Lumotia/docs/superpowers/plans/2026-05-12-engine-slop-residuals.md

---
name: Engine slop residuals (post-prognosis pass)
type: plan
tags: [engine, slop-cleanup, plan, deferred]
description: Roadmap for the deferred code-quality items that did not fit in the 2026-05-12 prognosis-fix pass. Each area is scoped for its own focused plan when ready to execute.
---

# Engine slop residuals — post-prognosis pass

> **For agentic workers:** This is a meta-plan. It catalogues independent residual work areas, each of which should get its own detailed TDD plan when scheduled. Do **not** treat the sections below as a single execution checklist.

## Context

On 2026-05-12 an external code review rated the codebase 4/10 and laid out a four-bucket de-sloppifying plan. That pass shipped (uncommitted in working tree at time of writing). It hit the *prognosis-level* items only:

- DSP: naive decimation removed, `StreamingResampler`/rubato in place, regression tests at 12 kHz (40 dB) + 9 kHz (26 dB).
- Core errors: `Other(String)` removed; `Io` is structured; `FileNotFound` quotes paths; `ProviderNotRegistered` introduced.
- Core types: `Cow<'static, str>` for `ModelId`/`EngineName`, `NonZeroU32` for `AudioSamples::sample_rate`, `Megabytes::from_gb` takes `u64`.
- Capture cleanup: regex-based `/proc/asound/cards` parser (test covers product names with embedded colons), constants documented, RMS sum idiomatic, Codex-review comments removed.
- Tauri startup: `unsafe std::env::set_var` removed (logs warning instead); function renamed `warn_if_x11_env_unset_on_wayland`; setup collapsed to single `block_on`; JS injection rewritten (`var p = {...}` direct, no `JSON.parse`); WebKitGTK mic auto-grant logs warning at startup.
- Counter: `RECORDING_COUNTER` uses `SeqCst`.
- Tracing: subscriber initialised at top of `run()` with sensible default filter, honors `RUST_LOG`, writes to stderr.

The code-proposal and macro-architecture captures from the same review session went further (10/10 demands and workspace-level critique). Those items are the residuals below.

## Residual areas

Each area is independent. They can be tackled in any order, but the suggested order minimises rework.

### A. Storage-layer typed errors

**Status:** identified, not started.

**Scope:** ~25 `StorageError(String)` sites in `crates/storage/src/{database.rs, migrations.rs}`.

**Why it matters:** The frontend cannot distinguish "schema migration failed" from "FTS search failed" from "transcript not found" — all surface as one stringly-typed variant. This is the same class of bug that motivated removing `Other(String)`.

**Approach:** Split `StorageError` into a `StorageError` enum (sub-error in its own right) with variants like `Connect { path, source }`, `MigrationFailed { version, source }`, `TransactionFailed { phase, source }`, `QueryFailed { kind, source }`, `NotFound { table, id }`. Wrap `sqlx::Error` (or whatever `rusqlite::Error` is in use) as `#[source]`.

**Decisions needed:**
- Is `sqlx::Error`/`rusqlite::Error` serializable across the Tauri boundary? If not, what is preserved (kind + display)?
- Do we keep one flat error type for the crate or nest by domain (`MigrationError`, `TranscriptError`, `TaskError`)?
- Backwards-compat for any existing error-handling code in frontend.

**Files in scope:**
- `crates/core/src/error.rs` — add `StorageError` variant if changing top-level shape
- `crates/storage/src/error.rs` — new file (likely)
- `crates/storage/src/database.rs` — 19 call sites
- `crates/storage/src/migrations.rs` — 6 call sites

**Acceptance:** All `StorageError(format!(...))` calls replaced with typed variants. Frontend can `match` on storage-error kind. `cargo test --workspace` green.

### B. eprintln → tracing sweep (rest of repo)

**Status:** identified, not started.

**Scope:** ~30 `eprintln!` sites in production code paths (excluding tests and tooling binaries — see below).

**Why it matters:** Now that the subscriber is wired, every `eprintln!` is a missed observability signal. The startup/capture paths are clean; the runtime hot paths are not.

**Approach:** Mechanical migration, one file at a time, with appropriate target and structured fields.

**Files in scope (production):**
- `src-tauri/src/commands/live.rs` — 8 sites (highest priority — session lifecycle + inference errors)
- `src-tauri/src/commands/models.rs` — 5 sites (Whisper warmup)
- `src-tauri/src/commands/power.rs` — 4 sites (macOS App Nap)
- `src-tauri/src/commands/transcription.rs` — 1 site
- `src-tauri/src/commands/diagnostics.rs` — 1 site
- `src-tauri/src/commands/tasks.rs` — 1 site
- `crates/hotkey/src/linux.rs` — 2 sites

**Files explicitly out of scope:**
- `crates/mcp/src/main.rs` — separate binary, its own stdio protocol; keep eprintln as-is unless that binary gets its own subscriber.
- `crates/*/tests/*.rs` — test diagnostic output, `--nocapture` semantics depend on stdio.

**Decisions needed:**
- New tracing target names for `commands::live`, `commands::models`, etc. — extend the default filter in `src-tauri/src/lib.rs::init_tracing`.

**Acceptance:** No `eprintln!` in `src-tauri/src/commands/` or `crates/hotkey/src/linux.rs`. Default filter shows live-session events at INFO. `cargo test --workspace` green.

### C. Actor-model refactor for capture + inference

**Status:** identified, not started. **Largest scope.**

**Scope:** Replace `Arc<Mutex<Vec<f32>>>` shared-state pattern in `CaptureWorker` (and analogous patterns in live inference) with isolated actor loops + `mpsc` channels.

**Why it matters:** Reviewer's core architectural critique. Shared mutable state across async boundaries with `std::sync::Mutex` risks executor stalls under contention. Architecturally also makes it impossible to add features like multi-source capture, parallel inference, or recording-while-streaming without lock contention.

**Approach:** Spawn dedicated tokio tasks for the capture worker and the inference worker. Communicate exclusively via bounded `mpsc` channels. Tauri command handlers hold only channel senders.

**Decisions needed (before plan):**
- Bounded vs unbounded mpsc? Recommend bounded for backpressure visibility.
- Channel capacity? (current ad-hoc cap is `AUDIO_CHANNEL_CAPACITY = 32`)
- Backpressure policy when capture outpaces inference: drop oldest, drop newest, await consumer?
- Worker shutdown protocol: drop-channel-and-await, explicit stop message, or cancellation token?
- How are runtime errors (already typed via `CaptureRuntimeError`) propagated through the actor boundary?
- Migration plan: parallel implementation behind a feature flag, or in-place replacement?

**Risk:** This is a substantial refactor. Likely worth a brainstorming session before plan-writing.

**Files in scope:**
- `crates/audio/src/capture.rs` — worker lifecycle rewrite
- `src-tauri/src/commands/audio.rs` — command handler refactor
- `src-tauri/src/commands/live.rs` — inference worker integration
- Likely new files: `crates/audio/src/capture_actor.rs`, `crates/transcription/src/inference_actor.rs`

**Acceptance:** No `Arc<Mutex<T>>` across `.await` points in capture or inference code paths. Capture and inference are independent tokio tasks. End-to-end recording flow works in dev mode. `cargo test --workspace` green. Manual dogfood pass.

### D. Property-based DSP testing

**Status:** identified, not started.

**Scope:** Add `proptest` (or similar) generators for audio inputs and assert invariants of the resampling and VAD pipelines.

**Why it matters:** The current DSP tests are point checks at 12 kHz and 9 kHz. The reviewer's 10/10 demand is property-based testing that covers the input space (any frequency, any chunk size, any noise floor) and asserts invariants (output is bounded, energy is bounded, duration is approximately preserved).

**Approach:** Add `proptest` dev-dependency. Write generators for `(sample_rate, signal_frequency, amplitude, chunk_size, total_secs)` tuples. Assert:
- Output samples are finite (no NaN/Inf).
- Output peak ≤ input peak + small tolerance (no amplification).
- Output duration within ±5% of expected.
- Output energy at frequencies above Nyquist of target rate is at least 30 dB below input energy at those frequencies.

**Decisions needed:**
- `proptest` vs `quickcheck`? Recommend `proptest` (more flexible shrinking).
- Run as part of standard `cargo test` or as a separate `cargo test --features slow-tests`?

**Files in scope:**
- `crates/audio/Cargo.toml` — add `proptest` to `[dev-dependencies]`
- `crates/audio/src/streaming_resample.rs` — extend `#[cfg(test)]` module
- Likely also `crates/transcription/src/streaming/rms_vad.rs` — VAD has equally testable invariants

**Acceptance:** `cargo test -p magnotia-audio` includes property tests. Property tests have run with at least 1000 cases per invariant.

### E. Frontend/backend error boundary cleanup

**Status:** partially complete (orchestrator's `ProviderNotRegistered`).

**Scope:** Audit every Tauri command's return type and ensure errors propagate with full type information across the IPC boundary. Currently many commands return `Result<T, String>` instead of `Result<T, MagnotiaError>`.

**Why it matters:** Reviewer's frontend-backend boundary critique. The frontend can't `switch` on error kind if the backend serialises errors as strings.

**Approach:** Inventory all `#[tauri::command]` functions, identify those returning `Result<T, String>`, convert to `Result<T, MagnotiaError>` (or a command-specific error enum that serializes cleanly). Update frontend handlers to use the typed shape.

**Decisions needed:**
- One `MagnotiaError` for all commands, or per-command enums?
- How do we surface the new shape to the Svelte side without manually maintaining TypeScript types? (specta? hand-rolled?)

**Files in scope:**
- All files in `src-tauri/src/commands/`
- All call sites in `src/lib/` (Svelte)

**Acceptance:** No `Result<T, String>` in `src-tauri/src/commands/`. Frontend has typed error access. `cargo test --workspace` green; frontend builds.

## Suggested sequencing

1. **B (eprintln sweep)** first — mechanical, low-risk, immediate observability win.
2. **A (storage typed errors)** next — well-scoped, no architectural change, unblocks E.
3. **D (property-based DSP)** in parallel with A — independent surface, can be done by anyone any time.
4. **E (FE/BE error boundary)** after A — depends on storage error shape.
5. **C (actor model)** last — biggest surface, deserves its own brainstorm + plan + likely its own branch.

## Cross-cutting deferrals (Phase 10a-blocking)

Two things are explicitly **not** in these residuals because they belong to higher-level workstreams:

- **Lumotia rebrand cascade** (Wyrdnote → Lumotia in code, paths, GitHub repo, bundle ID). Hold until A–B done so rename diff doesn't hide regressions.
- **Phase 10a QC dogfood** — `docs/superpowers/audits/2026-05-10-phase10a-dogfood-notes.md` is queued. Best done after B at minimum so live-session logs are visible.

## What this plan does not do

- No fix recipes for bugs not yet found. Each area's plan should start by reading the current state and writing failing tests against the current behaviour.
- No timeline commitments. Sequencing is suggested, not mandatory.
- No architectural decisions deferred to "Phase B" earlier are pre-committed here. Each area writes its own brainstorm + plan when scheduled.