The tree opened into nothing, and 1003 tests could not see it

[dmitrat]

Phase 0 of the fix plan: the probes that need a running Studio. Studio 1007 green, 1003 before —
and those numbers are the point, because every defect below was invisible to all of them.

What the run confirmed

Every phase-1 answer Studio could show was right: the recent list carried the LSM folder
database, the tree said 7 indexes where it used to say 8, the footer said page 1, the drop
confirmation quoted with straight quotes. The phase-3 menus were all there — Edit ▸ Find and
replace…, View ▸ Database…, a folder offering only Create and Refresh, a connection root offering
Close Database. A table opens into Id INTEGER, CustomerId INTEGER, … with the NOT NULL names in
bold.

Finding 20, answered by measurement

A parse error is marked: a squiggle one character wide, under the first letter of the token —
while a missing table is underlined across its whole name. That is why the report said nothing was
marked. The fix is the token's length, not the wiring, and it is not in this PR.

What the pass broke open

the tree opened into an empty row	two links missing, one in each direction: the row's IsExpanded is bound to the node's in a style setter, which does not push back — so the model never heard; and Children was a List, so replacing the placeholder notified nobody. Every test read Children directly and set IsExpanded itself — the ViewModel's side of a binding that only ever worked one way
Create ▸ hidden on a table	phase 3's halves contradicted each other: the trigger item is offered on a table, and the submenu holding it was not
the read-only banner cut off	the shape of findings 26 and 31: text that wraps or trims inside a horizontal StackPanel, which measures with infinite width. A rule now names it — and found four more (export, import and query error banners, query read-only banner). An element with its own width is exempt: a width is an edge
the strikethrough did nothing	a decoration set from a row style never reaches the cell text. What a deleted row gets — dimmed, warning colour down its left edge — is what the screen shows; the style claiming the rest is deleted rather than left lying

What automation could not answer

The MCP click prefers the Invoke pattern, which bypasses real pointer routing — so the palette's
click-outside and click-to-run (finding 18) and the double click that opens a table's data are
unverified. The double click was seen to be broken by the placeholder (the row toggles its own
expansion); the change for it is written but unmeasured. Both need a hand on the mouse.

🤖 Generated with Claude Code