Replace the GUI Builder with a Maven-first Codename One app - #5523

The guibuilder job hit the one hour limit because of two tests I added. Both
focused a real text component, and focusing one starts native editing on the
simulator, which blocks the EDT that callSeriallyAndWait needs -- so they hung
rather than failed, which is why I never got a result out of them locally
either. They are replaced by the branch that can be driven here: with nothing
editable focused the key belongs to the designer, and with no editor open the
Save shortcut belongs to the form. The focused-surface half is not covered, and
the tests say so rather than pretending otherwise. LiveCssTest is 8 tests in
132 seconds again.

Class annotations are reapplied on every save, not only during migration --
same shape as the imports and the trailing declarations before them, and the
third thing that ordinary branch has had to be taught to carry.

The accessors Object already declares are reserved, so a bindable component
named "Class" no longer generates a getClass() that cannot override the final
one and will not compile.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…guments

"extends com.acme.Form" passed the check because only the simple name was
compared, and the rebuilt declaration then resolved Form out of
com.codename1.ui.* -- the developer's base class, and everything the carried
members inherit from it, gone without a word. A qualified name is compared
whole, and com.codename1.ui.Form is accepted as the real one.

initGuiBuilderComponents(registerAndGetResources()) was accepted as generated
work, so removing the constructor dropped a call the developer made for its
side effects. Only the argument forms the scaffold wrote are accepted: nothing,
Resources.getGlobalResources(), or the constructor's own parameter.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The ordinary save branch rebuilt the declaration from the template, so a class
migrated with "implements Observer" quietly stopped implementing it on the next
save, taking any @OverRide among the carried members with it.

That is the fourth thing this branch has needed teaching to reapply, after the
imports, the trailing declarations and the annotations. I named the pattern
last time and still did not act on it, so this time I went through what
migration does rather than waiting: everything else it adds lands inside the
user region, which is carried verbatim, so this is the last of them.

constructorAt() matched on the name and a parenthesis alone, so a legal
"void LoginForm(Resources r)" was taken for a constructor and deleted -- its
callers broken by a migration they had nothing to do with. A return type in
front of the name now rules it out, and the match is a whole word.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

A cleared Event field on a toolbar command emitted "this::null", which does not
compile. An unhandled command is a reasonable thing to want, so it is added
with an empty listener rather than refused.

An Advanced inset the layout cannot parse was swallowed: the canvas stayed on
the previous constraint while the document kept the text, and the same text
reached LayeredLayout.setInsets() in the generated form, which threw while the
form was being constructed. The value is taken back out of the document and the
status bar says why.

A Form's content pane arrives scrollable on Y and the generator emits no setter
when the attribute is absent, so a canvas that forced it off disagreed with the
running form. The gutter stays hidden when the scrolling is that default rather
than something the document asked for -- showing it cost a few pixels of width
the running form does not lose, which moved every centred component, and a
centring test caught it before this went out.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The inspector checkbox still carried the old false, so a Form whose content
pane scrolls showed scrolling as off, and saving without touching the box
looked like it would disable it and did not. Both the preview and the inspector
now ask GuiDocument.defaultScrollableY(), so there is one answer rather than
two that agreed until one of them changed -- which is what happened here, and
the test asserts the two are equal rather than each being right on its own.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

"extends Form" says nothing on its own, and carriedImports() keeps whatever
resolution the legacy file had, so an "import com.acme.Form" survived while the
regenerated constructor called super(title, layout) on a base class that had
only ever supported the implicit super().

The name is resolved the way javac would from this file: an explicit import
decides it, failing that a com.codename1.ui.* import does, and a second
wildcard makes it genuinely ambiguous. No import at all means a class in the
same package, which is also not ours.

I looked at this case when the qualified one was reported and talked myself out
of it -- refusing every unqualified extends would have turned away ordinary
companions, which was true, but resolving the name was the option I did not
consider. One test covers all five outcomes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

declaresActionHandler() scanned the whole file, so a trailing helper class
declaring onSubmit(ActionEvent) satisfied the form's own check: the stub was
left out while buildUI() still emitted this::onSubmit. It is scoped to the
form's class body at member level, which also rules out an inner class or a
local declaration inside a method.

hasReturnType() read the annotation in "@Inject LoginForm(Resources r)" as a
return type, since that declaration has no access modifier. Neither constructor
path then recognised it, and the old constructor was carried into the generated
class still calling the initializer that is gone. Annotations are blanked
before the token in front of the name is judged -- the same helper the other
declaration scanners use.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The Code pane recorded its buffer as the on-disk baseline, but editing outside
the markers is allowed and companionSourceFor() regenerates those regions on
the way out -- so the pane reported a clean save while showing text that was
never written, and closed without a word about it. The baseline is the text
that reached the file, and the pane is reloaded onto it, which is also the only
honest thing to show.

Model regeneration was armed only by a strategy change, so adding, renaming or
deleting a bindable control after the model existed left it naming the old set.
UiBinding binds by exact name, so those controls stopped binding with nothing
failing to say so. Save now notices when the model does not mention a bindable
component and rewrites it. Only the set of names is compared: the model is
generated from them, so a difference there is this generator's business, while
anything else in that file is the developer's.

isBindable() is extracted rather than repeated -- the type list existed once
inline and this needed the same answer.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…rameter

The divergence check searched the model for any occurrence of the name, so
"value" matched the parameter of every generated setter and a model missing
that control looked current. Both strategies write the bound name as a string
literal -- new Property<>("email", ...) and @Bind(name = "email", ...) -- so the
literals are what is compared.

initGuiBuilderComponents(res) was accepted from a hard-coded list of likely
parameter names, so a no-argument constructor passing a field called res read
as scaffolding and the developer's resource selection went with it. An
identifier is accepted only when it is that constructor's own parameter.

Which made an earlier test wrong: it asserted "res" migrates from a no-argument
constructor, and that is exactly the case being reported. Its expectation is
corrected rather than the rule loosened to keep it green.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Regenerating a diverged model on save replaced a file saveModelSource() keeps
exactly as typed -- fields, validation annotations and methods this generator
knows nothing about, thrown away to add one property, without a word. It asks
first, remembers a refusal for that form so every later save does not ask
again, and MCP refuses rather than putting a modal in front of a client that
cannot answer it.

The divergence check counted every string literal, so a validation message
equal to a component's name made an unbound control look bound -- the same
false positive as the setter parameter, one step further out. Only the two
forms this generator writes are read: the Property constructor's name and the
@Bind name attribute.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

javaEscape() deleted \r instead of escaping it. The document keeps one and the
preview shows it, so saving quietly joined the surrounding text in the running
application and nowhere else. Tab had the same exposure without being reported
and is escaped with it.

build-and-run.sh called /usr/libexec/java_home unconditionally, which exists
only on macOS, and under set -e the failed substitution ended the script before
any build began -- on Linux the helper appeared to do nothing at all. It uses
JAVA8_HOME/JAVA21_HOME when set, then java_home where it exists, then the usual
/usr/lib/jvm layout, and otherwise says which variable to set. All three paths
exercised by hand, including the one where nothing is found.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…guard

An MCP redo that reapplied a binding-strategy change reached syncBindingModel()
and opened a modal. MCP runs synchronously on the EDT, so the client waits for
a human who is not there and the call never returns.

Guarding sites one at a time is what let this through -- save and refresh were
each fixed when reported, and redo was not among them. Every MCP entry point
now runs inside unattended(), and the confirmations ask cannotAskTheUser()
before opening anything, taking the branch that keeps the user's work: the
strategy prompt keeps the existing model, and the divergence prompt declines.
Neither loses anything, and the next interactive save asks properly.

unattended() restores the flag in a finally and reports a runtime failure as
text, since an exception escaping it would leave every later confirmation
suppressed. The test covers both.

Callable is outside the Codename One bytecode subset -- the compliance gate
caught it -- so the command is a small interface of our own.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The delegate dropped the throws clause the original declared, so a caller with
a dedicated catch stopped compiling -- the same broken-callers problem the
delegate exists to prevent. It carries the clause, which is legal on a body
that only chains.

A type in the same package beats an on-demand import, so "package com.example;
import com.codename1.ui.*;" makes Form mean com.example.Form when that file
exists. The project has that file to look at, so it is looked at rather than
the case being refused: a wildcard with no such sibling still migrates.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The code pane checked only a pending strategy rewrite or a missing model, so
adding a bindable control and saving from Code wrote the .gui and the companion
against a model that does not name it -- the control stayed unbound until some
later toolbar save happened to notice.

The two save paths needing the same thing done twice is the shape behind a
string of these reports; both now run the divergence check and the same
confirmation.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…mbers

-Xdock and the eawt exports go on only for Darwin. A Linux JVM refuses to start
on the first, so every default build-and-run.sh reached its failure check
twelve seconds later with nothing but a log -- the same split
OpenGuiBuilderMojo.desktopIdentityArgs() already makes.

"public <T> LoginForm(Resources r)" is a legal constructor whose type parameter
was read as a return type, so neither constructor path recognised it and the
old constructor was carried into the generated class still calling the
initializer that is gone. A bare <T> is skipped while List<String> still marks
a method, which the test pins in both directions.

Columns, Maximum length, Rows and the slider values went through the
unrestricted field, so "garbage" was stored: the preview kept its previous
value and appendIntSetter() emitted nothing, leaving the canvas at 20 while the
saved form used the runtime default and Save reported success. They use the
validated numeric field the table sizes already used.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>