[themuffinator]

Compiling the attached hub2.map from #181 requests 2,344,661,760 bytes for unshared vertices in ShareMapTriVerts. The legacy allocator rejects that request and the writer dereferences the null buffer.

Allocate vertices after sharing, using a spatial hash that preserves the original epsilon matching and first-match/index order. The reported surface needs only 21,559 shared vertices (1,379,776 bytes). Check rejected geometry allocations and count overflow, and retain triangle-list tails during T-junction repair and surface assembly to avoid repeated scans. Add a production-code regression fixture to both validation workflows. Allow 150 minutes for the full Windows x64 PR validation job, matching Push Verification, so its existing regression suite, build and dedicated-server checks can finish.

Validation:

• Windows optimized build and staging pass.
• Native fixture passes with Windows MSVC/Clang and Linux Clang/GCC ASan/UBSan. The original implementation reproduces the null write with the same allocation-limit fixture.
• Existing dmap geometry-lifetime, AAS compiler-contract and Linux ARM64 portability checks pass.
• The supplied map completes geometry (503 seconds), collision and all five AAS outputs, then exits normally.
• Staged OpenGL single-player gameplay on stock game/airdefense1 passes, including an engine screenshot.

Independent validation of PR head a7e7a0a08ea285d1c574457b458b81eff57c93dd:

• The production-code vertex-sharing fixture passes with Windows Clang and MSVC, and with Linux Clang and GCC under ASan/UBSan (Ubuntu on WSL).
• The dmap geometry-lifetime, AAS compiler-contract and Linux ARM64 source-portability checks pass; the diff has no whitespace errors.
• Local Windows LLVM 20.1.7 sanitizer execution fails while handling a C++ exception. The same failure reproduces in a standalone throw/catch program without engine code; that Windows sanitizer result is not used as qualification. No additional engine build or gameplay run was performed during this independent check.

Remaining limitation: the rebuilt hub2 still cannot enter gameplay. Its 36,635,340-index surface exceeds the renderer's existing 16,777,216-index limit (R_ParseModel: bad surface counts); loading also reports MAX_SIL_EDGES. Keep #181 open for that follow-up. This PR qualifies the compiler crash fix and records the remaining limit in the developer documentation and candidate release notes.

The original runtime validation used the staged working-tree build, which includes pre-existing UI work. The PR contains nine files covering this compiler fix and its validation.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T09:41:55.983145Z	8a320db	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.