In Manifest V3 every message handler is a cold start, and the fix for lost state hides new failures

The worker being killed is the cheap half of the problem. Once the token lives in `chrome.storage.session` or `chrome.storage.local` ^[5], every handler that needs it opens with an asynchronous read, and an asynchronous read means the reply to the sender is asynchronous too. That is precisely the case covered by the guide's second rule: return `true` from the `onMessage` listener, or the channel shuts and the caller is left with nothing and no error ^[7]. The recommended cure for lost state therefore lands you in the other failure mode ^[16]. A null token at least produces a stack trace somewhere. A closed message channel produces a promise that never settles, which no console error sweep will catch.

The structural consequence is worth stating plainly: no handler may assume it is the second one to run. Chrome, not your code, decides when the worker goes away, and the guide puts that window at roughly thirty seconds of idle time ^[3]. That number arrives without any citation to Chrome's own documentation in the material supplied ^[17], and it is the sort of figure that gets repeated until it is treated as a contract. The pattern does not actually depend on it. Whether the real threshold is twenty seconds or five minutes, a handler that reads its state from storage on entry survives; a keep-alive hack tuned to a thirty-second folklore number does not.

Two soft spots in the guide itself. It announces three isolated execution contexts as the core mental model ^[1] and then, in the text we have, names none of them ^[15]. And the Shadow DOM section argues for a closed root on the grounds that it stops host page scripts tampering with your UI ^[11], while the sample code it ships calls `attachShadow({ mode: 'open' })` ^[12]. Copy the block and you get the boundary the comment told you to avoid ^[13]. For an overlay injected into a site you do not control, that is the difference between a UI the page can reach into and one it cannot.

The three-billion-installs framing ^[14] is the reason anyone builds this at all, and it has no bearing on any of the above. What the distribution buys you is a small distributed system, running inside somebody else's process lifecycle, with strict CSP and hard context walls as the price of entry ^[8].