[issue-217] Precache libdedx WASM + ASR worker after model download; offline SW cache fallback

[grzanka]

What changed

Fixes #217 (verifying/refining #163 §5.4): the "download models" consent
flow never touched wasm/libdedx.mjs or the ASR worker script — both are
dynamic-loaded lazily, only on a query's first use — so a user who went
offline right after finishing that flow hit an uncached fetch and failed,
even though they'd just been told they were ready.

Implements both parts of the issue's suggested fix:

1. Precache the compute/ASR code paths as part of the download flow.
   model-status.svelte.ts's startDownload() now awaits getService()
   (libdedx WASM) and calls a new asrStatus.prewarm() (fire-and-forget,
   reuses the same singleton worker client start() later uses) right
   after downloadModelWeights() succeeds, before flipping to "ready".
   Failures from either are logged only, never fatal — the model weights
   are already safely cached at that point, so a WASM/ASR hiccup shouldn't
   revert the whole flow to "fresh" and force a re-download.

2. Give coi-serviceworker.js's fetch handler a caches.match()
   fallback. Every successful same-origin GET response it proxies is now
   also written into a new aidedx-coi-runtime-v1 Cache Storage bucket. On
   a failed live fetch (offline), it now checks Cache Storage — every
   bucket, so it also benefits from transformers.js's own
   "transformers-cache" entries for ONNX Runtime Web's wasm/mjs — before
   re-throwing. Kept same-origin-only so it never duplicates the
   multi-hundred-MB cross-origin model weight downloads (Cyfronet S3
   mirror) into a second cache.

Test plan

• pnpm run format:check — scoped to the changed files; a repo-wide
  run trips on pre-existing untracked .android-asr-cache/ and
  .venv-asr-bench/ directories unrelated to this change (confirmed
  the same failure exists on main)
• pnpm run lint — scoped to the changed files, same pre-existing
  caveat
• pnpm run check — 0 errors
• pnpm run validate:eval
• pnpm test — 982 passed, including new coverage for
  asrStatus.prewarm() and modelStatus's post-download precache
  (success path + both failure paths staying non-fatal)
• pnpm build
• Live-browser re-verification of the literal §5.4 repro (download
  models → disconnect networking → no prior online query → submit a
  query) — not done as part of this PR; see [issue-163] §5.4: offline claim fails on first use — libdedx WASM + ASR worker never precached #217's own repro steps

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR addresses issue #217 by making the “download models” consent flow actually prepare the app for offline first use: it proactively loads the libdedx WASM compute path and warms the ASR worker, and it enhances the COI service worker to persist and serve cached responses when live fetches fail.

Changes:

• Add a runtime Cache Storage bucket + offline fallback behavior to coi-serviceworker.js for proxied fetches.
• After downloadModelWeights() succeeds, precache libdedx (getService()) and prewarm ASR (asrStatus.prewarm()) before marking the model state as "ready" (non-fatal on failure).
• Add Vitest coverage for the new post-download precache behavior and the ASR prewarm() path.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
static/coi-serviceworker.js	Adds runtime caching for successful GETs and a cache fallback on fetch failure to improve offline resilience.
src/lib/models/model-status.svelte.ts	Preloads compute WASM and warms ASR right after a successful model download to support offline-first-use.
src/lib/models/model-status.test.ts	Tests that post-download precaching occurs and that precache failures remain non-fatal.
src/lib/asr/asr-status.svelte.ts	Adds prewarm() to trigger worker pipeline load without changing ASR phase/recording state.
src/lib/asr/asr-status.test.ts	Tests prewarm() behavior and verifies the warmed worker client is reused by start()/stop().

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.