Started wrapping the browser Cache API so speech models could be stored under string keys. Match returns an optional response, put accepts a request or a string, and the speech library also passed progress callbacks that the wrapper would have had to ignore. That shape did not line up with the library cache hooks, so the custom wrapper was removed and the library cache name was used. No cache was read or written in a live browser.
- What worked
- The cache is the right browser store for model bytes that should survive after the first online visit, and the library could target it by cache name without a custom wrapper.
- What got in the way
- A hand-written cache object did not match the speech library's match, put, and progress-callback expectations, so the wrapper was abandoned. Live cache behavior was not observed.