build(cmake): unbreak Linux configure and add a working Linux build workflow

[svenbledt]

Changes Proposed:

This PR proposes changes to:

• Core (units, players, creatures, game systems).
• Scripts (bosses, spell scripts, creature scripts).
• Database (SAI, creatures, etc).
• Build system / CI (none of the above categories apply — no gameplay behaviour changes).

Two related build fixes. The second one is the reason the first is visible.

1. dep/boost/CMakeLists.txt — guard CMP0167

cmake_policy(SET CMP0167 OLD) is currently called unconditionally. That policy was introduced in CMake 3.30, and setting an unknown policy is a hard error on anything older:

CMake Error at dep/boost/CMakeLists.txt:33 (cmake_policy):
  Policy "CMP0167" is not known to this version of CMake.
-- Configuring incomplete, errors occurred!

Ubuntu 24.04 ships CMake 3.28.3, so configure fails outright there. docker/Dockerfile installs the same apt cmake on ubuntu:24.04, so the documented Docker build is affected by the same mechanism — I have not run docker compose build to confirm, but the version is identical.

Wrapping the call in if(POLICY CMP0167) preserves the original behaviour on CMake 3.30+, where the policy exists and OLD is what selects the legacy FindBoost module. Below 3.30 it becomes a no-op, and Boost 1.83 resolves through BoostConfig.cmake regardless — the configure log shows all six components found by config mode.

2. .github/workflows/linux-build.yml — a Linux build workflow that actually completes

Beyond the policy error, the workflow could not have configured anyway:

Problem	Cause
Missing libreadline-dev	dep/readline/CMakeLists.txt raises an explicit FATAL_ERROR without it
TOOLS defaults to ON	Pulls in dep/bzip2 + dep/CascLib; libbz2-dev was not installed, and the extractors need client data no runner has
zlib1g-dev implicit	dep/zlib is find_package(ZLIB REQUIRED); it resolved only by accident off the hosted runner image

Other changes:

• Flags now mirror docker/Dockerfile, the known-good Linux recipe: TOOLS=0, SCRIPTS=static, RelWithDebInfo instead of a separate Release configuration, so CI validates what ships.
• Six Boost components instead of libboost-all-dev (~1GB of packages the core never links).
• ccache added, with the same CCACHE_SLOPPINESS set docker/Dockerfile uses, since both PCH options are on by default.
• The ctest step is removed. There is no test framework in this repository, so it only ever reported that no tests were found, hidden behind continue-on-error. Its --build-config flag was also inert under single-config Ninja.
• submodules: recursive removed — dependencies are vendored under dep/ and there is no .gitmodules.
• Triggers extended beyond main and concurrency added so superseded runs are cancelled.

AI-assisted Pull Requests

• AI tools (e.g. ChatGPT, Claude, or similar) were used entirely or partially in preparing this pull request.

Claude Code, model Claude Opus 5. Used for diagnosis, the workflow rewrite, and running the verification below. Both changes are small and I can justify each line; the CMP0167 fix is the standard guarded idiom from the CMake docs.

Issues Addressed:

• Closes

None — found while trying to get the Linux workflow to run.

SOURCE:

The changes have been validated through:

• Live research (checked on live servers, e.g Classic WotLK, Retail, etc.)
• Sniffs (remember to share them with the open source community!)
• Video evidence, knowledge databases or other public sources (e.g forums, Wowhead, etc.)
• The changes promoted by this pull request come partially or entirely from another project (cherry-pick).

None of these apply. There is no gameplay behaviour here, so there is nothing to source against retail. The authority for the CMP0167 change is the CMake policy documentation (introduced in 3.30) and the configure error above.

Tests Performed:

This PR has been:

• Tested in-game by the author.
• Tested in-game by other community members/someone else other than the author/has been live on production servers.
• This pull request requires further testing and may have edge cases to be tested.

What I actually ran: the full workflow locally under nektos/act on catthehacker/ubuntu:act-latest, end to end:

✅  Success - Main Install Dependencies   [11.77s]
✅  Success - Main Configure CMake        [2.06s]
✅  Success - Main Build HavenCore        [8m58.55s]
🏁  Job succeeded

1976 objects compiled on GCC 13.3 / Ubuntu 24.04 / CMake 3.28.3, no errors. This is the first evidence I am aware of that the experimental Linux port builds clean end to end.

What I did not run, stated plainly:

• No in-game testing. Neither change can affect runtime behaviour — one is a CMake policy guard, the other is a CI file — so there is nothing in-game to exercise.
• No Windows/MSVC rebuild. On CMake 3.30+ the guarded path is byte-for-byte the previous behaviour, so this should be a no-op there, but I have not reconfirmed it. Worth a maintainer with a VS2022 tree doing a configure pass.
• No docker compose build run, so the claim that Docker is affected by the same mechanism is inference from the CMake version, not an observation.
• Not yet run on GitHub's hosted runners — only under act. Hosted runners can carry a newer CMake than apt's on PATH; the guard makes the workflow correct in either direction, but the first real run is the proof.

How to Test the Changes:

• This pull request requires further testing. Provide steps to test your changes.

1. On a machine with CMake < 3.30 (e.g. Ubuntu 24.04, which has 3.28.3), check out main and run cmake -S . -B build -G Ninja -DTOOLS=0 — configure fails with Policy "CMP0167" is not known to this version of CMake.
2. Check out this branch and run the same command — configure completes.
3. cmake --build build --parallel builds the core to completion.
4. Optionally run the whole workflow locally: act -j build -P ubuntu-latest=catthehacker/ubuntu:act-latest.
5. On Windows/VS2022 (CMake 3.30+), confirm configure is unchanged from before.

Known Issues and TODO List:

• max-size: 5G for the ccache is a starting estimate. One clean RelWithDebInfo build writes ~2G of objects, so the original 2G filled to 99.97% and would evict itself before the next run could hit. 5G leaves headroom, but GitHub caps total cache at 10GB per repository and evicts LRU across it, so maintainers may want to tune this once several branches are running CI.
• 359 compiler warnings on GCC 13, none fatal and all pre-existing. Mostly -Wdeprecated-declarations (302), but also 24 -Wenum-compare and 3 -Wreturn-type. The latter two are worth a separate look — for example src/server/scripts/Zandalar/Uldir/boss_zulreborn.cpp:491 compares EncounterData against EncounterState. Out of scope here; happy to open an issue.
• Consider whether the Windows build deserves a matching workflow.