tests: parallel-safe sqlite test paths (#21)

[hunnyboy1217]

Closes #21.

Summary

• Implements all 10 acceptance items of beta: invest in harness reliability (mock voice cycle, CI matrix, sqlite test isolation) #21 (6 in-scope + 4 acceptance criteria).
• Adds voice_loop::process_transcript extracted from voice_cycle, so the integration test can drive the orchestration end-to-end with mocks.
• Adds SttEngine::mock, LlmClient::mock (new MockLlmBackend impl of the existing LlmBackendClient trait), and TtsEngine::silent test doubles.
• Fixes the two named flakes: tools::parser Memory available: (Linux-only via #[cfg(target_os = "linux")]) and
  memory::promotion_redacts_person_memory_in_namespace_note (root cause was shared canonical_dir = path.parent().join("memory") — every memory test now gets its own parent dir).
• Adds .github/workflows/ci.yml with a Linux + macOS matrix (fmt, check, clippy -D warnings, cargo test --lib, cargo test --tests, cargo test --doc) plus a separate cargo-llvm-cov coverage job.

Scope

This PR closes #21 in full. Intentionally out of scope, tracked under #34:

• aarch64-unknown-linux-gnu cross-compile job.
• cargo-audit and cargo-deny jobs.

Tests

• cargo check -p genie-core
• cargo test -p genie-core --lib
• cargo test -p genie-core --tests voice_loop_integration
• cargo clippy --workspace --all-targets -- -D warnings
• cargo fmt --all -- --check

Real Behavior Proof

• I have built and run the affected code locally (or noted why I could not).
• I have verified the change end-to-end on Jetson hardware OR explained the equivalent verification path I used.

What I ran

This branch was prepared without a local Rust toolchain (the system cargo from apt is 1.75; the workspace is on edition 2024 and needs rustc ≥1.85). Verification was therefore driven through CI per CONTRIBUTING.md option 2 + 3:

• Cross-compile (aarch64 / Jetson) job — builds the real aarch64 release binaries (genie-core, genie-ctl, genie-governor, genie-health, genie-api) with the same make jetson recipe that ships to Orin Nano.
• CI / fmt — cargo fmt --all -- --check.
• CI / ubuntu-latest / check + clippy + test — cargo check --workspace --all-targets, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace --lib, cargo test --workspace --tests, cargo test --workspace --doc.
• CI / macos-latest / check + clippy + test — same matrix on macOS so the "fails on macOS" class of bug surfaces on every PR.
• CI / cargo clippy + test (--no-default-features) — chat-only build axis with voice + telegram disabled, scoped to -p genie-core -p genie-ctl.
• CI / coverage (cargo-llvm-cov) — non-blocking baseline coverage report, lcov.info uploaded as a 14-day artifact.

What I observed

• The new voice_loop_integration.rs (10 #[tokio::test] cases) drives process_transcript with mocks and asserts on the three beta: invest in harness reliability (mock voice cycle, CI matrix, sqlite test isolation) #21 AC-B observables: the user transcript text appears in the conversation history, an assistant message is appended after the LLM call, and a get_time event lands in the tool-audit.jsonl audit log.
• The two named flake tests (tools::parser Memory available: and memory::promotion_redacts_person_memory_in_namespace_note) now pass reliably on every CI run — the parser test only runs on Linux (where /proc/meminfo exists), and every memory test gets its own ${tmpdir}/geniepod-mem-${label}-${pid}-${id}-${nanos}/ parent dir so the canonical-dir markdown pipeline cannot collide under parallel execution.
• The Cross-compile (aarch64 / Jetson) job has been green on every push to this branch, including this one — concrete proof the changes still build for the deployment target even though the integration test file is now gated behind #![cfg(feature = "voice")].
• No Jetson hardware run on my end. A maintainer with an Orin should sanity-check the voice loop's process_transcript extraction on real hardware before merge; the extraction is a pure refactor (no behavior change in voice_cycle's audio prelude → STT → process_transcript split), but mics in the loop are the kind of thing that surprises you.

[hunnyboy1217]

Hi, @ai-hpc ,
Could you please review my PR?

[hunnyboy1217]

Hi, @ai-hpc ,
Could you please review this PR?
Thanks.

[ai-hpc]

@hunnyboy1217 Thanks for the contribution. This issue is related to the beta version and is currently blocked until the alpha version is completed first.

[ai-hpc]

The work splits cleanly into "valuable for #21" and "conflicts with main's CI evolution". Asking for the second part to be dropped on a rebase so the first part can land.

Keep — these are real value-add for #21

• crates/genie-core/tests/voice_loop_integration.rs (+511) — the integration test that #21's acceptance criterion B specifically asks for. This is the high-value piece of the PR.
• crates/genie-core/src/llm/mock.rs (+153) and mod.rs (+14) — the test mock that lets the integration test run without a real LLM. Clean and reusable. Right design point.
• crates/genie-core/src/memory/mod.rs (+22/-13) — the parallel-safe sqlite test paths that #21's title specifically calls out. Most relevant to the project's actual stability.
• **crates/genie-core/src/voice_loop.rs (+102/-9), voice/stt.rs (+79), voice/tts.rs (+36)** — supporting changes for the integration test. Need to look at whether these have any interaction with PR #61 (real streaming TTS, just merged) or PR #87 (concurrent server / LocalSet`), but the test-only changes should rebase cleanly.
• tools/parser.rs (+8) — small assist; should rebase.
• CHANGELOG entry (+112) — most of it can stay once split into the right ## Unreleased location.

Drop on rebase — .github/workflows/ci.yml rewrite

#21's CI acceptance criterion (C — "CI runs on every PR" — and IS-5 — "matrix runs cargo check / cargo test --lib / cargo clippy --deny warnings") has already been delivered on main by a different shape:

• PR #37 (8ad9d11) landed the fmt / clippy / test workflow.
• PR #62 (e8b0ef1) added the --no-default-features axis specifically to guard the chat-only build for PR #57's voice Cargo feature. This PR's rewrite removes that axis — that's a real regression on existing functionality.
• PR #89 (8045d35) split out a Contribution / PR body checklist workflow with pull_request_target so the check fires on every PR including stale-base ones.
• PR #91 (6158951) added the AI-attribution rule on PR bodies.

Plus, this PR's check-and-test matrix adds a macos-latest axis. macOS isn't a deployment target — GenieClaw is a Jetson appliance — so the macOS leg adds maintenance burden without reflecting the production stack. The coverage (cargo-llvm-cov) job is interesting as a follow-up, but adding it now alongside the matrix rewrite that conflicts with main is too many things at once.

Concretely:

git fetch origin
git checkout -b feat/issue-21-resubmit origin/main
git checkout perf/21-sqlite-test-paths -- \
    crates/genie-core/tests/voice_loop_integration.rs \
    crates/genie-core/src/llm/mock.rs \
    crates/genie-core/src/llm/mod.rs \
    crates/genie-core/src/memory/mod.rs \
    crates/genie-core/src/tools/parser.rs \
    crates/genie-core/src/voice/stt.rs \
    crates/genie-core/src/voice/tts.rs \
    crates/genie-core/src/voice_loop.rs \
    crates/genie-core/tests/tool_dispatch_test.rs
# Reconcile the voice_loop.rs changes against PR #61 (just merged — real
# streaming TTS) and PR #87's `LocalSet` server. Carry just the parts of
# the CHANGELOG entry that match what's actually changing.
# DO NOT touch .github/workflows/ci.yml — that file is now owned by the
# CI evolution that landed across PRs #37 / #62 / #89 / #91.
cargo fmt --all
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace --lib
# … if green, open as a fresh PR.

Re: the cargo-llvm-cov coverage job

If you still want to land that as a separate piece of work after the integration test is in, it's clearly worth its own PR — and the existing CI is structurally fine to graft a coverage job onto without rewriting the fmt / clippy / test shape.

Soft norm note (carries over from #61 era — not blocking)

Four of five commits on this branch are authored as Hunnyboy1217@gmail.com (personal Gmail) rather than the GitHub-noreply form. Future PRs would be cleaner with git config --global user.email "<numeric-id>+hunnyboy1217@users.noreply.github.com". Same single-vector flag as PR #61 (which I merged anyway); just for next time.

Why this isn't a close

PR #61 was excellent — SentenceStreamer was genuine engineering — so this contributor is clearly capable of delivering quality work. The integration test in this PR is the same quality. The CI rewrite is a casualty of timing (this PR predates the CI evolution on main), not of effort. A clean re-submission against current main should be a quick re-review and merge.

[hunnyboy1217]

Hi, @ai-hpc ,
Could you please review my PR?
Thanks.

[ai-hpc]

Merged at 4eb867a

Thanks @hunnyboy1217