feat(security): wire ModelHost lifecycle + pf-analyze IPC bridge

[graydawnc]

PR 2/7 — PF engine wiring: ModelHost lifecycle + pf-analyze IPC bridge + scan worker participation

PR 1 set up the scaffold; this one wires it to the scan engine. After this PR, flipping pfEnabled actually spawns the inference window, the scan worker consults a pfProvider, and the profile string drifts to regex@N,pf@1.5b-q4 so the backfill loop will re-scan stale sessions. Real ONNX inference still stubs out to [] — that lands in PR 3.

IPC flow

sequenceDiagram
  participant W as scan-worker (worker_thread)
  participant M as main process
  participant H as Inference window
  participant S as Settings UI

  S->>M: setPrefs({pfEnabled: true})
  M->>M: onPfEnabledChanged(true) → syncPfRuntime
  M->>H: spawn BrowserWindow
  H-->>M: pf:ready (runtime, adapterLabel)
  M->>W: notifyPfOnline()
  M->>W: backfill()  ← profile drift triggers rescan
  W-->>W: pfProvider.available() returns true
  W->>M: pf-analyze-req {reqId, text}
  M->>H: PF_ANALYZE_REQUEST {reqId, text}
  H-->>M: PF_ANALYZE_RESULT {reqId, matches[]}
  M-->>W: pf-analyze-res {reqId, matches[]}
  W-->>W: mapPfMatches → mergeMatches → DB insert

Loading

The round-trip is correlation-ID based: each pf-analyze-req carries a unique reqId from the worker; main routes the response to the same reqId. ModelHost holds a Map<reqId, {resolve,reject,timer}> and ignores results from the wrong sender.

What's in this PR

inference/types.ts — wire-protocol surfaces

• PfAnalyzeRequest, PfAnalyzeResponse, PfAnalyzeError
• ANALYZE_REQUEST / ANALYZE_RESULT channel constants

pf-inference.ts (renderer)

• Stub handler: returns [] for now (real ONNX in PR 3)
• Preserves the detectRuntime → pf:ready handshake from PR 1

model-host.ts — analyze pipe

• analyze(text): posts PF_ANALYZE_REQUEST to the renderer, registers a (resolve, reject, timer) slot, awaits the response
• Per-call 30 s timeout (generous for WASM cold start)
• Rejects every pending request on scope close
• 5 new analyze test cases (not-ready, happy path, wrong sender, timeout, inference error)

pf-runtime.ts (NEW)

• Main-side singleton that owns the ModelHost Scope
• start() / stop() idempotent
• analyze() returns [] when inactive so callers don't gate
• pfModelInstalled() helper

main/index.ts — toggle wiring

• syncPfRuntime(pfEnabled): brings the runtime up/down to match the user's preference, then calls scanWorker.backfill() so existing sessions rescan
• Skipped if model isn't installed yet

scan-worker-thread.ts — pfProvider

• available() reads a sync pfOnline flag + the user's pfEnabled pref
• analyze() posts a pf-analyze-req over parentPort and awaits the matching pf-analyze-res
• After the round-trip, runs a cheap regex pass over the same text so class-mapping suppression rules (url/secret boost, DOB gating) have the context they need
• currentProfile() folds pfEnabled + PF_PROFILE_VERSION into the canonical profile string

i18n + e2e + cleanup

• 8 new keys for PfDownloadCard (Download / Cancel / Retry / Ready / WebGPU / WASM / progress / error)
• detector_coming_soon orphan key dropped (PR 1 replaced the inert toggle)
• security-pf-download-card.spec.ts — e2e for the card lifecycle

Test plan

• App suite green: 270/270
• e2e covers download surface
• Locale parity test still green (ja/ko/de/fr get English placeholders, same pattern as backups PR feat(security): Settings → Maintenance with selectable SQLite backups manager #277)

Builds on PR 1 (foundation). Followed by PR 3 (real ONNX inference + runtime info chip).

🤖 Generated with Claude Code

[graydawnc]

• feat(security): hide PF UI entry points; move to opt-in Experimental section #298
• feat(security): restrict PF to high-precision PII categories #297
• feat(security): forward PF inference window console to main + dev devtools #296
• feat(security): in-page Privacy Filter discovery callout #295
• feat(security): real ONNX inference via openai/privacy-filter #294
• feat(security): wire ModelHost lifecycle + pf-analyze IPC bridge #293 👈 (View in Graphite)
• feat(security): hidden inference window skeleton for Privacy Filter ML #292
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.