[ryangswe]

Two behavior changes for 0.8.0, which hasn't been uploaded yet.

Load at play (7bc7256). The content script makes the compute frame, which loads the model, when the page's main video starts playing with card hovers on. If the video is already playing when hovers turn on, it loads at once. The first hover then usually meets a loaded model.

• The tooltip frame still waits for the pointer, and nothing is analyzed before a hover.
• A video that never plays loads on the first pointer move over its picture, as before.
• Hovers off, pages with no video and installs load nothing.

Idle unload (33a6113). After 10 minutes with no move over the picture and no card showing, both frames are removed, which frees the model's memory.

• The next pointer move anywhere in the page while the video plays reloads the model, and so does the video playing again after a pause.
• A playing event with no pause first (a buffering stall, a loop, an ad) does not reload.
• A dev build can shorten the idle time with --idle-unload-ms; other builds refuse the flag.

The privacy page and the docs now say the release download happens when a video plays. The settings refresh and the card details file still wait for a hover.

Measured on an Apple-silicon Mac (native Chrome for Testing 152, dev builds):

• With no hover, a playing video costs about 320 MB of memory, and 1.5–2.2 CPU-seconds a minute while idle.
• A hover after the load shows the card in 0.2–0.4 s, against 2.5 s when the load runs inside the hover.
• After the unload, memory returns to within about 30 MB of having nothing loaded.
• After an unload, moving the pointer across the page to a card shows it about 1.2 s after the first move.

Checks: 595 tests pass (type check and vitest), and lint is clean.