table): dispatch drag gestures that leave the widget, and measure travel against the content - #377

…e widget

A drag that starts in a rendered cell and is released outside the widget
never delivered a click to the widget root — per UI-events the click
retargets to the nearest common ancestor of the press and release targets,
measured here as .cm-content — so the gesture was lost and the editor was
left with whatever caret its observer happened to park at the release point.

Add a second dispatch seam: a document-level mouseup, armed at mousedown,
that dispatches only when the release landed outside the root. An inside
release still belongs to the click listener, so the modifier-click
open-external path is untouched.

Four things disarm the seam, each pinned by a test that reddens when its
guard is removed: the release itself, a native drag-and-drop (which ends in
dragend, never mouseup), WidgetType.destroy, and the next press anywhere.
That last one is registered in the CAPTURE phase deliberately — four sibling
widgets stop mousedown propagation while leaving mouseup alone, and a
bubble-phase disarm starved by one of those would let a stale gesture
dispatch a range the user never drew.

The happy-dom suites drive the widget with hand-built MouseEvents, so none
of them observes what a real pointer gesture delivers. These two rows use
Playwright's own mousedown/mousemove/mouseup/click over real geometry.

The emptiness assertion comes first deliberately: before the fix the editor
already parked a collapsed caret at the release position, so asserting the
head alone would have passed without the seam existing.

The second row covers the overshoot below the last line, which the unit
suite cannot see — posAtCoords clamps a point past the document to
doc.length rather than answering null, so that gesture is a real range.

Type soundness: isDrag's false arm narrowed on conditions that are not
properties of its argument (detail and travel), so TS's negative inference
lied. Split into a pure isArmed guard plus armedDragFor, which returns a
value and therefore has no false arm at all.

Coordinate and offset spaces: contentPoint returned a bare {x, y},
structurally identical to the viewport points passed to posAtCoords and
cellPointAt two lines away. Named ContentPoint makes the mis-pass a compile
error. releaseRange was the one absolute-space entry in this family with no
asAbsoluteOffset mint, which the module's own policy requires be said out
loud; DragSelection now names the shape both seams produce.

Diagnostics: the release-lookup catch logged {err} alone, below the standard
this file sets twice. It now carries the cell offset and the coordinates.

Phase: the dispatching mouseup was bubble while both disarms were capture,
resting on an enumeration of today's siblings — the reasoning the adjacent
comment rejects. It is capture now. CodeMirror registers no document mouseup
for a gesture that starts in an ignoreEvent widget, so the ordering change is
inert; the browser suite measures unchanged at 64.

Coverage: four paths had no test — the forward arm of the outward snap
(a mutation to a constant left every row green), the release-lookup catch,
the multi-widget disarm ordering, and the fixture's posAtCoords misuse
channel. A byte-identical duplicate row that could not pin what it named is
replaced by one pinning re-arm after a completed gesture.

Comments: the file header documented only the click seam while the code had
two; 'sole dispatch seam' sat 35 lines above 'The SECOND dispatch seam'; and
several internal line references had rotted. Bare :NNN refs are replaced by
file-and-symbol references, which do not rot.

The two disarm listeners each spelled out the same abort-and-clear pair,
which must stay identical — an abort without the pendingDrag clear leaves an
anchor to be paired with someone else's release. One named local makes that
structural rather than a convention, and records why the mouseup seam
deliberately does not use it: an inside release must leave pendingDrag armed
for the click listener that owns it.

travelSince measures a distance, so it now takes the pressed point rather
than the whole PendingDrag; the anchor's cell mapping is no business of it.

…t they missed

The second review round found that the fixes from the first had reproduced
the defect they were fixing: prose asserting things the code does not do.

Two test comments claimed pins that measurement disproves. The two-table row
said it pinned the WeakMaps' root-keying; making both maps module-wide
singletons leaves the whole suite green, because the capture-phase disarm
runs before the second root's bubble arm, so even a shared map sees
clear-then-write-then-read. It pins the coexistence of disarm and arm, and
says so now — with two new rows that do pin the keying, verified red under
that mutation.

The re-arm row claimed no single-line mutation could redden it. Two can. One
works here; the other should work and does not, because happy-dom registers
listeners on a pre-aborted AbortSignal where the DOM spec requires a no-op —
which is why the earlier claim was believed. Both are recorded, with the
general rule that a green mutation is not evidence of a missing pin in this
environment until the event fidelity is ruled out.

The release-lookup payload widened in cycle 1 was asserted so loosely that
all three new fields could be deleted with the suite green; it now pins the
values, which also catches logging the wrong cell.

Also: a mint comment claimed to be the family's only unbranded entry while
the caret path beside it is unbranded too (unchanged by this PR, and now
tracked separately); asAbsoluteOffset's docblock enumerated its call sites
and the enumeration had already gone stale; isArmed read as 'a gesture is
armed' when it means 'the press landed on a cell'; and the ordinal, unit and
scope of three further comments were wrong.