AD-98's fixed-canvas stretch (73041d70) scales every retained-UI quad at
TextRenderer.AppendQuad, but the live gate reported it JAGGED — text
especially. Cause: dat-font glyph atlases and IconComposer's composited
icons upload nearest (TextureCache.UploadUiTexture's UiNearestRepeat
sampler) — correct at the native 1:1 scale (pixel-exact retail art), but
aliased once magnified 2.4x1.8. Chrome/background art was already fine:
it uploads through GpuSamplerDescription.WorldRepeat (linear) by default.
Retail's own fixed-canvas presentation is a single bilinear-filtered
frame blit, never a per-texture stretch — this closes that gap one step
earlier, at the source texture, without adding RHI surface area.
- TextureCache.GetOrCreateLinearUiTwin: lazily registers a SECOND table
slot for a nearest handle's IGpuTexture, sampled WorldRepeat (linear)
instead of nearest — no re-decode, no re-upload, no extra memory-ledger
bytes. Returns the handle unchanged for anything never registered
nearest (chrome, UiTextureTableHandle.None), so it's a cheap
unconditional probe. Twin slots are released in Dispose without
double-disposing the shared texture.
- TextRenderer.LinearTwinResolver + the DrawSprite chokepoint: swaps a
sprite's texture handle through the resolver only while
CanvasScale != One. At CanvasScale == One the resolver is never even
called — zero overhead on the ordinary in-world/UI path.
- InteractionRetainedUiComposition wires the resolver to TextureCache
right after every UiHost acquisition (the lease can hand back a host
from a prior session against a fresh TextureCache).
- AD-98's register row gets one added sentence recording the fix.
Tests: TextRendererLinearTwinTests pins the renderer-side handle-swap
seam GPU-free (segment handle selection); TextureCacheLinearTwinTests
pins twin creation/reuse/dispose against RecordingGpuDevice. App suite
5097/3 skips (Release, ACDREAM_PROBE_LIVE_MOUNT=1 live-DAT probes
included). Full solution builds clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Third iteration on the screen, completing AD-98. The previous substitution
stretched only the root BACKGROUND while the child widgets stayed at their
authored 800x600 pixel positions - and the background painting carries
visual anchors (the World/Characters captions are art), so the user gate
showed captions overlapping the listbox and every widget misaligned
against the stretched art.
Retail model (established at 71bf24fb): fixed-canvas pre-world screens
render at authored 800x600 and the whole composed frame stretches once at
presentation; the blitter has no stretch mode. Our equivalent now does the
same one stage earlier:
- UiRoot.FixedCanvasSize: while the char-select screen is active, the
retained tree lays out in its authored canvas and Draw scopes a uniform
scale onto TextRenderer.CanvasScale; the mouse entry points apply the
exact inverse so MouseX/MouseY and every hit test live in canvas space.
- TextRenderer.AppendQuad is the single emission chokepoint - sprites,
rects, AND glyphs scale together, including retail-authentic non-uniform
aspect distortion and stretched text. World-space HUD stays native (the
scale resets outside UiRoot.Draw).
- CharacterManagementUiController stops resizing Root to the viewport;
activate/deactivate/dispose set and clear the host canvas.
- UiDatElement returns to retail-pure copy-or-tile; the interim
StretchOwnBackgroundToFill flag is deleted.
- AD-98 updated to describe the completed substitution.
Tests: canvas-scale quad math, inverse input mapping (window click lands
on the canvas-space widget), degenerate-size guards, controller keeps
authored extent + sets/clears the canvas. App suite 5085/6 skips; live-DAT
char-select probes 3/3.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The LA8 char-select root (0x1000039A) authors LeftEdge=TopEdge=RightEdge=
BottomEdge=0 ("no anchor") in the installed DAT — confirmed via the new
CharacterManagementLiveDatTests.RootAuthorsNoEdgeAnchors_RetailNeverResizesItSelf
gate — so retail's own UIElement::UpdateForParentSizeChange (0x00462640) never
resizes this element; it stays a fixed 800x600 rect in retail's own tree.
Retail's generic UI sprite blit, Graphic::Draw (0x00693b20) dispatching to
Graphic::PutImage (0x00693a30) for an exact/undersized destination or a
modulo-wrapped tile loop otherwise, has no third "stretch" mode — confirmed
against BlitMode (acclient.h ~3135) and MD_Data_Image::m_drawMode/DrawModeType,
both COLOR-blend selectors, not tile-vs-stretch geometry modes. The prior
"Normal -> tile, matching ImgTex::TileCSI" citation in UiDatElement was a
mis-attribution: ImgTex::TileCSI (0x0053e740) is called exclusively from
TexMerge::CopyAndTile/ImgTex::CopyCSI for LAND-SURFACE terrain texture
compositing, never from the UI element system.
Given the dat authors zero resize anchors and the blitter can only copy or
tile, the only way retail's whole pre-world scene (background + buttons +
listbox together) fills an arbitrary window resolution is that these
fixed-canvas "flow" screens render at 800x600 and the WHOLE FRAME is
stretched once at presentation — outside the UI sprite system entirely.
acdream has no offscreen fixed-resolution UI render target / present-time
scale pass; CharacterManagementUiController's constructor instead resizes
the MOUNTED ROOT element itself to the live viewport, which is why its own
background tiled (Width/tw > 1 at any resolution above 800x600, wrapped by
GL_REPEAT).
Fix: UiDatElement gains StretchOwnBackgroundToFill (default false, every
ordinary chrome/container element keeps tiling) — when set, the element's
own DirectState background draws as one UV-0..1 quad instead of the native
tile formula. CharacterManagementUiController sets it on Root right where
Root is resized to the host viewport, reaching the same visual result as
retail's present-time stretch (no tiling, no aspect-preserving letterbox)
through a different mechanism. Divergence register row AD-98 records the
substitution.
Tests: three new UiDatElementTests pin the UV-span mechanism generically
(tile past 1.0 when unset and rect exceeds native size; clamped to 1.0 when
set; byte-identical to the old tile formula when rect equals native size,
so every unaffected panel is untouched). CharacterManagementUiControllerTests
pins Root.StretchOwnBackgroundToFill == true post-construction. The live-DAT
gate confirms the root's zero edge-anchors and Type=3 against the installed
DAT. AcDream.App.Tests: 5084 passed / 3 skipped with ACDREAM_PROBE_LIVE_MOUNT=1
(5081/6 skipped without it — the 3 live-DAT-gated tests skip).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Root cause: the LA8 character-select screen's root background RenderSurface
(0x06007576, LayoutDesc 0x21000004 element 0x1000039A) is PFID_CUSTOM_RAW_JPEG
— a complete JFIF byte stream (confirmed live: 414,230 bytes, FFD8...FFD9,
Width=0/Height=0 on disk) that SurfaceDecoder.DecodeRenderSurface had no case
for, so it fell through the switch's `_ => DecodedTexture.Magenta` default arm
with nothing logged. Retail's RenderSurface::CreateFromSourceData
(named-retail decomp @0x004440a0) hands this exact byte stream to the Intel
JPEG Library (`_ijlInit`/`_ijlRead`/`_ijlFree`) at runtime and reads the real
pixel dimensions from the JPEG's own SOF header rather than this
RenderSurface's Width/Height fields, which are legitimately 0 for this
format — the same reason the decoder's generic non-positive-Width/Height
guard was also wrong to apply here.
A per-id media sweep of the installed DAT (new EveryDeclaredMediaId_
ResolvesToADecodableTexture test) showed this was the ONLY unresolved id
among the screen's 25 distinct media ids — the listbox (0x1000039D) and every
button face resolve fine. The listbox interior and the ENTER button's
circular fill are both transparent regions layered on top of the root, so
the one broken root background bled through everywhere nothing opaque
covered it, producing all three symptoms (full-screen background, listbox
interior, ENTER circle) from one cause.
Fix: SurfaceDecoder now special-cases PFID_CUSTOM_RAW_JPEG before the
Width/Height guard and decodes it with StbImageSharp (dual Unlicense/MIT,
pure managed, no native dependency — works on the Linux headless/graphical
targets Slice K/L commit to). JPEG is ITU T.81-standardized, so any
conforming decoder reproduces the pixels IJL would; round-tripped a
synthetic fixture through the real decode path to confirm. Verified against
the live DAT: 0x06007576 now decodes to 800x600, exactly the screen's
LayoutDesc-authored size.
Guard: per claude-memory/feedback_ui_resolve_zero_magenta.md, an unresolved
id reaching the draw path should be loud. That memory's existing guard
("guard on the id, not the handle") only covers a DIFFERENT trap — a
zero/absent id — and could not have caught this one, which has a real,
non-zero, DAT-resolved id. No guard existed for "id resolves but can't
decode" or "id doesn't exist in either dat" before this change, so both were
silent. SurfaceDecoder now logs once per surface id on every magenta-return
path (null data, JPEG decode failure, unsupported format, no-palette
paletted format, decode exception); TextureCache.GetOrUploadRenderSurface
logs once per id when a RenderSurface isn't found in Portal or HighRes at
all.
Tests: CharacterManagementLiveDatTests.EveryDeclaredMediaId_
ResolvesToADecodableTexture (installed-DAT gate, ACDREAM_PROBE_LIVE_MOUNT=1)
sweeps every StateMedia id in the char-select root + listbox row template
and asserts none decode to the magenta placeholder — this class of gap now
fails the gate instead of shipping silently. SurfaceDecoderTests adds
PFID_CUSTOM_RAW_JPEG coverage (real decode via a synthetic from-scratch
JPEG fixture — not retail art, generated with StbImageWriteSharp and
round-tripped before being pasted in as a literal; corrupt-data and
null-SourceData magenta paths) plus PFID_P8/PFID_INDEX16 no-palette cases
that now flow through the same logged path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A launcher-spawned client showed the world with NO interface at all -
character screen included. RetailUi rode ACDREAM_RETAIL_UI (the dev-era
opt-in), FromSessionConfig inherited the env parse, and the launcher
strips ACDREAM_* from children by design, so every product launch got
the dev default. A session-config launch IS a product launch: RetailUi
is now forced true on that path; the env flag remains the dev-launch
opt-in. Pinned by the session-config options test with a null env.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
F1: the crash reporter comment claimed the launcher never holds a password
in any field - false (ProfileEditorDialogViewModel, AccountProfile.Password,
StartRequest.Password). Reworded to the true, narrower invariant (no throw
site interpolates a credential VALUE into an exception message) and pinned
it with CrashReportNeverContainsAStoredPassword: a real STJ failure over a
profiles document containing a known password, corrupted after the
credential, must yield a crash file with the stack and without the value.
F2: the co-deploy Inputs covered only Bake own sources; a Content edit
never refreshed the 83 MB exe. Now the full reference closure. Fixing it
surfaced two more incrementality traps, both fixed and comment-documented:
SkipUnchangedFiles left the output older than the triggering input (target
re-ran forever - added an explicit Touch), and %(Item.Metadata) in a plain
Include does not batch (the literal percent-text became a permanently
out-of-date phantom input - globs are now spelled per project). Verified:
Core edit retriggers, then two consecutive clean incremental builds.
F3: RID publishes ran BOTH co-deploy paths (two self-contained bake
publishes). Build-time target now guarded on _IsPublishing; verified a
real win-x64 publish runs zero build-target co-deploys and still ships
both exes.
F4: comment misattributed PublishBakeTool=false to CI lanes; it is
target-local recursion guarding. F5: the x:Name reflection sweep now walks
the markup as XML and tolerates template-scoped names (no generated field
exists for those). F6: dead using removed. Hardening: the crash reporter
positional --data-dir fallback requires a fully-qualified path so a
relative or flag-shaped value cannot create ./crash-reports at an
arbitrary CWD.
Launcher 67/67, Launcher.Core 317/317.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
#398 was a crash on every modal open/close caused by MainWindow's
constructor calling AvaloniaXamlLoader.Load(this) instead of the
generated InitializeComponent() — only InitializeComponent assigns the
x:Name backing fields, so every named control was null and the first
Dispatcher.UIThread.Post callback in OnViewModelPropertyChanged threw
NullReferenceException, killing the process. It reached the user gate
because no test in tests/AcDream.Launcher.Tests (ViewModel-only) ever
constructed a MainWindow. #399 is the process gap that let that class of
defect through 14,012 green tests.
Adds Avalonia.Headless.XUnit 12.1.1 to the launcher test project. Its
net10.0 dependency group targets xunit v3, so the project migrates
xunit 2.9.3 -> xunit.v3 3.2.2 (drop-in: all 54 pre-existing tests compile
and pass unchanged under dotnet test via xunit.runner.visualstudio 3.1.4,
which already supported v1/v2/v3; two call sites needed
TestContext.Current.CancellationToken per the new xUnit1051 analyzer).
TestAppBuilder.cs wires [assembly: AvaloniaTestApplication] to a headless
AppBuilder.Configure<App>() so the real App.axaml FluentTheme is live in
tests.
MainWindowViewTests.cs adds 12 [AvaloniaFact]/[AvaloniaTheory] tests:
- an explicit non-null + type check of every x:Name field the
code-behind dereferences (ProfilesTree, ServerNameTextBox,
AccountNameTextBox, CharacterNameTextBox, EditorSubmitButton,
FirstRunDatDirectoryTextBox, FirstRunCloseButton, UpdateCloseButton)
- a reflection sweep over every x:Name found in MainWindow.axaml, so a
future named control without a matching non-null field fails loudly
- one open+close round trip per ProfileEditorKind (all seven, including
Remove), plus the first-run wizard and the update prompt, each pumping
Dispatcher.UIThread.RunJobs() so the queued focus callback actually
executes instead of just being asserted vacuously
- a dedicated test for the _focusBeforeModal-restore branch (not just
the ProfilesTree.Focus() fallback), anchored on a real focusable
button since ProfilesTree (TreeView) has Focusable="False" under
FluentTheme — its own tab stops are TreeViewItem rows, so the
close-path assertions check "no exception escaped the dispatcher"
rather than "focus landed on ProfilesTree"
Falsification (required evidence): reverting MainWindow's constructor to
AvaloniaXamlLoader.Load(this) and rerunning gives 12 failed / 0 passed —
10 tests throw NullReferenceException at MainWindow.FocusActiveModal,
propagating cleanly out of Dispatcher.UIThread.RunJobs() (confirming
dispatcher exceptions are not silently swallowed), and the 2 reflection
tests fail on an explicit "x:Name 'ProfilesTree' was null after
construction" message. Restoring InitializeComponent() gives 12 passed /
0 failed. Full launcher suite: 66 passed / 0 failed, reproduced on both
Windows and native Ubuntu (WSL, no display/Xvfb — Avalonia.Headless needs
none). AcDream.Launcher.Core.Tests: 317/317 unaffected.
No CI workflow change needed: .github/workflows/headless-portability.yml's
portable-launcher job already runs dotnet test on the launcher test
project on both windows-latest and ubuntu-latest with no display setup,
which is sufficient for Avalonia.Headless.
Closes#399.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Prove idle play remains passive and live until cancellation, then converges through one truthful status teardown. Keep probe mode string-only so numeric enum aliases cannot expand the pinned v1 contract, and record the Windows/WSL gates.
Agent was stopped for token budget. Landed here: probe flag through
LiveSessionConnectOptions + the StartCore short-circuit, the mode field
with JsonRequired-to-semantic-validation move, host exit-code mapping,
and 34 passing tests including 3 new probe tests (agent last reported
green before the stop). NOT DONE: the idle-policy unit tests (next
step), full-suite verification, and the WSL run.
Build/test state UNVERIFIED at this commit. Next session: finish idle
policy tests, run Runtime+Headless Release suites Windows and WSL, then
dispatch the Opus dual-lens review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Agent was stopped for token budget partway through the LA1 review fix
round. Landed here: F1 best-effort SessionStatusWriter, F2 App reader
tolerance (paths/mode), F5 argument-parsing hardening, plus new tests.
NOT DONE: F4 shared-fixture production shape (was the next step), F3
reconnect disconnected edge + recorded limitation, F6 exited
idempotency/reasons, F7 structural redaction test, F8 platform-guard
test + comment fix, optional RuntimeOptions PrintMembers redaction.
Build/test state UNVERIFIED at this commit. Next session: finish the
remaining findings, run the suites, then narrow re-review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Opus review of LA3 returned FIX FIRST; this addresses every finding in
scope (F1-F5, F7-F12; F6 CI-lane addition excluded per instructions):
- F1 (CRITICAL): SessionProcessSettings.Paths is now nullable and left
null by SessionConfigComposer unless a caller supplies overrides, so
the JSON key is entirely absent instead of "paths":{} — the App-side
loader's strict UnmappedMemberHandling.Disallow would otherwise reject
every gui/guiSelect session-config document at load.
- F2: added SessionConfigComposer.ComposeProbe and a nullable
SessionDescriptor.Mode field ("probe", omitted for normal play) per
the pinned contract — no character/policy/plugins/loginCommands.
- F3: LauncherProcessSupervisor.Stop now tries
ILauncherChildProcess.TryRequestGracefulStop (Linux: libc SIGINT via
LibraryImport, K4-proven graceful headless logout) before
CloseMainWindow. Windows has no reliable no-window-console equivalent
today; filed docs/ISSUES.md #397 with the CREATE_NEW_PROCESS_GROUP +
CTRL_BREAK fix direction. Stop()'s blocking-timeout contract is now
documented for LA4.
- F4: LauncherProfileStore.Save chmods the Linux temp file to 0600
immediately after creation, before any credential is serialized;
failure paths and Load() clean up a stale .tmp.
- F5: added LauncherCoreDependencyBoundaryTests asserting Launcher.Core
references exactly AcDream.Platform and no packages.
- F7: StatusEventParser.Parse no longer throws on a whitespace/null
line; StatusFileTailer.ReadNewEvents swallows the File.Exists/open
TOCTOU window (FileNotFoundException/DirectoryNotFoundException/
IOException) instead of throwing.
- F8: Start() now kills (entire process tree) and disposes a child that
started successfully but failed while being fed its stdin password,
instead of orphaning it.
- F9: SetState is monotonic — once Exited, no later transition applies
or fires StateChanged, closing a Start()-path race where a
synchronously-exiting child could be "resurrected" to Running.
- F10: CharacterIdFormat.TryParse now requires the "0x" prefix (an
unprefixed hand-typed decimal id is also valid hex and was silently
misread); a parsed id of 0 is treated as unusable and falls back to
the name selector; LauncherProfileStore.MergeRoster normalizes both
sides through TryParse/ToHexString instead of raw string equality, so
a legacy unprefixed-hex row self-heals via name match instead of
duplicating.
- F11: StatusCharacterEntry.SecondsGreyedOut is now uint, matching
CharacterRosterEntry and the host writer.
- F12: added MalformedStatusEvent, returned for a recognized `e` whose
payload doesn't match its shape, distinguished from UnknownStatusEvent
(an unrecognized `e`).
AllowUnsafeBlocks was added to AcDream.Launcher.Core.csproj — required
by the LibraryImport source generator's function-pointer marshalling
stub for F3's Linux SIGINT P/Invoke.
Verification: dotnet build AcDream.slnx -c Release green (0 errors);
dotnet test tests/AcDream.Launcher.Core.Tests -c Release green at 94/94
on native Windows and under WSL (Ubuntu, verified across multiple runs
for the timing-sensitive SIGINT/sharing-violation tests, no flakes
observed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The LA3 Opus review process note was right: the contract both sides
implement lived only in orchestrator prompts, which is exactly the drift
mode the pin exists to prevent (and it produced the paths-key CRITICAL).
The schema, field rules, probe-mode discriminator, and status vocabulary
are now a binding plan section; amendments change this text first,
implementations second. Ledger: LA3 fix round dispatched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Opus retail-lens review decoded the PDB-paired binary at
CPlayerSystem::RestoreCharacter@0x0055d760 and refuted the
uninitialized-edx justification: the two extra arguments are real
push imm32 of a constant PStringBase (BN mis-renders them, but they
pack to >=4 bytes each), so retail 0xF7D9 is >=16 bytes where ours
is 8. The guid-only CODE stands (ACE reads only the guid; holtburger
consensus) but it is an adaptation, not a corrected decompile — filed
as divergence register AD-97 and the doc comment now states the true
mechanism.
Also from the review: the 0xF643 conditional-parse doc now names BOTH
ACE flag-only failure branches (NameInUse + Corrupt); CharacterError
0x08 doc corrected (ACE misnames it ServerCrash2 — the port corrects
an ACE misnaming; ACE omits three values, not four); LA7b hazard notes
added (ACE silent no-reply on unknown restore guid; retail SendToLogon
vs SendToControl routing; NumErrors never rendered); two review-nit
tests (flag=0 Undef flag-only, non-Ok body with trailing bytes
ignored).
Core.Net suite: 953 passed / 0 failed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New AcDream.Launcher.Core (BCL-only, ProjectReference: AcDream.Platform
ONLY) plus tests/AcDream.Launcher.Core.Tests, both registered in
AcDream.slnx. This is the file-contract orchestrator core the Avalonia
launcher (LA4) will bind to — the game solution (Core/Runtime/App/
Headless) stays entirely out of this dependency graph, so the launcher
can never accidentally grow a game-protocol coupling.
- Profiles/: LauncherProfileStore owns launcher-profiles.json (spec §5
schema: version 1, servers[]/accounts[]/characters[]), strict
camelCase System.Text.Json (UnmappedMemberHandling.Disallow), typed
CRUD (add/edit/remove server; add/edit/remove account; edit character
settings), and MergeRoster (fold a reported roster into an account's
characters[] while preserving user-owned launchMode/plugins/
loginCommands, adding new rows with default guiSelect, and retaining
rows absent from the roster — they may be pending-delete). 0600 on
Linux via File.SetUnixFileMode after save.
- Launching/: SessionConfigComposer builds the pinned session-config
contract (Headless K1 shape + plugins/loginCommands/
loginCommandDelayMs/statusFile) from a profile character + install
record — character selector omitted entirely for guiSelect, policy
{id:"idle"} only for headless, credential always standardInput/
session. Passwords never enter this document (proven by a dedicated
test). LauncherProcessSupervisor spawns a host, feeds the password to
stdin then closes it, and exposes Starting/Running/Exited lifecycle;
Stop calls CloseMainWindow falling back to Kill after a timeout, both
reachable through an injectable ILauncherChildProcess/factory seam so
the state machine is unit-testable without real OS process timing.
- Status/: StatusEventParser decodes the v1 status.jsonl vocabulary
(started/connected/characterList/enteredWorld/pluginLoaded/
pluginFailed/disconnected/exited); an unrecognized "e" or a malformed
line degrades to a typed Unknown event rather than throwing.
StatusFileTailer incrementally reads new lines, tolerating a
not-yet-existing file and a partial trailing line (only advances its
read position past confirmed '\n' boundaries; a truncated tail is
simply re-read next poll, never parsed early).
- Integrity/: streaming SHA-256 + hex verify for later pak/download
checks (LA9/LA10).
Tests: 71 passed (profile CRUD + roster-merge matrix + strict-schema
rejection; composer golden-shape tests for gui/guiSelect/headless +
password-absence; supervisor tests against both an injected fake child
(state-machine determinism) and a real spawned `dotnet --version`
child (genuine cross-platform stdin/exit-code proof); tailer tests
incl. partial-line and not-yet-existing-file; SHA-256 tests). Verified
green on Windows (Release) and native WSL/Linux (Release) — the Linux
0600 test executes its real assertion body under WSL rather than
early-returning.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
