agent: engine slop pass — DSP, typed errors, regex parsing, tracing, audit fixes

External code review on 2026-05-12 rated the codebase 4/10 across audio DSP, error typing, JS injection, env-var safety, ALSA parsing, and async logging. This commit lands the prognosis-level fixes plus three audit follow-ups.

Audio/DSP:
- StreamingResampler/rubato confirmed in the live capture path
- regression test at 12 kHz (rms < 0.01, ~40 dB) catches naive decimation
- near-Nyquist test at 9 kHz (rms < 0.05, ~26 dB) exercises transition band

Core errors:
- Other(String) removed; ProviderNotRegistered introduced
- Io variant restructured as struct with kind/message/raw_os_error
- FileNotFound display quotes paths
- Configuration variant removed (unused)

Core types:
- ModelId, EngineName backed by Cow<'static, str>; const borrowed ctor
- Megabytes::from_gb takes u64 (was f64)
- AudioSamples::sample_rate is NonZeroU32; zero-rate defensive branch removed

Capture:
- /proc/asound/cards parsing rewritten as anchored regex (OnceLock)
- regression test covers product names with embedded colons
- monitor_pattern_detection test restored alongside the regex test
- DEAD_SILENCE_FLOOR promoted to module-level with rationale
- DEVICE_VALIDATION_MS, SILENCE_RMS_FLOOR documented with field-observation rationale
- RMS validation loop made idiomatic
- eprintln! migrated to tracing with structured fields and targets

Tauri startup:
- unsafe std::env::set_var removed; ensure_x11_on_wayland renamed to warn_if_x11_env_unset_on_wayland (launcher/wrapper owns env-var contract)
- DB init + log prune + preferences load collapsed to one block_on
- build_preferences_script rewrites JS injection from JSON.parse string to direct object literal plus malformed-JSON guard and unit tests
- WebKitGTK microphone auto-grant logs warning at startup
- tracing subscriber initialised at top of run() (warn,magnotia=info,... on stderr; honors RUST_LOG); previously eprintln→tracing migration was silent because no subscriber existed

Filename counter:
- RECORDING_COUNTER uses SeqCst

Tests: cargo test --workspace --lib green (322 passed, 0 failed across 10 crates).

Three independent audits (original cleanup → Wren → fresh Codex subagent) concur on no critical findings.

Deferred to docs/superpowers/plans/2026-05-12-engine-slop-residuals.md: storage-layer typed errors, remaining eprintln→tracing sweep, capture actor-model refactor, property-based DSP testing, frontend/backend error boundary cleanup.

src-tauri/src/commands/models.rs

@@ -25,14 +25,14 @@ fn whisper_model_id(size: &str) -> ModelId {
         "distil-large" | "distil-large-v3" | "distillarge" => {
             ModelId::new("whisper-distil-large-v3")
         }
-        other => ModelId::new(other),
+        other => ModelId::new(other.to_string()),
     }
 }
 
 fn parakeet_model_id(name: &str) -> ModelId {
     match name {
         "ctc-int8" => ModelId::new("parakeet-ctc-0.6b-int8"),
-        other => ModelId::new(other),
+        other => ModelId::new(other.to_string()),
     }
 }
 
@@ -121,7 +121,7 @@ pub async fn ensure_model_loaded(
     model_id: &str,
     concurrent: Option<bool>,
 ) -> Result<(), String> {
-    let model_id = ModelId::new(model_id);
+    let model_id = ModelId::new(model_id.to_string());
     let entry = model_registry::find_model(&model_id)
         .ok_or_else(|| format!("Unknown model: {model_id}"))?;