feat(jarl-atoms): 778 — veto any navigation with navigationGuardAtom

[randomdevpete]

Public API

// jarl-atoms
type NavigationGuardAtom = Atom<string | null>;
const navigationGuardAtom: (guard: (get: Getter) => string | null) => NavigationGuardAtom;
const enforceNavigationGuards: (store: Store, guards: ReadonlyArray<NavigationGuardAtom>) => () => void;

// jarl-react
const useNavigationGuard: (guard: NavigationGuardAtom) => void;

A guard atom returns the confirm message to block a navigation with, or null to allow it.
Several guards compose: the first that returns non-null wins. enforceNavigationGuards is the
effect that makes registered guards bite, for one store — call it once near the root, or per
component via useNavigationGuard, which enforces for as long as the calling component is
mounted.

Two choke points

Every in-app navigation (Link, useNavigate, a route atom or locationAtom write) funnels
through locationAtom's writer, so it's vetoed there, synchronously, before jotai-location calls
history.pushState — no history entry, no rollback, no flicker.

That single choke point isn't enough: it only sees writes that go through jarl's own atoms. A
third-party history.pushState, and the browser's own back/forward buttons, never touch
locationAtom. Those are caught by the second choke point — the
Navigation API's navigate event,
which fires before commit for every same-document navigation regardless of source and whose
preventDefault() genuinely cancels it. locationAtom also now passes a subscribe override to
atomWithLocation that listens to navigation.currententrychange (falling back to popstate
where the Navigation API is absent) — closing a pre-existing gap where a third-party
history.pushState was invisible to jarl entirely.

A module-level reentrancy flag (approvingOwnNavigation) suppresses the second choke point for a
navigation jarl's own locationAtom write already approved, so it isn't asked to confirm twice for
one navigation. It's module-level rather than per-store because window.navigation/history are
one global per page — there's only ever one in-flight "our own" pushState to track, not one per
store.

What each navigation source gets

Source	Outcome
In-app: Link, useNavigate, a route atom or locationAtom write	Vetoed at locationAtom's write. No browser support required.
Same-document, from outside jarl: third-party history.pushState, a fragment change, same-document back/forward	Vetoed through the Navigation API's navigate event. Unguarded in a browser without it.
Leaving the document: reload, a cross-document link, closing the tab	beforeunload. The browser shows its own wording, not the guard's message.
Cross-document back/forward	Never cancelable by platform design (anti-trapping) — un-vetoable.
A back/forward traversal repeated without interacting with the page in between	Consumes the user activation that permits cancelling — un-vetoable.
Browser-initiated navigation: URL bar, a bookmark, the reload button	Fires no navigate event at all; reaches beforeunload and nothing else.

Design notes carried from the ticket

• Navigation API only, no History-API fallback — OWNER's call, "fine until it's not". A
  History-based fallback would need pushState/replaceState patching and a rollback path;
  revisit only if a real consumer reports an unsupported browser.
• window.confirm, synchronously — also OWNER's call. preventDefault() must be called
  synchronously, so an async custom modal would need precommitHandler, which Safari 26.2 doesn't
  yet ship.

Stacking

Stacked on #98 (task-675-coalesce-resolvedatom-and-asyncrouteatom) per depends_on
resolution: merge-tree found a conflict between the two branches in
e2e/fixture-app/src/routes.ts (675 renames resolvedAtom to asyncRouteAtom in that file's
import list; 778 adds navigationGuardAtom alongside), and CLAUDE.md's ticket-order tiebreak puts
the lower id first. Resolved during the restack — the two changes are independent edits to the same
import list, so the resolution is a plain three-way merge with nothing to reconcile logically.

#96 (ticket 676) is held pending this PR. It ships a GuardedLink built on useLink, guarding
only click-throughs on that one component — the userland workaround this ticket replaces with a
real primitive. Once this merges, 676 is reworked onto it: the fixture page swaps to
navigationGuardAtom/useNavigationGuard, and GuardedLink is dropped.

Style-guide exceptions

None in this diff.