2026-10-13 13:26:16 +02:00
# macOS runtime preparation
Normal macOS `auto` builds select Metal with CPU fallback. The worker links
llama/ggml statically and uses Apple system libraries/frameworks. Searchi now
forces `GGML_METAL_EMBED_LIBRARY=ON` so upstream Metal kernel sources reside
inside the worker rather than requiring loose kernel files from the build tree.
After compilation, the builder inspects the artifact using Apple `lipo` and
`otool` ; it does not execute the worker or initialize a GPU. Inspection requires:
- Supported arm64/x86_64 architectures, including both for a universal artifact.
- Only explicitly allowed Apple system libraries/frameworks, including Metal.
- A nonempty `__DATA,__ggml_metallib` section for each architecture.
- Current build-input provenance and a SHA-256 of the resulting worker.
The generated inventory declares the `static-metal-embedded` layout with an
empty packaged-library list. Normal macOS staging now requires this inventory;
the previous optional-inventory exception is removed for `auto` /`metal` .
Staging verifies source identity, worker bytes, Mach-O magic and the static
layout instead of requiring Windows DLLs or Linux shared libraries. Explicit
internal CPU tooling keeps its optional-inventory behavior.
On a native Mac with the required Apple build tools:
```sh
python3 tools/embeddinggemma2/build_runtime.py --backend auto --jobs 2
python3 tools/embeddinggemma2/stage_runtime.py --backend auto
```
2026-10-23 09:43:42 +02:00
The initial preparation was fixture-tested on 13 October; native Apple Silicon
execution was subsequently verified on 23 October as recorded below. This does
not create a macOS release download. Follow the [README installation guide ](../README.md#macos-build-and-install-from-source )
for the tested source-build and local installation steps.
2026-10-13 13:26:16 +02:00
## Evidence, 2026-10-13
Four controlled Mach-O metadata-fixture tests passed, covering static system
dependencies, foreign library rejection, missing Metal data and per-architecture
section coverage. Staging tests exercise the empty-library macOS layout and the
mandatory normal-profile policy. These fixtures are not executable Mac builds.
Across macOS metadata, staging, builder, provenance and Windows/Linux packaging
suites: 75 tests passed with one platform-specific skip, in normal and optimized
Python modes. Actual Windows and Ubuntu WSL normal runtime builds passed after
the recipe change. Windows staging and Linux inventory verification passed.
Worker hashes remain identical to their previously tested CPU deployments:
Windows `c01cb3ec13cfca722ee768e8350e1399a8b9a4d14a24a35adc3473a5cfc82a73` ;
Linux `b52d50a1e676bd4bf64124daff1c4ce83d86a8bde01e4eaf09838cbb24a4c28f` .
No GPU inference or native macOS execution was performed.
2026-10-23 09:43:42 +02:00
## Native Apple Silicon installation, 2026-10-23
Source `9a93ee6375d610d580ef16c5ec52218727fab187` , Apple M4/16 GiB,
macOS26.6.2, stable Rust, CMake4.4.3 and Node24.19.0. Normal `auto` worker
build/staging passed: arm64, embedded Metal library, Apple system-framework
linkage and artifact provenance verified. Both bundled Google assets passed
size/SHA-256 checks. The genuine query-role worker reported `device: gpu` ,
512-token context and normalized 768-dimensional English/German embeddings.
The release Tauri app was installed outside the checkout with both helpers and
the model files. Local ad-hoc signing and deep/strict signature verification
passed. Installed UI checks covered model readiness, English/German document
retrieval, code-only retrieval, source previews, explanations and filters.
Seven actual native screenshots were captured. All66frontend tests passed
under Node24; global Node26 failed one Unicode-highlighting test.
Homebrew FFmpeg supplied local media helpers with external Homebrew library
dependencies. Speech recognition was unavailable during this initial check; the full installation below supersedes that limitation. Intel/universal execution,
GPU failure recovery, shortcut/tray interaction, clean-machine distribution,
Developer ID signing and notarization remain unchecked. These results do not
replace those gates or establish broad media accuracy.
## Full media and speech installation, 2026-10-23
The installed native app now includes the auto Metal/CPU whisper.cpp runtime at
`d1be6fde11ac6e0407606b4e42fe72d34add8037` , with verified multilingual base
and Silero6.2 model assets. All speech shared libraries use loader-relative
paths. `tools/speech-runtime/macos_smoke.py` checked installed linkage and real
recognition of a public-domain recording; the observed backend was `MTL0` .
An owned13-file corpus produced38indexed content segments, including8recognized
speech segments, with no reported media/speech failures. Native UI checks found
and played a waterfall video at4seconds, a spoken sentence in audio and video
near2seconds, and a tagged piano track by artist. A semantic piano-description
query ranked the track first among related suggestions. These checks do not
establish general lyrics recognition or large-library retrieval quality.
CLI and stdio MCP use the same normal query worker. A three-file fixture passed
semantic search, code filtering, source excerpt reading and invalid-handle
rejection. The worker reported GPU,768dimensions and512query-context tokens.
Ten distinct queries through an already-running MCP process had median8.28ms
request wall time; ten repeated cached queries median0.29ms. Three fresh CLI
processes including worker startup had median1018ms. These are small-index local
measurements, not desktop UI timings or production-scale latency guarantees.
Commit validation: all57native desktop tests pass, including real filesystem
create/update/rename/delete events using canonical Mac paths;66frontend tests,
format/build checks,3speech builder tests and installed Metal recognition pass.
The broader workspace run exposes an existing stale permission-denial rollback
expectation in `targeted_reconciliation` (current engine skips unreadable inputs).
Strict current-toolchain Clippy also flags unchanged permission-literal and
ANN-cache boolean expressions. These failures are not qualified as passing gates.
2026-10-23 09:47:08 +02:00
### Native document extraction policy
Mac parent/worker DOCX roundtrip and malformed-PDF rejection tests now pass.
Darwin rejects finite `RLIMIT_AS` and `RLIMIT_DATA` on the tested system. The
worker retains the30-second CPU limit; the parent checks resident memory through
`proc_pidinfo` during its existing5ms polling loop and rejects usage above512MiB,
killing/reaping its owned process group. Input/output bounds and the parent
wall deadline remain active. This sampled RSS guard may overshoot between polls;
it is not equivalent to a kernel-enforced hard address-space limit. Linux and
Windows retain their existing resource policies. An owned live-child test
verifies budget rejection; five process-owner tests and two actual parent/worker
roundtrip/rejection tests pass on the Mac.
