macOS 13 Ventura floor cannot be met with upstream llama.cpp prebuilts (target 14.0+)

[charneykaye]

Context

While fixing #469 (Steam macOS depot shipped a llama.cpp build targeting macOS 26), we bisected upstream llama.cpp releases for their declared macOS minimums (via the new scripts/check-macho-minos.py, PR #471):

• b9428 and newer: target macOS 26.0
• b9415 and earlier (back through at least b8500): target macOS 14.0

No upstream prebuilt release satisfies macOS 13. We pinned b9415 (14.0) in PR #471, which fixes the crash for every macOS 14+ player — but docs/QA_STEAM_PLATFORM_MATRIX.md lists macOS 13 Ventura as the oldest supported release, and a Ventura player would still hit the dyld kill (surfaced as "Model loaded but could not start").

The decision

Two ways to close the gap — maintainer's call:

1. Build llama.cpp from source for the Steam macOS depot with MACOSX_DEPLOYMENT_TARGET=13 (and -mmacosx-version-min=13 for the Metal backend), instead of repacking upstream prebuilts. More release engineering, but keeps the documented floor.
2. Amend the QA matrix to macOS 14 Sonoma as the supported floor. Zero engineering; drops Ventura (final release of the last Intel-only support window — worth checking player telemetry/Steam survey share before dropping).

Until decided, the effective floor shipped to players is 14.0 (enforced in CI by the llama-minos-gate job from PR #471, so a future pin bump can't silently raise it further).

References

• Steam macOS depot bundles a llama.cpp Metal backend built for macOS 26 — engine dies on dyld on older macOS (misreported as "not enough RAM") #469 (the packaging bug this gap was discovered under)
• PR Pin llama.cpp to b9415 and gate macOS deployment target #471 (pin + gate; CONVSIM_MACOS_MINOS_FLOOR env var is the knob if the floor changes)

🤖 Generated with Claude Code

https://claude.ai/code/session_01C9WmcidRL3ABq7qhWHUsrG