searchi/docs/LINUX-SPEECH-0.1.6.md

Linux speech client portability check

2026-10-13, Ubuntu 24.04 WSL x86-64. Immutable source snapshot from commit 9bc7221; existing isolated Cargo cache and target, locked offline dependencies. SEARCHI_FORCE_CPU=1 throughout. No model inference or GPU execution occurred.

Executed checks:

cargo test -p searchi-speech --locked --offline
cargo check -p searchi-cli --locked --offline
SEARCHI_TEST_PYTHON=python3 cargo test -p searchi-speech actual_subprocess_timeout_and_overflow_are_reaped --locked --offline -- --ignored --nocapture

Five default unit tests passed. The real owned-subprocess transport test also passed: timeout and oversized framing reaped their child processes; only the typed transport failure was eligible for recovery. The genuine native speech model test remained ignored. CLI compilation passed, including the shared speech/service/assets dependencies, without desktop dependencies.

Recorded command output: speech compilation 11.67 seconds; headless dependency check 7.28 seconds; real transport test 0.14 seconds. These are build/test durations, not inference or search performance measurements.

Snapshot: /home/karl/.cache/searchi-validation/linux-source-9bc7221. This closes a client portability check only. It does not prove the Linux native speech worker, model setup, runtime packaging, recognition quality, desktop UI or GPU operation. Published Linux downloads remain version 0.1.4.

Native runtime preparation

A normal auto runtime diagnostic build from that immutable snapshot was started in /home/karl/.cache/searchi-validation/speech-auto-016. It included Vulkan compilation and no GPU inference. It was deliberately cancelled after inspection confirmed that it predates the relocation correction, rather than due to a wait timeout. The exact owned builder command line was verified before terminating its process tree; its original tool session then reported terminal exit 1. No runtime from that incomplete build is releasable.

Inspection exposed an unverified release path: speech packaging flattened shared libraries without checking ELF dependency closure or their build-tree lookup paths. The builder now checks every staged ELF for origin-only RUNPATH and declared dependencies, allowing a system Vulkan loader only for the dynamic Vulkan module. Three builder tests passed, including rejection of build-tree paths, missing libraries and a Vulkan dependency on the CPU worker.

The CMake configuration now gives Linux shared targets their own $ORIGIN RUNPATH; this is required because executable RUNPATH is not transitive. Actual configuration and a separate native ggml-base build passed in speech-rpath-check-016. readelf verified that the resulting library has RUNPATH $ORIGIN; generated executable link flags use the same path. This narrow library check does not establish complete runtime relocation. The original snapshot does not contain this correction. Its output must not be released as corrected evidence. A corrected normal auto build from immutable commit 61afa08 is now running in /home/karl/.cache/searchi-validation/speech-auto-61afa08. Next: wait for its terminal result, inspect the complete ELF closure, then validate the runtime outside the build tree with LD_LIBRARY_PATH unset and forced CPU recognition.

Linux staging now invokes the same ELF closure verification after manifest, architecture and CPU/acceleration checks. A valid inventory cannot bypass this check; missing readelf also fails staging. Six targeted staging tests passed, including inspected ELF coverage, failure propagation and Windows independence. The same six tests subsequently passed directly on Ubuntu 24.04 (0.409 seconds).

The upcoming real CPU test fixture was prepared from the unchanged verified English WAV (source SHA-256 0aa1fa97d367d987bfb3bb01d7eca11db77a544552ccb8dc0a78ad74d0b496ea). Exact mono 16-kHz signed PCM16 samples were converted to float32 by division by 32768, preserving all 360192 samples. PCM SHA-256: 56948226ac5a0ad4b70523a856dad0851929c7067957f339e9a5b30fc9b369ce. Fixture preparation succeeded; native recognition has not yet been executed.

Headless release preparation

The locked offline Linux release build of searchi-cli and searchi-extract-worker from source 61afa08 passed in 58.38 seconds. The extraction helper was staged with its current source fingerprint and verified. CLI executable: 21921224 bytes, SHA-256 f7602d41eeab80c74e00014ab51297223eae651755707b341951c2e624807d94. This is compilation and provenance evidence, not a new published archive.

tools/package_linux.py --speech-runtime PATH now includes a verified normal speech runtime in its isolated speech-runtime/ directory. The worker is executable; inventory, libraries and approved notices are included. Archive instructions explain speech-model setup and indexing, and metadata identifies speech as experimental. Omitting this parameter preserves reproduction of older speech-free packages. New speech-enabled releases must provide it.

Twenty-one focused package tests passed, including archive hashes/permissions, instructions, and rejection of internal CPU artifacts, tampering, absent notices and invalid ELF closure while preserving an existing output. These use fixture artifacts and do not replace the pending genuine native-runtime verification.

The same 21 packaging tests also passed directly on Ubuntu (1.317 seconds). Package preparation from immutable source 36bd510 verified 21 normal Google runtime records, four pinned decoder/notice records and current extraction-worker provenance. The stripped release CLI is 17560544 bytes, SHA-256 51523b1b20e5afcd2ea81a9c824148262f79ea8360ad411c1ce8a97278290d2f. Its actual setup --help and embed --help accepted the speech options used in archive instructions. No model inference was performed in these preparation checks. At this preparation point, native compilation and the queued relocated test were still pending. Their completed evidence follows below.

Completed normal runtime and genuine CPU recognition

The corrected normal auto build from source 61afa08 subsequently completed successfully. All 25 runtime records passed origin-only ELF dependency closure. The runtime was staged outside its build directory with 29 verified records (including inventory and approved notices); its self-test passed with LD_LIBRARY_PATH and LD_PRELOAD unset.

The actual Rust native integration test also passed in 79.42 seconds. It copied the runtime to supplementary-Unicode paths, ran real recognition on all 360192 English samples twice with forced CPU, verified stable passages/source bounds, and confirmed persistent model reuse. First/warm ASR was 29.879/38.614 seconds; warm model loading was zero and fallback cost zero. These single shared-VM functional timings are not a clean performance benchmark or broad quality score.

Inventory SHA-256: a033ed36d33c43601d219ec17e57c3e561ad6c25eee1556ffbb054f306c5e4a9. Worker SHA-256: b801b733c4212efa0afd28146df5bc6e85e9419314f5c9616355c38ebd4e7110. The 25 inventoried runtime files total 59234912 bytes. Actual test log SHA-256: e651a80e2a980adeb309af28643cf3a33bc184899510fb7054a79327fb5900b7. Ignored raw evidence is in speech-auto-61afa08/relocated-verification.json and relocated-native-cpu.log under the WSL validation cache. Linux GPU execution, shared-service packaged speech/Google retrieval and public archive delivery remain open. Archive construction is now underway.

The first complete archive passed all 60 extracted file/permission records and both installed-layout self-tests, but at 109314948 bytes it exceeds Gitea's 100-MiB upload limit. SHA-256: 9ae7fa9a203c6471cfd6bdd82ac5bf81683813c15fe652595976344fbbf96084. It is retained as an unpublished candidate. XZ packaging now uses a 64-MiB dictionary (preset 9) to compress duplicate decoder/runtime code across large ELF files; client decompression needs roughly 65 MiB. All 21 package tests passed after this compression change. No runtime files are removed.

Genuine packaged shared-service retrieval

The first archive's complete extracted layout then passed a genuine CPU-only CLI integration: index the unchanged English recording, encode raw media and recognized passages using the bundled workers, and search through Google EmbeddingGemma 2. The known-positive request returned semantic_speech with a timestamp_ms citation to the indexed recording and a nonempty snippet. No build-tree loader paths or helper overrides were used; ordinary bundled decoder discovery was exercised. This one-source derived request proves the packaged integration, not independent relevance or precise alignment quality. Raw report SHA-256: f1104fe526d3908066e5030e5ff989fc57e0aae6dea171a7b1ed20e639c38d08. Evidence: headless-speech-functional-016/report.json in the WSL validation cache. Final recompression is underway and must preserve every packaged byte and mode before this evidence can be applied to the upload candidate.