Repository navigation
Releases
Celtium 0.2.3
Celtium 0.2.3 — Multi-OS verified release
Released 2026-11-16 · tag v0.2.3 · previous v0.2.2
0.2.2-rc1 fixed a Windows-only failure whose root cause was a Linux-only assumption in the readable-range
guard. This release finishes that job: the whole repository was audited for OS-specific assumptions, the
defects found were fixed, and CI now runs the test suite on Windows and macOS as well as Linux, so the next
assumption of that kind fails a build instead of shipping.
What "multi-OS verified" means here
Exactly this, and no more: the complete celtium-vulkan test suite (94 suites / 462 tests) and the
repository's OS-portable checks (check_agent_docs.py, instrument-index.py --check) pass with identical
results on CI runners for all three operating systems — Ubuntu, Windows (windows-latest) and macOS
(macos-latest) — in the same workflow run
(run 43877915571). It does not mean the
game itself was run on real hardware on those systems; see Known limits.
Fixed
- The test build no longer assumes a Linux host. The
celtium-vulkantest runtime pinned the LWJGL
natives classifier tonatives-linux, so on a Windows or macOS host the test JVM extracted a Linux.so
and the shaderc binding failed its initialiser — the suite could not run off Linux at all. The classifier
now follows the host OS and architecture. - The shader compiler is findable on Windows and current installs. The PATH lookup probed only the bare
program name, which never matches a Windows.exe, and the candidate list knew only the pre-rename name
glslangValidator— the current Khronos Windows release ships onlyglslang.exe. The probe now tries the
.exeform on Windows,glslangis a candidate on every OS, and the not-found message names the macOS
and Windows install routes alongside Debian/Ubuntu. - The readable-range guard now covers macOS. The guard behind the 0.2.2-rc1 fix had providers for Linux
(/proc/self/maps) and Windows (VirtualQuery) only; on macOS every client-memory draw not covered by an
owned-buffer publication was refused — the same failure class as the Windows bug, one OS over. A macOS
provider now walks VM regions withmach_vm_region(FFM), behind the same testable seam as the other
providers. The new CI leg caught this before it shipped: its first run was green on Windows and red on
exactly the three client-array tests on macOS. - The audit's remaining findings were classified as by-design with safe degradation (for example, the guard
still refuses by default on any OS beyond the three above) or checked and clear; the full classification
is recorded with the change history.
CI
- New
vulkan-multi-osmatrix job (Windows, macOS) running:celtium-vulkan:testplus the OS-portable
check scripts, with glslang 16.6.0 pinned from the official Khronos release archives. The legs
deliberately rely on the build's own PATH resolution for the shader compiler, so the Windows.exe
lookup is itself under test. The existing Ubuntu jobs are unchanged, and the Forge packaging build
(:forge122:packageJar) stays on the Ubuntu legs — the shipped jar is platform-independent.
Known limits
- No real-machine in-game verification on Windows or macOS: actual rendering under a real install (for
example EfficientEntities on Windows, or MoltenVK on macOS), frame-time impact, and the FFM bindings
under the game's own JVM remain unverified on hardware. The CI legs exercise the bindings in a test JVM
on hosted runners; they narrow the class of bug that can ship, they do not close it.
Celtium 0.2.2-rc1
Celtium 0.2.2-rc1 — Windows entity rendering fix
Released 2026-11-16 · tag v0.2.2-rc1 · previous v0.2.2
Release candidate for a single fix on the 0.2.2 line. On Windows, entity and player models drawn from client
memory silently failed to render — only the shadow appeared. This candidate fixes the readable-range guard
behind that failure. Windows real-machine verification is still pending, hence the rc.
Fixed
- Entity and player models render on Windows again. The GL stub's client-pointer fault guard determined
readable memory from/proc/self/maps, which exists only on Linux. On Windows the guard therefore refused
every draw from client memory that was not covered by an owned-buffer publication, so models drawn from their
own direct buffers (for example EfficientEntities'ModelBiped/ModelPlayerandModelSlime) never
appeared, while the shadow — drawn through the BufferBuilder path — remained. - The guard now uses an OS-independent readable-range check (
ReadableMemoryRanges): Linux keeps the existing
maps parser with unchanged behavior; Windows resolves committed, readable regions viaVirtualQuery(FFM);
other OSes keep the previous refuse-by-default behavior. If the Windows binding fails, the guard degrades
safely to refusal, as before. The guard's other call sites (drawElements indices/vertices, split-array
repack) go through the same entry point and benefit identically. - Regression tests cover the range arithmetic, the Windows protection attributes and region walking (against
fake functions), and the stub's accept/refuse behavior.celtium-vulkan: 94 suites / 459 tests, 0 failures.
Known limits
- No Windows real-machine run yet: the
VirtualQueryFFM binding, the census reporting readable ranges, and
actual rendering under EfficientEntities (plus frame-time impact) are unverified on hardware.
Celtium 0.2.2
Celtium 0.2.2
Released 2026-11-07 · tag v0.2.2 · previous v0.2.1
English: 0.2.2 release notes (English)
0.2.1 のホットフィックスに、round 15 の成果(ウィンドウpollの省略と、mod framebufferのdepth破棄)を
含めたチェックポイントです。-Dceltium.vulkan.validation=true を付けた環境で、validation layer が無い/到達できない
場合に起動できなかった問題も修正します。
修正
- validation を要求しても、使えなければ起動を続ける。これまでは
ValidationBridge.enableの
InvocationTargetExceptionが「validation support is present but could not be reached」としてデバイス作成を
8回失敗させ、ゲームが起動しませんでした(テスターの AMD 環境で再現)。 - 起動前に layer /
VK_EXT_debug_utils/ LWJGL binding を確認し、使えない場合は
[Celtium] CELTIUM_VALIDATION_UNAVAILABLE reason=... (continuing without validation)を出して続行。 - messenger の作成に失敗した場合も同じ扱い(起動を止めない)。
- 反射呼び出しの
InvocationTargetExceptionを原因メッセージ付きで報告するようにし、次回は
「layer が無い」「binding が古い」などがそのままログに出ます。
round 15 の変更
- ウィンドウpollの省略:
FrameCoordinator.submitCommands()がRECORDED(バッチ経路でフレームのコマンドバッファに
残った)またはSKIPPED(デバイス停止)を返した場合、そのウィンドウの終わりにceltium$pollCompletedSlots()を
呼ばない。キューに何も渡っていないので、pollは前回と同じフェンスを読み直すだけ。確認は次フレームの
beginCommandsが行う(遅延であって喪失ではない)。KPI scene でslotPolls20.6 → 1.03/フレーム、
frameFenceWaitMillis68 / 55 で新しいストールなし。WindowPollDecisionTestが規則を4ケースで固定。 - mod framebufferのdepth破棄: modのbloom chainはdepth attachmentを持たないため、そのパスはpairのdepthを
load/storeする必要がない。VulkanTerrainPipeline.createChainPairRenderPassが同じattachment/subpassで
depthをCLEAR/DONT_CAREにしたパスを作る(互換規則はformat・sample・referenceのみで、load/storeは含まない)。
-Dceltium.vulkan.chainPairDiscardDepth=falseで旧パスに戻せる。celtium$windowBeginInfoはCLEAR用の
clear value(depth 1.0)を渡す(r15depth-val-amdの VUID-00902 を解消)。
検証
- build:
:celtium-vulkan:test :forge122:packageJar :forge122:validationRemapJar成功。 scripts/tests183 OK、route verifier 259/87 failed=0。- AMD RX 6400 の leg
r16val2(validation=true、plain、240 s): ゲーム起動、
CELTIUM_VALIDATION messenger=0x... layer=VK_LAYER_KHRONOS_validation、validation メッセージ 0。
fixture は plain scene の較正なし(UNCALIBRATED_PRESENTED_IMAGE)だが、起動失敗ではない。 - AMD RX 6400 の leg
r17wip(world、guiScale 3、validation on、240 s): PASSED、
CELTIUM_RECORD chain-pair render pass created (depth cleared and discarded)、
validation メッセージ 0(VUID-00902なし)、picture gate PASSED。
インストール
- Cleanroom Relauncher(0.7.12-alpha 以降、0.7.14 検証済み / LWJGL 3.4.3 / ゲーム JVM Java 25+)。
celtium-forge-mc1.12.2-0.2.2.jarをmods/に置く(Celeritas は入れない)。- validation を使う場合も、layer が無ければ警告を出して通常起動します。
既知の課題
0.2.1 と同じ(p99 の pair 外れ、modpack のログ全文/画像比較 pending、葉の mipmap、ATM3 の空、
RLCraft の ordering flush、GTCEu bloom の既定 off)。本リリースは起動修正のみです。
Celtium 0.2.1
Celtium 0.2.1 — faster than Celeritas in a real modpack
Released 2026-11-06 · tag v0.2.1 · previous v0.2.0
English: 0.2.1 release notes (English)
0.2.0 からの 189 commits。実modpack(Supersymmetry 0.2.0.2、188 mods)で Celeritas を p50/p95 で上回り、
有名modpack 7本が AMD/Intel で world に到達(validation 0)。mod 自前の GLSL program の native 実行、
shaderc natives の同梱、描画・同期の修正、巨大クラスの分割を含みます。
FPS 目標はまだ未達: p99 はほぼ同等(AMD 0.993 / Intel 0.993)だが、各GPU 1対が p99 で 1 を超えるため
goalCondition=NOT_MET。p50/p95 の優位は再現済み。
ハイライト
- 実modpackで Celeritas を上回った。Supersymmetry 0.2.0.2(188 mods、qmagnet、1280×720、validation off、
3対×240 s):- AMD RX 6400: p50 0.976 / p95 0.935 / p99 0.993(Celtium/Celeritas の frame time 比)
- Intel UHD 630: p50 0.940 / p95 0.970 / p99 0.993
- 2回目のAMD run(
l28kpi-amd)は p50 0.918 / p95 0.905 / p99 0.991。
- 有名modpackの互換拡大: SkyFactory 4 4.2.4、All the Mods 3 6.1.1、RLCraft 2.9.3、SevTech: Ages 3.2.3、
Nomifactory CEu 1.7.7、Project Ozone 3 3.4.11F、E2E-E 1.87.3(≈430 mods)が AMD/Intel で world 到達・validation 0。 - mod 自前 GLSL の native 実行(opt-in
-Dceltium.vulkan.modPrograms=true)。GTCEu bloom を native パスで実行し
validation 0。既定は program ごとに拒否・計数(安全側)。LWJGL の **shaderc natives(linux/windows x64)**を同梱。 - 描画・同期の修正: raw framebuffer bind が main target に描く、mod particle layout の PARTICLE への再パック、
SkyFactory 4 の空が平坦な青になる不具合、atlas の mip level の VUID-09600、acquire semaphore の再利用、
renderer 起動失敗時の crash report。 - コード整理:
VulkanTerrainGpuBackend22,576 → 21,038 行。IndirectImmediateRuns / ModPrograms /
FrameStatsReport / GroupLighting / GroupChange / PixelSignatures / HighlightProbe を分離。 - 検証ハーネス:
pack_verify.py(ログ分類と Celeritas 画像比較)、modpack preset(sf4/atm3/rlcraft/sevtech/nomi/po3)、
modpack-reverify.sh。
性能(Performance)
測定環境: Supersymmetry 0.2.0.2(GregTech CEu ベース、188 mods)、Cleanroom Relauncher、save qmagnet、
pose x=217.5 y=48 z=217.5 yaw=180 pitch=10、render distance 12、GUI scale 3、1280×720、vsync off、validation off。
build r107(5ff004c26)、3対×240 s/GPU(l27kpi-amd / l27kpi-intel、2026-11-06)。
| GPU | Celtium | Celeritas 2.5.0-dev | 比(pair中央値) | |
|---|---|---|---|---|
| AMD RX 6400 | p50 | 427 FPS | 407 FPS | 0.976 |
| p95 | 227 FPS | 214 FPS | 0.935 | |
| p99 | 174 FPS | 179 FPS | 0.993 | |
| Intel UHD 630 | p50 | 296 FPS | 273 FPS | 0.940 |
| p95 | 167 FPS | 159 FPS | 0.970 | |
| p99 | 139 FPS | 138 FPS | 0.993 |
- 比は frame time の Celtium/Celeritas(小さいほど良い)。FPS は参考値で、pair中央値の中央値。
- p99 は AMD pair 2(1.121)と Intel pair 2(1.081)が 1 を超え、
goalCondition=NOT_MET。
6 AMD pair の p99 は 0.924 / 1.121 / 0.993 / 0.991 / 0.942 / 1.058(中央値 0.99)。 - 0.2.0 リリース時点は同packで p50 約16%負け。main でここまで回復。
- 証拠: animated-sprite-uploads-2026-11-05.md。
modpack 互換
| Pack | Version | Mods | 結果 |
|---|---|---|---|
| Supersymmetry | 0.2.0.2 | 188 | KPI pack。validation-on legs pass |
| Enigmatica 2: Expert - Extended | 1.87.3 | ≈430 | world・validation 0(AMD/Intel)、Xaero raw-FBO identity も 0 |
| SkyFactory 4 | 4.2.4 | 198 | world・validation 0(AMD/Intel)。ログ全文と画像比較は pending |
| All the Mods 3 | 6.1.1 | 250 | world・validation 0(AMD/Intel)。pending |
| RLCraft | 2.9.3 | 165 | world・validation 0(AMD/Intel)。pending(Bloodmoon は Cleanroom 非互換で除去) |
| SevTech: Ages | 3.2.3 | 247 | world・validation 0(AMD/Intel)。pending |
| Nomifactory CEu | 1.7.7 | 149 | world・validation 0(AMD/Intel)。pending |
| Project Ozone 3 | 3.4.11F | — | 画像一致・Celtium由来の報告なし |
「verified」の定義(validation 0+ログ全文+Celeritas画像比較)と pack ごとの置換は
docs/modpacks.md を参照。pending は validation 0 のみで、ログ・画像は未確認。
追加・変更(Added / Changed)
- mod program:
glUseProgram系のルート、program ごとの拒否・計数、opt-in の native 実行、shaderc natives 同梱。 - harness:
pack_verify.py、modpack-reverify.sh、pack preset(sf4/atm3/rlcraft/sevtech/nomi/po3/e2ee)、
CELTIUM_E2E_PACK、CELTIUM_WARM_WINDOW、leg ごとの pack copy 解放。 - fixture: world switch の title screen 対応、getter 経由の counter 読み、presented counter の再起動、
keepEntities、cameraY、worldType。 - 診断:
CELTIUM_RECORD(重複排除)、CELTIUM_PRESENTheartbeat、CELTIUM_PROGRAM_BOUND。 - コード: 巨大クラス分割、孤立 javadoc 整理、未使用 switch/分岐の削除。
修正(Fixed)
- raw framebuffer bind が main target に描く(ATM3 の一人称手)。
- mod particle layout を PARTICLE に再パック(FBP 粒子の拒否)。
- SkyFactory 4 の空が平坦な青になる(非ITEM群の照明payloadがfogを失う)。
- atlas の mip level 書き込み後の layout(VUID-vkCmdDraw-None-09600)。
- acquire semaphore の再利用(image 再取得後)。
- renderer 起動失敗時に crash report を出す(Loading terrain で止まらない)。
- mod program pipeline のキー(参照)と world teardown での解放。
- indirect run の group 合流条件(pipeline/descriptor/buffers/alpha)。
- 空の arena mesh の stride、transformed vertex の一括コピー、window の index 上限 32768。
- 描画前の published-output layout 検査、mod pass の入力可視性、window prefix の target。
既知の課題(Known issues)
- p99 の pair 外れ: AMD pair 2 / Intel pair 2。
goalCondition=NOT_MET。 - modpack の完全検証: SF4/ATM3/RLCraft/SevTech/Nomifactory は validation 0 のみで、ログ全文と
Celeritas 画像比較が pending。 - 描画欠陥: 葉の暗化(mipmap)、ATM3 の空グラデーション、RLCraft の ordering flush / destroyed mesh、
GTCEu bloom は既定 off(-Dceltium.vulkan.modPrograms=trueで有効)。 - 未実装: cube map / 3D texture / MSAA / compute / blit / copyImage / 一部 DSA / stencil。
- 影・shaderpack: Demonica fork の別 addon。
検証(Verification)
- build:
:celtium-vulkan:test :forge122:packageJar :forge122:validationRemapJar成功。 - tests:
celtium-vulkan422、scripts/tests183。 - route verifier: 259 declarations / 87 classes / failed=0。
check_agent_docs/instrument-index --check: OK。- JAR:
celtium-forge-mc1.12.2-0.2.1.jar。
インストール
- Cleanroom Relauncher(Cleanroom 0.7.12-alpha 以降、0.7.14 検証済み / LWJGL 3.4.3 / ゲーム JVM Java 25+)。
celtium-forge-mc1.12.2-0.2.1.jarをmods/に置く(Celeritas は入れない)。- 起動し、ワールド・GUI・modpack を確認。
- 問題報告には
logs/latest.log、GPU、pack/version、JAR の sha256 を添付。
統計
- commits: 189(docs 76 / diag 20 / perf 18 / fix 16 / test 11 / feat 7 / harness 6 / refactor 5 ほか)
- tests: 422 + 183
- route verification: 259 declarations / 87 classes
- 証拠: docs/evidence、
docs/plan.md
Celtium 0.2.0
Celtium 0.2.0 — GL-less surface, measured
Released 2026-11-04 · tag v0.2.0 · previous v0.1.4-alpha-rc.26
English: 0.2.0 release notes (English)
0.1.4-alpha-rc.26 からの 393 commits(fix 177 / perf 125 / feat 30 / docs 32 / chore 10 / measure 6 /
refactor 5 / test 3 / revert 1 / profile 1 / glow 1)。GL-less レンダラの「起動して絵が出る」土台を、
実測・CPU検証・シーン比較で固めたリリースです。
FPS 目標は未達: Celeritas との paired KPI は 1.189(
NOT_MET)。描画忠実度とエラーゼロは
大きく前進しましたが、フレーム時間ではまだ負けています。詳細は「性能」と「既知の課題」。
ハイライト
- ビルドが通る状態に戻した。上流由来の未コンパイル(
RenderSectionの metrics アクセサ、
mixin の注釈位置、build-1122に Lombok のバージョンが無く annotation processing が動いていなかった)
を修復し、packageJar/validationRemapJarが成功する。 - transformer 生成バイトコードの CPU 検証器を追加。実 LWJGL 宣言 241 件と実 GL クラス 87 個を
ロード検証し、実行時 VerifyError を 4 件修正。うち 1 件は GL20C 自体がロード不能で、
シェーダ経路が起動できない状態だった。 - インデックス描画を実装。クライアント/VBO 頂点、クライアントインデックス、baseVertex、
multi-draw、instance=1 の派生形、12/16/20/24/32 バイトのレイアウトと 32 バイトの再パック。 - レガシー即時モード(glBegin/glEnd)を既定で描画。タイトル画面・メニュー背景・パックメニューの
アートがフレームに載ることをシーンで確認。 - クリエイティブ画面のシーンとアイコン、ツールチップの移動フレーム、モデルリプレイのホットパスを
計測・修正。 - GL 表面の大規模修復: 到達不能だった display list / pixelStore / buffer / texture / FBO /
shader getter 群の再接続、pname 取り違え、未初期化出力、固定幅書き込みのオーバーラン、
VAO 別属性、push/pop 属性スタック、ライト/フォグ/マテリアル、バインディングクエリ。 - 性能計測基盤: 同一 JAR・同一 leg 内 A/B、control arm によるマシンドリフト相殺、
JFR プロファイル、ウィンドウ/グループ/受理経路の内訳。
破壊的変更・移行
- バージョン:
0.2.0。-Pceltium_release_versionで明示ビルドする運用は従来どおり。 - ビルド環境: Gradle を JDK 25 で起動(RetroFuturaGradle 2.0.4 の要件)。
JAVA_HOME=/usr/lib/jvm/java-25-openjdk-amd64 scripts/build.sh。
build-1122/gradle.propertiesにlombok_versionが必須(無いと@Getterが生成されない)。 - glstub の適用範囲は不変: gl を直接呼ぶ mod 以外には適用しない。
- 設定: 既存の
-Dceltium.*は維持。新規の実験スイッチは既定で安全側
(例:-Dceltium.vulkan.mainClearInQueueは既定 off、-Dceltium.vulkan.replayValueVersionsは既定 on)。 - 配布物:
celtium-forge-mc1.12.2-0.2.0.jar(sha256bf1193b6ba21dc51cd36a5b013607dd58a9ef8bcbc516ca8f24b0c247e9cccee)。
追加(Added)
描画経路
- インデックス描画:
glDrawElements/glDrawRangeElements/glDrawElementsBaseVertex/
glDrawElementsInstanced/glDrawElementsInstancedBaseVertex/glDrawElementsInstancedBaseInstance/
glDrawElementsInstancedBaseVertexBaseInstance(instance=1 は通常描画へ転送)。 - マルチドロー:
glMultiDrawArrays/glMultiDrawElements(サブ描画をデコード経路で再生)。 - レガシー即時モード: glBegin/glEnd の全頂点綴り(float/double/int/short/byte)、glRect*、
マルチテクスチャ座標、全描画モード。 - 描画トポロジ: lines / line strip / line loop / triangle strip / quad strip / polygon の受理と変換。
- 属性:
glVertexAttribPointer/glVertexAttribIPointer(index 0)、normalized/integer フラグ、
VAO 別の属性記録、glPushClientAttrib/glPopClientAttrib/glPushAttrib/glPopAttribの復元。
GL 表面
- display list:
glNewList/glEndList/glCallList/glCallLists/glListBase/
glDeleteLists/glIsList。 - pixelStore: UNPACK/PACK の記録・クエリ・アップロード変換への反映。
- buffer: storage / sub-data / copy / readback / map(DSA 含む)/ clear(零埋め)/ binding クエリ。
- texture: storage(DSA 含む)/ sub-image(DSA)/ parameter(f/iv/fv/Iiv/Iuiv、DSA)/
border colour / level parameter(DSA)/ image readback のゼロ化。 - framebuffer: EXT/ARB 綴り、Named 形式のアタッチメント・状態・パラメータ、renderbuffer の extent。
- shader: info-log(ARB 含む)、active uniform/attrib、program binary、precision、resource 系、
compile/link status、not-found ルックアップ(-1)。 - 同期:
glGetSyncivは SIGNALED、クエリ結果は可視(1)。 - デバッグ:
glGetObjectLabel/glGetObjectPtrLabelの空ラベル。
シーン・検証
- メニュー/タイトル/パックメニュー、クリエイティブ画面、ツールチップ移動、モデル、空、雲、
フォグ、発光 outline のシーンと参照比較。 - ルート検証器(
scripts/cpu-contract/transformer-route-verify.sh)を CPU 契約に追加。 - KPI の control arm(同一 leg 内で candidate / Celeritas / 過去ビルドを交互計測)。
変更・改善(Changed / Improved)
描画忠実度
- タイトル画面・メニュー背景・パックメニューのアートをフレームへ(シーンで確認)。
- クリエイティブ画面のアイコン欠落(空の item plan キャッシュ)を修正。
- ツールチップ: 移動フレームで 1 フレーム 1 ツールチップ、mod FBO バインド中の取りこぼし/残像を修正。
- モデル/カスタム描画: ミラー供給の共通化、属性 0 の記述、VAO 別ポインタ、インデックス経路。
- 空/雲/フォグ/ライト/深度/ブレンド/シザー/ポリゴンオフセットの状態反映を整理。
- 半透明 terrain の背面 cull、combine 中の alpha test 維持、色付きガラスのちらつき修正。
状態・クエリ
- pname の取り違え(SKIP_PIXELS/ROWS、SPOT_CUTOFF、VIEWPORT_DIMS、SHADE_MODEL 等)を修正。
- 未初期化出力の一掃(sampler/renderbuffer/multisample/uniform/object label 等)。
- 固定幅書き込みのオーバーラン修正(light/material の成分数、boolean の 1 バイト)。
- バインディングクエリ(element/pixel/uniform/copy buffer、texture 1D/3D/cube、renderbuffer)。
- 能力値の整合(texture units、anisotropy、attachment/draw buffers、uniform vectors、stack depth)。
性能
- モデルリプレイ: 入力キャッシュ、値バージョン化、ライトペイロードの共有、行列/色/シザーの共有。
- グループ化: 状態キー、フォグ/シザー/ポリゴンオフセットのキャプチャ、ウィンドウ計測。
- クリア: 順序 flush のサイト計測、メイン FBO クリアのキュー搭載(既定 off、拒否 census 付き)。
- ホットパス: 単一参照化、プロパティ読みの定数化、アロケーション削減。
- フォント: バッチの行列保存をスクラッチ化、colour witness の絞り込み。
修正(Fixed)— 主なもの
- 起動/ビルド:
RenderSectionの metrics アクセサ欠落、@Injectの位置 2 件、@Shadowの二重付与、
Lombok の annotation processing 不動作。 - 実行時クラッシュ:
glClearBufferSubDataのスロット、glPushClientAttribの引数欠落、
glDrawElementsInstancedBaseInstanceのスロット、nglGetShaderInfoLogのスロット(GL20C ロード不能)。 - テクスチャ: ミラー未供給、PBO アップロード、DSA アップロード、サブイメージのパッチ、
削除後の名前再利用、サンプラ推論(wrap/filter)。 - FBO: アタッチメント記録、EXT/ARB/Named 綴り、削除時のクリーンアップ、読み出しのルーティング。
- 状態: raw の depth/alpha/lighting enable、push/pop、VAO 属性、ライト/フォグ/マテリアル、
ピクセルストア、バッファバインディング。 - クエリ: 未初期化出力、pname 取り違え、幅の誤り、not-found の 0(-1 へ)。
- 描画: インデックスのレイアウト/オフセット/再パック、クライアントメモリ、instance=1、
ライン/ストリップ/ポリゴン、属性 0。 - GUI: メニュー背景、クリエイティブタブ、ツールチップ、HUD の色/深度。
性能(Performance)
対 Celeritas paired KPI(modpack 環境、AMD RX 6400 / RADV NAVI24、1280×720、3 対)
測定環境(重要): 素のバニラではなく、テスターと同じ Cleanroom Relauncher の modpack インスタンス
celtium-world-accept(192 mods、save qmagnet)で実行しています。シーンは world、pose は
x=217.5 y=48 z=217.5 yaw=180 pitch=10、guiScale=3、renderDistance=12、maxFps=260、vsync off、
1280×720。1 leg 240 秒、warm window は world tick 400–600(hold 620)。
KPI leg は計測税を除くため validation off(-Dceltium.vulkan.validation=false、
presentReadback=false、presentedDumpEvery=0)で、Celeritas 2.5.0-dev(sha256 567dcf8a…)と
同一シーン・同一 JAR 条件で交互に計測しています。control arm は round 58 のビルド(de8ed2e7…)。
validation 0 の証拠は別 leg(validation on、r64c-carry 等)で取得したもので、KPI leg の
validationMessageCount=0 は validation off のため、クリーン性の証拠には数えていません。
| rep | Celtium p50 | Celeritas p50 | 比 | 過去ビルド比 |
|---|---|---|---|---|
| 1 | 2.742 ms | 2.332 ms | 1.176 | 0.985 |
| 2 | 2.869 ms | 2.413 ms | 1.189 | 1.063 |
| 3(負荷混入) | 3.620 ms | 2.311 ms | 1.566 | 1.380 |
| 中央値 | 1.189 | 1.063 |
goalCondition=NOT_MET。clean な 2 rep は 1.176 / 1.189 で、過去ビルド(round 58)比は −1.5% / +6.3%。- rep 3 は
finalLoadAverage0.96 → 3.98 の混入で、同型の汚染は round 63 でも観測済み。 - フレーム内訳(出荷コード、AMD): frame 1322 µs = world 861 + gui 421 + hand 50。
85% は Celtium の即時経路(受理 161 offer × 4024 ns = 648 µs + flush 392 + guiFlush 81)。
単位は offer 160.8 / group 131.9 / 窓 23.4。 - 窓の起点は framebuffer bind(
flushSiteCounts=[bind-framebuffer=792]、flushReasonsの 65%)。
clear 起点という仮説は棄却済み。 - 残る構造レバー: bind ごとの「frame image を実際にサンプルするか」census と不要 flush の削除、
accept path の未計測区間(offPreflight〜offAppend の約 470 行)の bracket、
bindless immediate textures、不透明・depth-write・非 blend 用の第 2 open group。 - 棄却済み(再試行しない): sticky pass 統合(2.85→11.47 ms)、font run の行列 bake(94→8 run)、
clear 起点の flush 削減(窓/フレーム不変)、stateKey の format 要因(hit 21.9% のまま)。
検証(Verification)
- ビルド:
:celtium-vulkan:test :forge122:packageJar :forge122:validationRemapJar成功。 - テスト:
celtium-vulkan422 tests / 0 failures、scripts/tests177 tests OK。 - ルート検証:
transformer-route-verify.sh→ 実 LWJGL 宣言 241、実 GL クラス 87/87、
colour 32/32、texcoord 64/64、query 8/8、failed=0。 - 文書:
check_agent_docsOK、instrument-index --checkcurrent。 - 実機 leg: KPI 9 legs はすべて
fixtureStatus=PASSED、gpuRequirementMet=true、problems=[]
(KPI leg は validation off。クリーン性は validation on の別 leg で確認)。world 画像
(terrain/blocks/items/HUD/minimap/hotbar/hand)を目視確認。 - JAR 検査:
verify_forge_jar.py/verify_version_consistency.pyともverified: true。
既知の課題(Known issues)
- FPS は Celeritas 比 1.189 で未達。accept path とウィンドウが残。Intel の p99(1.048)も未解決。
- メニュー/ツールチップ/クリエイティブタブはシーンで確認済みだが、テスター環境での再確認が必要。
- 未実装: cube map / 3D texture / MSAA / compute / blit / copyImage / 一部の DSA /
stencil(ModularUI の Stencil#reset は mask 0x400 で未実装)。 - 実機 leg は環境依存: 検証ハーネスの Prism インスタンス構成に依存。配布 JAR は別途テスターで確認。
- 影/シェーダパック: Demonica fork の addon として別成果物。core には取り込まない。
インストール
- Cleanroom Relauncher 構成(Cleanroom 0.7.12-alpha / LWJGL 3.4.3 / ゲーム JVM は Java 27)を用意。
celtium-forge-mc1.12.2-0.2.0.jarをmods/に置く。- 起動し、ワールドに入り、GUI/メニュー/クリエイティブ/ツールチップを確認。
- 問題報告には
logs/latest.log、-Dceltium.vulkan.validation=trueの有無、GPU、シーン、JAR の sha256 を添付。
統計
- commits: 393(fix 177 / perf 125 / feat 30 / docs 32 / chore 10 / measure 6 / refactor 5 / test 3 / revert 1 / profile 1 / glow 1)
- 変更ファイル:
celtium-vulkan/forge122/common/scripts/docs - テスト: 422 + 177
- ルート検証: 241 declarations / 87 classes
- 主要証拠: docs/evidence、
docs/plan.md
Celtium 0.1.4-alpha-rc.26 — the frame target is not rebuilt under an open recording
Celtium 0.1.4-alpha-rc.26 — the frame target is not rebuilt under an open recording
This is the build that actually fixes the tester's black screen. rc.24/rc.25 did not, because two wiring faults
kept the guard blind:
frameBufferOpenwas declared to be "set at its begin" but only ever cleared; no assignment set it true, so the
retire-drain guard never fired for the frame's own recording.- Even with the flag live,
ensureFrameTarget()releases the old pair's images and framebuffer in place on a size
change. The real launch changes size once - the window opens at 854x480, then goes fullscreen at the monitor size -
and that change can land while the frame's unified command buffer is still recording.
The flag is now set where the frame's command buffer begins, and a size change that arrives under an open recording is
refused for that recording and retried on the next one (CELTIUM_FRAME_TARGET_DEFERRED names it).
Evidence
On the tester's launch of this build: VUID 0, device loss 0, GPUVM 0, and
CELTIUM_FRAME_TARGET_DEFERRED from=468x1018 to=1920x1080 fired twice - the deferral doing what it was written for.
Previously the same launch produced VUID-vkEndCommandBuffer-commandBuffer-00059 (two images and a framebuffer
destroyed while bound to a recording buffer) and the black window.
Canonical plain gate: PASSED with floors enforced and the presented image judged.
Also carries rc.25's celtium-device-lost.log report and rc.24's retire-safety rules.
Known open issues (fidelity, not validity)
With the black screen gone, four visual defects are visible and reproducible: main-menu background missing, tooltip
afterimage, blocks flickering / structures appearing in the sky (persist with the GUI hidden), and only some creative
inventory tabs rendering wrong. The validation layer reports no violations for them, so they are renderer fidelity
bugs, and they are the next work.
Packaging
JAR sha256 6e7cba3b08a022b68000d3d2e45579e45449515aab2224530eaab0dd9bdcff87.