Merge branch 'rule-builder': a row builder for the tagging rules

Item 76 replaces the four free-text fields with a row builder: field and
operator dropdowns per condition, +/- to add and remove them, a match
all/any choice, and a separate "but not" block. The stored format does
not change, so mailctl needs no edit. The query string stays
authoritative and remains visible, and a rule the builder cannot
represent opens in a text mode every rule carries.

Along the way, items 75, 77 and 80, and a data-loss defect released in
0.16.0 (item 79): opening the dialog and pressing Save destroyed the
first rule with nothing edited. That one damaged a real rule in the
user's own file, which was repaired by hand.

Four defects in this work were found by hand rather than by the suite,
and each is recorded where it was missed: a lost note, a one-way text
mode toggle, a geometry save on a path neither button takes, and a rule
list squeezed to one row by a long rule. Two Qt traps and one about the
user's compositor went into CLAUDE.md.

1 files changed, 30 insertions, 0 deletions

diff --git a/CLAUDE.md b/CLAUDE.md
index 67c0794..6ab47b3 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -317,6 +317,36 @@ current index invalid when nothing was current. A test that calls `selectAll()`
 view therefore passes against a missing selection guard, because no signal ever fires. Test
 multi-select from a row that is already current, which is also how a user reaches it.
 
+**A `QDialog`'s buttons do not send a `QCloseEvent`.** `accept()` and `reject()`
+go through `done(int)`, which hides the dialog without ever closing a window, so
+a `closeEvent` override runs only for the window manager's X button. Anything a
+dialog must persist on the way out belongs in a `done(int)` override, which both
+buttons and `close()` reach. This shipped wrong in the rules dialog and the test
+covering it passed, because the test used `close()` and the user used Cancel:
+one route out of three. Assert every route. Underneath sits a second trap:
+`close()` on a widget that was never shown returns early WITHOUT reaching
+`done()`, so a test for the closed path has to `show()` the dialog first or it
+asserts nothing at all.
+
+**Under a tiling compositor a window's size is not the application's to
+restore, and the user's desktop is Hyprland.** `saveGeometry` stores
+`frameGeometry` and `normalGeometry`; `restoreGeometry` restores the NORMAL
+one. When the compositor tiles the window to fill its slot, the size the user
+drags is the tile's, and `normalGeometry` keeps whatever the code last passed to
+`resize()`. Measured against the real state file after a hand test: frame
+2248x806, normal 760x664, so the dialog correctly restored 760 and correctly
+looked broken. A whole session went into "the geometry restore is broken" before
+the blob was decoded. Decode the stored geometry before theorising, and expect
+`maximized` to read as a value no bool should hold, which is the tiled state Qt
+records and does not round-trip.
+
+The corollary for tests: **the offscreen platform cannot test window sizing at
+all.** It prints "This plugin does not support propagateSizeHints()" and returns
+an identical frame for a correct restore and a broken one, verified in a
+standalone program containing none of this project's code. A size assertion
+there passes against both, and a mutation putting the bug back leaves the suite
+green. Assert on the stored value, and leave the frame to a hand test.
+
 **A queued load can outlive the state that started it.** `loadThread` crosses to the worker
 on a queued connection, so its reply lands after whatever the UI did in the meantime. The
 generation counter covers a superseded *query*, not a superseded *selection*: blanking the