freefrank/LostOdysseyRecomp v0.6.3

on GitHub

English

• Revert large vertex-cache hits to bounded sampled content comparison to prioritize draw-time CPU cost. Small vertex buffers remain exact; large buffers compare the head and tail plus strided samples, while index-cache hits retain exact source-byte verification. The focused LoVertexCacheTest passed 3,668,957 checks, including sampled changes, small-buffer exactness and the selected large-vertex blind-spot policy. The synthetic blind spot is known; no known game bug has been caused by sampling, and no game or release-binary performance result is claimed. A reliable low-cost vertex invalidation mechanism remains backlog work.
• Harden Issue #54 language-menu selection: invalid language-table counts and indices no longer read or rewrite a selection, and unavailable entries display as — instead of fabricating a language. The menu capacity follows the native parser limit (16 entries), while the existing USA/Europe 82481BE8 host-language mapping and independent text/voice handling remain unchanged. An opt-in, bounded LO_TRACE_LANGUAGE=1 trace is available for future reports, including the menu apply/close boundary snapshot. The specific cause of the reported cutscene voice issue remains unconfirmed; no save was available and no real-game reproduction was performed.
• Apply a defensive file-I/O locking fix for Issue #53: read, write and scatter paths keep the per-file mutex only while updating seek, transfer, position and size state, then publish completion results through the retained handle reference. The original Disc 2 hang was not reproduced, so its root cause remains unconfirmed.
• Add opt-in, bounded I/O diagnostics for future Issue #53 reports, including file-handle lifecycle, mutex wait/acquisition, transfer, completion publication and API-return stages, plus manual JSONL snapshots.
• Add deterministic guest I/O lifetime, APC/event ordering, independent-file and duplicate-handle regression coverage, and reuse the multi-disc checks on Windows and Linux. These checks do not establish a story transition or player acceptance.
• Use the platform archive format for asynchronous F1 render-state exports: Windows produces .zip, while Linux uses the system tar and gzip tools to produce .tar.gz. Focused WSL Manjaro g++ C++20 -Wall -Wextra -Werror checks pass for byte-identical 8 MiB/log extraction, asynchronous preparation, archive collisions, missing tar, unreadable sources, symlink-root rejection and shutdown joining. AppImage runtime and in-game/user acceptance remain pending.

Published at GitHub Release v0.6.3 on 2026-09-19T23:56:08Z from tag/source commit 93bdbc1ccae7652e38dc80db24a9d25a34a72a47. Release CI 35476569158 passed Windows and Linux packaging on its first attempt. The four public assets are the Windows ZIP and Linux AppImage plus their .sha256 sidecars; both package hashes match their sidecars and GitHub digests, and public sidecars returned HTTP 200. The Windows ZIP is 211,084,312 bytes with SHA-256 9737befe09bc011c17327d5a83df8e437ab7e03a0f107b4acf40621dfd5fee08; the Linux AppImage is 220,813,816 bytes with SHA-256 b4f0448e6e661f51a2cab07209289188c786710929e4c4334f09786f125e7ed7. The Windows manifest reports version/source version 0.6.3, commit 93bdbc1 and dirty=false; all 49 payload hashes and the embedded shader were verified. Linux native GPU, Steam Deck, AppImage runtime and broader gameplay remain unverified.

简体中文

• 为优先降低绘制阶段 CPU 成本，将大顶点缓存命中恢复为有界 sampled-content 比较。小顶点缓冲仍逐字节精确比较；大缓冲比较头尾片段和跨区采样；index cache 命中继续保留完整源字节校验。定向 LoVertexCacheTest 通过 3,668,957 项检查，覆盖采样变化、小缓冲精确性和已选择的大顶点采样盲区策略。已知存在合成盲区；目前没有任何已知游戏 bug 由采样引起，也没有据此宣称实机或发布二进制性能结果。可靠且低成本的顶点失效机制仍列入 backlog。
• 加固 Issue #54 的语言菜单选择：语言表数量或索引无效时不再读取或改写选择，无效条目显示为 —，避免伪造语言；菜单容量遵循原生解析器的 16 项限制。既有 USA/Europe 82481BE8 宿主语言映射以及文字／配音独立处理保持不变。新增默认关闭且有界的 LO_TRACE_LANGUAGE=1 追踪，包含 menu apply/close 边界快照，供后续报告使用。具体过场配音问题的根因尚未确认；没有存档，本次未做实机复现。
• 为 Issue #53 应用防御性文件 I/O 锁范围修正：读取、写入和 scatter 路径只在更新 seek、传输、位置和大小状态时持有文件互斥锁，随后通过保留的句柄引用发布完成结果。原 Disc 2 卡住未能复现，根因仍未确认。
• 增加默认关闭且有界的 I/O 诊断，记录文件句柄生命周期、互斥锁等待／取得、传输、完成发布和 API 返回阶段，并支持手动导出 JSONL 快照。
• 在 Windows 与 Linux 补充真实客户机 I/O 生命周期、APC／event 顺序、独立文件和复制句柄回归覆盖，并复用多盘测试。这些检查不代表已完成剧情换盘或玩家验收。
• 为异步 F1 渲染状态导出使用平台归档格式：Windows 生成 .zip，Linux 使用系统 tar 和 gzip 生成 .tar.gz。WSL Manjaro g++ C++20 -Wall -Wextra -Werror 定向检查已通过，覆盖 8 MiB／日志逐字节一致解包、异步 prepare、归档冲突、缺少 tar、源文件不可读、拒绝符号链接根目录和退出时 join。AppImage 运行时及游戏内／用户验收仍待完成。

v0.6.3 已于 2026-09-19T23:56:08Z 从 tag/source commit 93bdbc1ccae7652e38dc80db24a9d25a34a72a47 发布到 GitHub Release。Release CI 35476569158 首次通过 Windows/Linux 打包。四个公开资产为 Windows ZIP、Linux AppImage 及各自 .sha256 sidecar；两个主体包 hash 与 sidecar 和 GitHub digest 一致，公开 sidecar 返回 HTTP 200。Windows ZIP 大小为 211,084,312 字节，SHA-256 为 9737befe09bc011c17327d5a83df8e437ab7e03a0f107b4acf40621dfd5fee08；Linux AppImage 大小为 220,813,816 字节，SHA-256 为 b4f0448e6e661f51a2cab07209289188c786710929e4c4334f09786f125e7ed7。Windows manifest 报告版本／source version 为 0.6.3、commit 93bdbc1、dirty=false；49 个 payload hash 和内置 shader 均已核验。Linux 原生 GPU、Steam Deck、AppImage 运行时和更广游戏流程仍未验证。