Two failure modes that both showed the user the "no" cursor with no
other feedback:
1. ResolveDropTarget returned null when e.Source had no
SharpTreeViewItem ancestor, so Explorer drops landing in the gap
beneath the last row (or onto an empty list) never reached
SharpTreeNode.InternalDrop.
2. The middle 50% of a row produces a DropPlace.Inside target on the
row's own node. Most concrete SharpTreeNodes inherit the base CanDrop
(returns false), so a literal "drop a .dll onto an assembly row"
refused even though AssemblyListTreeNode happily accepts that payload.
For (1), fall back to (Root, Children.Count, DropPlace.Inside) when no
row is hit. For (2), introduce PickAcceptingTarget which retries with
the empty-space (root) target when the initial CanDrop is false. Both
OnDragOver and OnDrop go through the same picker so the cursor and the
actual drop agree on the outcome. The retry skips when the initial
target already IS the root, so a real refusal still surfaces as "no".
The marker adorner stays hidden for empty-space drops because the
place is always Inside (Item is null for that case); the existing
DropPlace.Inside early return covers it.
DropTarget, DropPlace, and the two new helpers are internal-visible to
the ILSpy.Tests project for the unit tests that assert the fallback
chain.
Assisted-by: Claude:claude-opus-4-7[1m]:Claude Code
Frozen tabs each cache their own decompiled output, but the language
change handler only refreshed ActiveDecompilerTab — hard-wired to the
preview tab — and the display-setting refresh only reached the preview
tab's cached content model, so frozen (and floated) tabs kept showing
stale output until manually re-selected. The handler predates the
freeze feature, which introduced multiple sibling decompiler tabs
without widening it.
Both paths now iterate every decompiler tab across all document docks
(the walk CancelPendingOperationsAsync already used, extracted as
AllDecompilerTabs), and the language handler uses Redecompile instead
of the CurrentNode re-assignment hack.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
Three follow-ups so the WPF-era ILSpy.xml survives a load-save-load
cycle through the Avalonia host without losing data the user expected
to keep.
Preserve unknown children verbatim. LoadFromXml only interprets a
known set of element names (KnownChildren) and stashes everything
else; SaveToXml re-emits the stash after the known section. The
AvalonDock <DockLayout> blob the WPF host writes is incompatible with
the Avalonia Dock host and we don't want to interpret it, but blank-
ing it on save would discard a WPF user's saved layout the first
time they launched the Avalonia build. The same stash future-proofs
older builds against fields added in newer ones.
Emit the legacy on-disk shape on save. WindowBounds is written as a
CSV body "L,T,W,H" (the WPF Rect TypeConverter format) instead of
Left/Top/Width/Height attributes, and ActiveTreeViewPath <Node>
values are \xNNNN-hex-escaped on write. With both directions of the
conversion in this file a file written by the Avalonia build is
still readable by an older ILSpy 10.x install, and the diff against
a pre-existing ILSpy.xml stays small during the transition.
Wire SelectedSearchMode and ActiveAutoLoadedAssembly. The WPF host
persisted both: the search-pane mode picker so the user's last
choice survives restarts, and the file path of the auto-loaded
(dependency) assembly the saved tree-path lands in so the restore
can re-open it before walking. The Avalonia code read both into
properties that nothing wrote to. Now SearchPaneModel restores
SelectedSearchMode on construction and writes back on change via
SettingsService, and AssemblyTreeModel walks the selection's
ancestor chain to find the owning AssemblyTreeNode, stores its
LoadedAssembly.FileName when IsAutoLoaded, and re-opens that file
(when it still exists) before RestoreSelectedPathAsync.
Assisted-by: Claude:claude-opus-4-7[1m]:Claude Code
The new Avalonia SessionSettings reads WindowBounds as Left/Top/Width/Height
attributes and reads ActiveTreeViewPath node values raw, but ILSpy 10.x wrote
WindowBounds as a CSV element body ("L,T,W,H", the Rect TypeConverter format)
and \xNNNN-hex-escaped every non-letter-or-digit char in each Node value (so
"TomsToolbox.Wpf" round-tripped through XML as "TomsToolbox\x002EWpf"). On
first launch against an existing ILSpy.xml that means the saved window
position resets to the default and the selected tree node never restores --
the escaped path can't match a live node's ToString().
Accept the CSV body when the WindowBounds element has no attributes, and
decode \xNNNN escapes in tree-view node values. Both fall back through the
existing ParseDouble defaults if a piece is missing, so a corrupted entry
won't crash startup.
Lock the shape in with tests against the actual section the WPF host emits,
plus a forward-compat pair confirming LoadFromXml ignores unknown children
(DockLayout, SelectedSearchMode, ActiveAutoLoadedAssembly) and SaveToXml
never echoes them back -- the AvalonDock schema doesn't translate to the
Avalonia Dock host and would otherwise persist forever as dead state.
Assisted-by: Claude:claude-opus-4-7[1m]:Claude Code
WPF showed a row Offset column and decoded the Kind GUID into one cell
combining heap offset, friendly name, and raw GUID; the Avalonia table
only carried the bare heap offsets. Offset follows the GetRowOffset
convention every other table uses. Kind is plain text rather than a
hex column plus tooltip so the per-column filter matches the friendly
names.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
Selecting a row in the CustomDebugInformation table now previews its
Value blob beneath it, completing WPF row-details parity: state-machine
hoisted local scopes, compilation options, metadata references, and
tuple element names parse into typed sub-grid rows; source-link JSON
shows as text; everything else — unrecognized kinds, embedded source,
and malformed structured blobs — degrades to a hex dump, as in WPF.
The parser switches on the Kind GUID directly via KnownGuids; the
entry deliberately carries no decoded-kind state, leaving that to the
kind-column work that builds on this.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
WPF ILSpy used DataGrid row details in three places: the COFF and
Optional header views (flag-bit breakdown of the Characteristics /
DLL Characteristics words, permanently expanded) and the
CustomDebugInformation table (decoded Value blob, visible when
selected). The Avalonia port already carried the data — both header
nodes populate Entry.RowDetails and the column builder skips it — but
nothing displayed it.
ProDataGrid builds the row-details template once per details element
and recycles the built control across rows, swapping only the
DataContext, so a WPF-style template selector would go stale;
MetadataRowDetailsControl re-runs its content factory on every
DataContext change instead. Per-row initial visibility has no
SetDetailsVisibilityForItem equivalent and recycling resets
AreDetailsVisible, so the view re-derives it in LoadingRow. Row
activation now ignores gestures originating inside the details
presenter — selecting text in a details blob must not navigate away.
This lands the infrastructure plus the two header consumers; the
CustomDebugInformation details build on the EmbeddedSource decoding
work and follow with it. ConfigurePage on MetadataTableTreeNode is
the hook that consumer overrides.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
Clicking inside the open flags filter popup could act on the
DataGridColumnHeader underneath: unhandled press/release pairs bubbled
out of the popup into the header (which treats them as a sort click
and flashes its pressed visual), and on X11 overlay popups the
light-dismiss machinery could even re-target a click's press directly
at the header while the popup stayed open.
The filter UI is now hosted in a Flyout attached to the funnel icon
instead of a hand-rolled Popup parked in the header's panel. Avalonia
positions Popup as the low-level primitive and recommends Flyout for
attached pickers: light dismiss, Escape-to-close, focus handling, and
theme-correct presenter chrome come built in (the old hard-coded white
background was also wrong in dark mode). The flyout content swallows
unhandled wheel and press/release events, since its internal popup is
still logically parented to the funnel inside the header.
A headless end-to-end regression test drives a real DataGrid with
overlay popup hosts (the X11 OverlayPopups=true configuration the app
runs with) and raw input through the funnel and every visible flyout
control, asserting the header never sorts and never receives an
unhandled press or release.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
Opening a package from a feed previously meant downloading the .nupkg by
hand and using File > Open (#2313). The new File menu command searches any
public V3 feed (editable package-source list, persisted as an MRU in the
ILSpy settings), offers the latest 100 versions in a dropdown, and downloads
into the NuGet global packages folder so the cache is shared with every
other NuGet consumer on the machine and nothing is fetched twice. The cached
.nupkg is then opened exactly like the regular Open command. Feed access
sits behind INuGetFeedClient so the headless test suite covers search,
paging, version selection, download, cancellation, and error surfacing
without touching the network.
Assisted-by: Claude:claude-fable-5:Claude Code
When enabled, switch sections are ordered by their case label value
instead of by the underlying branch's IL offset. Default is false to
keep existing output unchanged. Useful when diffing decompiler output
across rebuilds of obfuscated assemblies, where IL block layout is
unstable but the case-to-value mapping is not.
Includes an ILPretty test that exercises a hand-written switch whose
table targets are placed at non-monotonic IL offsets (simulating
obfuscator block shuffling) and verifies the cases come out in
label-value order with the setting enabled. Also adds the
Resources.resx / Resources.Designer.cs entry so the WPF settings UI
shows a proper label instead of the raw key.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Fix anonymous-type lambda early-return emitting unresolvable cast
When a lambda's inferred return type contains an anonymous type and one
branch returns null, the decompiler emitted an explicit cast such as
`return (IEnumerable<<>f__AnonymousType0<int>>)null;`, which is invalid C#.
Skip the cast in IsPossibleLossOfTypeInformation for null literals whenever
the expected type contains an anonymous type:
null is implicitly convertible to any reference type, so no cast is needed,
and the anonymous type has no nameable form to cast to anyway.
Fixes#3751
ECMA-335 names the zero state of most mutually exclusive sub-ranges
(NotPublic, AutoLayout, Class, AnsiClass, PrivateScope, ReuseSlot, ...)
and the reflection enums declare those members, but the schema inferer
skipped zero-valued fields and synthesised a generic '(none)' entry
instead. Zero fits every mask, so the member is attributed to a group
by declaration order: ECMA-335 and the reflection enums declare each
mask directly followed by its members. Filter persistence stores
numeric values, so saved filters are unaffected by the label change.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
The schema inferer only walks the enum's declared fields, so the
ECMA-335 Forwarder bit (0x00200000, no member in
System.Reflection.TypeAttributes) was invisible to the per-column
flags filter even though the cell tooltip already shows it. An
explicit per-enum extras table appends it as an independent flag,
making forwarders filterable on the ExportedType table.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
The blob was tracked 100644, so every fresh checkout (and any branch
switch in a worktree hosting the shared .git/hooks symlink target)
produces a non-executable hook script, which git then silently skips
with only an 'ignoredHook' hint. With core.fileMode=false on this
clone a plain chmod is git-invisible; the mode has to be fixed in the
index.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
0x00200000 is the ECMA-335 Forwarder bit (II.23.1.15), set on
ExportedType rows that forward a type to another assembly.
System.Reflection.TypeAttributes has no member for it, so the
enum-driven flag enumeration left it invisible in the Flags group.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
The #Strings/#US/#Blob nodes each open-coded the same offset-chained
walk, whose subtlety (handle 0 is a real entry despite being the nil
handle) is now documented once in WalkHeap. The TypeDef and
ExportedType entries built byte-identical TypeAttributes flag-group
tooltips; FlagsTooltip.ForTypeAttributes is the shared factory.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
The pane carried four hand-rolled try/catch export lookups that
predate AppComposition.TryGetExport and were missed when the other
call sites were converted.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
All six entity nodes repeated the same registry walk in LoadChildren;
AddAnalyzerChildren gives analyzer registration a single home, with
subclasses adding their entity-specific rows (accessors, backing
fields) before calling it.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
All six member-level Decompile* entry points wired the same assembly
resolver and debug info onto a fresh disassembler before dispatching;
the module-aware CreateDisassembler overload now owns that setup.
Assisted-by: Claude:claude-fable-5[1m]:Claude Code
Two Options tests selected the whole System.Linq.Enumerable type and then
awaited WaitForDecompiledTextAsync. That decompile takes >15 s in headless and
overran the 60 s wait on a loaded CI runner, so the suite flaked intermittently
(green in one run of a commit, red in another). The behaviours under test --
that a re-decompile setting refreshes the active tab, and that the refresh does
not steal focus from the Options tab -- fire on whatever the tab shows, so a
single small method exercises them just as well and decompiles near-instantly.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Selecting a C# version, switching to a version-less language (IL), then back to
C# left the toolbar's version ComboBox blank even though LanguageService had
restored CurrentVersion. The MVVM-toolkit setter runs OnCurrentLanguageChanged
(which assigns CurrentVersion, pushing it onto the bound SelectedItem) before it
raises PropertyChanged for CurrentLanguage (which repopulates ItemsSource from
the new language's versions). So SelectedItem is set against the previous
language's still-stale, empty list, rejected, and never re-pulled once
ItemsSource updates.
Re-assert the version combo's SelectedItem from the bound CurrentVersion after
each ItemsSource swap settles, keeping the model untouched.
Assisted-by: Claude:claude-opus-4-8:Claude Code
CaretHighlightAdornerTests waited on the wall clock (3s budget) for the
adorner's 1s real-time DispatcherTimer to remove it. Avalonia.Headless has no
simulated clock, so under a loaded CI runner the dispatcher-thread timer plus
the poll loop slipped past the tight budget and the test timed out
intermittently.
Extract the teardown the lifetime timer runs into Dismiss(); the test grabs the
live adorner and calls it directly, verifying the unregister half without any
timing dependency. Registration is still exercised through the real
OnReferenceClicked path, and production behaviour is unchanged.
Assisted-by: Claude:claude-opus-4-8:Claude Code
These three project settings were present in the WPF build but lost in the
move to Avalonia. RollForward=major matters most: the default publish is
framework-dependent (publish.ps1 uses --no-self-contained) and the app targets
net10.0, so without it a machine that only has a newer major runtime installed
fails to start the app. DynamicAdaptationMode (DATAS) lets server GC shrink its
heap count to the live load for a smaller idle working set, and Debug-only
CheckForOverflowUnderflow traps arithmetic wrap during development.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Generate-PDB and assembly-node decompile failures reported only the exception
Message, which is too thin to act on: the runtime-async PDB-generation crash
surfaces as a bare "Object reference not set to an instance of an object" with
no indication of where it came from. Append the whole exception (type, message,
stack trace, inner exceptions) inside a default-collapsed "Exception details"
fold so the report stays scannable while the trace is one click away.
The fold span is counted in WriteLine() calls rather than embedded newline
characters, so the trace is emitted one line per WriteLine() -- a single
Write() of the multi-line string would collapse to one counted line and be
dropped as a single-line fold.
Also fixes the PDB tab title showing a literal "{0}": the GeneratingPortablePDB
resource is a format string whose placeholder was never filled.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Debug.Assert failures were routed straight to the single-button unhandled-
exception dialog and deduplicated by call site, so the only available action
was effectively "ignore". Restore the developer dialog the WPF host offered:
Throw raises the assertion as an exception, Debug breaks into the debugger,
Ignore continues, and Ignore All suppresses that call site for the session.
Asserts fire on the background decompiler thread, so the dialog marshals to the
UI thread and blocks the caller on a nested dispatcher frame until a choice is
made, matching how the WPF version stayed responsive.
Separately, cancelling a long-running operation tab left it blank; it now shows
"Operation was cancelled." so the frozen tab explains why it stopped.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Toggling a display setting that affects decompiler output (folding, member or
using expansion, debug info, IL detail, indentation) routed through
ForceRefreshActiveTab, whose ShowSelectedNode re-projects the tree selection
and so activates the preview tab -- yanking the user off whatever tab they were
on (e.g. the live Options page they were editing).
Add RefreshDecompilerOutputInPlace, which re-decompiles the cached decompiler
content directly without re-projecting or activating anything, and route the
Redecompile display-setting reaction to it. F5 reload and dependency-load keep
using ForceRefreshActiveTab, where re-projecting the selection is intended.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Project/solution export (and other RunInNewTabAsync work) opened a tab with a
static title and a permanently indeterminate progress bar, even though the
whole-project decompiler already produces per-file DecompilationProgress that
was being dropped.
Wire that progress through: DecompilationOptions carries an
IProgress<DecompilationProgress> that DecompileAsProject sets on the
WholeProjectDecompiler; a new RunInNewTabAsync overload hands the work a
UI-thread-marshaled Progress<T> feeding the tab. The tab now animates the
braille spinner over its title (the operation name) and shows a determinate bar
plus the file currently being written and an "N of M" count. Solution export
reports per-assembly rather than per-file, since its assemblies decompile in
parallel and per-file reports would race. The determinate bar is pinned to 75%
of the overlay width so it doesn't jitter as the status text changes.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The README still described ILSpy as a WPF app. Reflect the current reality:
the desktop UI is cross-platform Avalonia (Windows/Linux/macOS), the solution
filters are ILSpy.Desktop.slnf / ILSpy.XPlat.slnf (no ILSpy.Wpf.slnf), Unix/Mac
can build and run the UI, and the build requires the .NET 11 SDK. In CLAUDE.md,
the net472 add-ins / installer / Windows-bound tests are Windows-only frontends
and extra tests, not a legacy porting backlog.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Make the repo-root pwsh scripts the mandated path for restore/build/update-deps/
format/clean/publish, and explain why: the bare dotnet commands prune the
packages.lock.json files (the recurring spurious lock-file diff). Add two code
conventions -- comments must read without any knowledge of the session that
wrote them, and edits made outside the session must not be reverted without
confirmation. Drop the "Framework-specific gotchas" section, which only covered
a few narrow UI control quirks.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The Debug Steps pane previously only served the ILAst language (one node per
IL transform). The C# decompiler's intermediate AST states were instead
exposed as a row of extra "C# - after TRANSFORM" dropdown languages, which
cluttered the language list and only let you jump to one fixed step at a time.
Generalise the pane to an IDebugStepProvider interface: ILAst keeps its IL
steps, and the C# language now contributes one step per AST transform, mapping
a selected step onto options.StepLimit (already honoured by CreateDecompiler).
Each language also supplies its own options object, so the writing-options
checkboxes show for ILAst and nothing for C#. The per-transform dropdown
languages are removed in favour of this.
Assisted-by: Claude:claude-opus-4-8:Claude Code
ILSpy targets net10.0, not net10.0-windows; the VS add-in build paths, the
installer output dir, the local-dev publish script, and the VS Code launch
config still referenced the old net10.0-windows layout. Align them with the
actual TFM, matching publish.ps1 and the build workflow.
Assisted-by: Claude:claude-opus-4-8:Claude Code
In pwsh the bare '-xr!*.pdb' 7z argument is split into two tokens, '-xr!*'
(which excludes everything) and a bogus '.pdb' source, so the Release
binaries zip added 0 files and the step failed. Quote the exclusion so it
reaches 7z as one token, and pin the job's run shell to pwsh via defaults
so no step silently inherits a different default shell.
Assisted-by: Claude:claude-opus-4-8:Claude Code
ICSharpCode.Decompiler / ILSpyX / BamlDecompiler set RestoreLockedMode but
declared no RuntimeIdentifiers, so their lock files were RID-agnostic. A
RID-specific 'dotnet publish -r <rid>' of ILSpy (which declares all four
distribution RIDs) propagated the RID to these locked project references and
failed restore with NU1004 unless --no-restore or --force-evaluate was used.
Declare the same win-x64;win-arm64;linux-x64;osx-arm64 set on the three
libraries and regenerate the locks so a locked-mode publish works for every
platform out of the box.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Pull in the submodule's new root Directory.Build.props so the decompiler
test fixtures aren't subject to this repo's TreatWarningsAsErrors=true.
Assisted-by: Claude:claude-opus-4-8:Claude Code
A clean build currently has zero warnings; enforce that so new ones can't
accumulate unnoticed. Set TreatWarningsAsErrors in the root
Directory.Build.props rather than per project. ICSharpCode.ILSpyCmd keeps
its existing TreatWarningsAsErrors=false (project-level wins over the
imported default), and the three legacy net472 projects (the VS add-ins
and the installer builder) opt out explicitly since they only build on
Windows and aren't covered by the cross-platform CI that gates this.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The Microsoft.DiaSymReader / .PortablePdb packages pull NETStandard.Library
1.6.1, which drags in 4.3.0 builds of System.Net.Http, System.Private.Uri
and System.Text.RegularExpressions, all carrying known advisories (NU1902/
NU1903). They are framework-provided at runtime on net10.0; pin the patched
4.3.4 / 4.3.2 / 4.3.1 via central transitive pinning so NuGet audit passes.
Enabling transitive pinning also aligns a handful of other transitive
packages to their declared central versions.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The 23 typed table-row viewmodels each inlined the same
MetadataOffset + GetTableMetadataOffset + GetTableRowSize * (rid - 1)
arithmetic, and the four heap tree nodes each repeated an identical
lazy-cache / "{name} Heap (count)" header / CreateTab shell that differed
only in the heap name and how rows are walked. A single mistyped
TableIndex or a divergent cache shell would have been invisible across
that many copies.
Add MetadataTableTreeNode.GetRowOffset (with a MetadataReader+offset
overload for the one entry that lacks a MetadataFile) and a generic
MetadataHeapTreeNode<TEntry> base carrying the cache, header, and tab,
leaving subclasses to supply only the name and row materialiser.
Assisted-by: Claude:claude-opus-4-8:Claude Code
DockWorkspace carried the reflection-heavy metadata-grid resolution: read a
clicked token cell, resolve a double-clicked row to its tree node, and find
the metadata-table node a token points at -- with the "read the row's
metadataFile private field + int token" reflection block copy-pasted between
the cell-click and row-activation paths. As the WPF host kept this metadata
logic out of the main workspace, lift it into a MetadataNavigator service.
DockWorkspace keeps the dock-side actions (select, scroll-to-row, history
stamping, open-in-new-tab) and now delegates resolution; the public
NavigateToToken / TryResolveRowToTreeNode surface is unchanged, so the
external callers (ShowInMetadata, DecompileMetadataRowInNewTab) don't move.
The duplicated reflection collapses into one TryReadRow.
Assisted-by: Claude:claude-opus-4-8:Claude Code
AssemblyTreeModel carried a ~180-line node-finding subsystem -- resolve a
reference (entity / namespace / resource / assembly / metadata file) or a
saved path to its tree node -- that has nothing to do with the class's
selection and list-lifecycle responsibilities. Lift the implementation into
a TreeNodeLocator, with the locator taking the tree root and assembly list
as parameters so it stays stateless.
The model keeps its public FindTreeNode / FindNodeByPath / GetPathForNode as
thin delegators (they're called from DockWorkspace, SearchPaneModel,
AssemblyTreeNode and tests), so no external call site changes -- the god
class just sheds the implementation.
Assisted-by: Claude:claude-opus-4-8:Claude Code
ApplyDocument was a 116-line god-method and the constructor a 112-line wall
of unrelated wiring. Both are linear sequences with no shared control flow,
so split them into intention-named steps:
ApplyDocument (now ~40 lines) orchestrates RebuildFoldings,
RestoreOrResetViewState and SwapCustomElementGenerators; the constructor
(now ~30 lines) calls SetupElementGenerators / SetupBackgroundRenderers /
SetupHoverHandlers / SetupContextMenuAndHyperlinks / SetupRichPopup /
SetupZoomAndCopy. Order is preserved exactly. Pure refactor, no behaviour
change.
The five renderer/generator/popup fields lose `readonly` since they are now
assigned in the ctor-only Setup helpers rather than the ctor body; they are
still single-assignment in practice.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The C# 13 'allows ref struct' anti-constraint is emitted as a single
PrimitiveType token (TypeSystemAstBuilder), so it reached WritePrimitiveType
rather than WriteKeyword and fell through uncoloured. Add it to the
value-type keyword group alongside the other constraint, 'unmanaged'.
Assisted-by: Claude:claude-opus-4-8:Claude Code
WriteKeyword coloured every modifier the decompiler emits except
'required' (C# 11 required members), which fell through and rendered in
the default text colour. Add it to the modifiers group.
Assisted-by: Claude:claude-opus-4-8:Claude Code
WriteKeyword/WritePrimitiveType/WriteIdentifier/WritePrimitiveValue each
ended with the same begin-span / base-write / end-span guard. Replace it
with `using (Colored(color)) base.WriteX(...)`, backed by an allocation-free
ref-struct scope so the per-token hot path takes no closure.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Siegfried Pammer
James May
Christoph Wille
danbopes
Jakob Hellermann