The re-check inside the callback is the part implementations skip, and it is the part that does the deduplication. Three tabs queue on the name `token-refresh`. The first does the network round trip and writes the result. The second and third take the lock after it, see a token that is no longer expired, and return without a fetch [7]. Take the `isExpired` guard out and the lock still works exactly as specified: you get three sequential refreshes instead of three concurrent ones, queued politely, each rotating the token out from under the last.
The arithmetic is worth stating because it is the whole return on the change. Three tabs queue, one fetch runs, and the other two requests never leave the browser [1]. That number transfers only if all three documents sit on the same origin and read the same token store, because the lock namespace is the origin and the guarantee covers code this browser can observe [2]. Auth split across two origins, or a phone and a desktop tab calling the same refresh endpoint, get no coordination from this at all.
There is a reason the racy version survives review. Reproducing the failure needs at least two documents noticing expiry in the same tick [1], and a single-tab development loop produces zero contention [3]. The bug ships because the test never runs it.
`signal` takes an AbortSignal, so a tab can give up after 200ms rather than queue behind a slow holder [11]. That is the second option I would reach for, after the re-check, because a queued request has no timeout of its own and a tab parked on another monitor since yesterday is a plausible holder [1].
Deadlock stays your problem. A tab holding `a` while waiting on `b`, against a tab doing the reverse, waits forever [16]. Two lock names in one code path is where that starts, so acquire them in a fixed order or collapse them into one name.
Context for the opinion, since this is a tradeoff and not a rule: I would use this for token refresh and for deciding which tab owns the WebSocket [9], and I would not treat it as a queue. It is a handful of lines, and it removes a class of bug that a flag in storage cannot remove, because the atomic compare-and-set that would make that flag correct does not exist in `localStorage` [2].
