CI is broken: unpinned core clones have drifted away from their patches

[lepht]

Every CI run fails during build-core, so no base.zip is produced and no device-installable build can be made. Inherited from upstream; almost certainly failing there too. #17 is chipping away at it.

This turned out to be systemic rather than a single bad core.

Root cause

Cores are cloned unpinned and shallow, straight from third-party HEAD:

cd src && git clone --depth 1 $(if $($1_BRANCH),--branch $($1_BRANCH),) $$($(1)_REPO) $(1)

Then a local patch is applied to add the tg5040/tg5050 platform block. When the upstream repo changes that file, the patch stops applying and the build dies — with no commit on our side. Two have already rotted:

Core	Symptom	Broken by	Pinned to
mgba	Makefile.libretro: No such file or directory	8940477 (2026-08-04) deleted it — CI moved to cmake	925f0f0
fceumm	patch failed: Makefile.libretro:461 — context drift	8ed0cb2 (2026-07-23) replaced bundled zlib	becde7dc

Both verified by checking out the pinned commit and running git apply --check for the tg5040 and tg5050 patches, which differ per platform.

Why it kept looking like one core

fail-fast: true on the core matrix meant the first failure cancelled the other ~40 jobs, so each CI round surfaced exactly one broken core and fixing them was serial. #17 turns it off, so one run now reports every broken core at once.

Remaining work

• Pin mgba (fix(cores): pin mgba and fceumm, stop fail-fast hiding broken cores #17)
• Pin fceumm (fix(cores): pin mgba and fceumm, stop fail-fast hiding broken cores #17)
• fail-fast: false on the core matrix (fix(cores): pin mgba and fceumm, stop fail-fast hiding broken cores #17)
• Pin every remaining core. Roughly 22 of them, all still unpinned clones of third-party repos. Any one can break the build tomorrow. For cores whose patches currently apply, pin to today's HEAD — no archaeology needed, just record what already works.
• Port the mgba patch to cmake so it can track upstream again. The pin is a holding action; upstream is cmake-only now.

Automating the bumps

Pinning trades "randomly broken" for "deliberately stale", and stale is only acceptable if bumping is easy. Worth a scheduled workflow that:

1. for each core, resolves upstream HEAD
2. tries applying our patch against it
3. builds the ones that apply
4. opens a PR bumping the pins that built green, listing the ones that didn't

That makes staleness a review queue rather than silent drift, and it catches upstream breakage on its schedule instead of when someone needs a release. Dependabot can't do arbitrary git pins, so this needs to be a small bespoke workflow — but it's mostly the loop above, and the patch-check step is cheap (a blobless clone plus git apply --check).

Cadence matters more for some cores than others. FBNeo is the one to watch: arcade romset compatibility is tightly version-coupled, so a stale pin there is user-visible in a way a stale gambatte is not. pcsx_rearmed sees active ARM performance work. The mature 8/16-bit cores can drift for a year with little cost.

Impact

Blocks #1, #2 and #3 from producing installable artifacts, and blocks any release. Not caused by any of them — the same failure occurs on main.