fix: close the remaining link-loss path in the picker batching

[jnahian]

Follow-up to jnahian#14 — review fixes, targeted at that PR's branch so they land as part of it. Merging this updates jnahian#14.

Thanks for this; the batching design is sound and the seam tests are genuinely good. Review turned up one path that can still lose links, plus a set of smaller things.

The blocking one

finishOpening presents the links that arrived during a launch, and then the browser it just opened — launched with activates = true — takes key a beat later. windowDidResignKey read that as click-away and discarded a batch the user had never seen: not opened, not copied, not in the fallback. The comment at the pick call site asserts this can't happen, but nothing established it.

A batch surfaced by finishOpening is now marked unseen and ignores focus loss until the user proves they've looked at it — a key press or a click on the panel. Hover deliberately doesn't count: the panel opens at the cursor, so mouse drift would clear the flag in the same moment the browser takes key. A click can't come from a browser activating, so accepting it keeps click-away dismissal working the way 0.8.2 fixed it; without it an unseen panel would need a click and an Esc to clear. Pickers the user opened themselves behave exactly as before.

PickerPanelControllerTests covers all three states, and the unseen-survives test fails if the guard is removed.

The rest

• The no-browsers degrade path issued one NSWorkspace request per link — the same race the batching removes everywhere else. It goes through a single Dispatcher.openInBrowser(urls:) now, via a new AppState.openInFallback(urls:).
• Arriving links appended below the fold of the link list and were never revealed; the list scrolls to the newest arrival.
• The browser list was force-scrolled back to the selection on every arrival. That's gone. Worth flagging: it does not make a hand-scroll survive an arrival — SwiftUI restores its own remembered offset when the content changes. I wrote a test for that, watched it fail with the fix already applied, and dropped it rather than ship a test asserting something untrue. Separate problem, not addressed here.
• Row heights were hardcoded at 17/34pt and feed the panel's fittingSize, so Larger Text clipped the list. They're @ScaledMetric now, identical at the default text size (the layout test's 34n - 2 still holds).
• resizePanel ran twice per arrival, and read fittingSize in the same turn it set maximumHeight — a panel moved between screens could be measured against the constraint it just left.
• installPanelForTesting behind #if DEBUG; the never-read state field removed (the closures already capture what they need); Environment.live is a static let rather than rebuilding six closures per Dispatcher.
• JunctionAppTests declares JunctionMacKit explicitly instead of resolving it transitively, and DispatcherBatchTests moves to its own JunctionMacKitTests target.
• The layout tests' fixed 50ms settle() is now bounded polling — 50ms is not enough on a loaded runner and would flake.
• Dropped docs/picker-multiple-links.png; nothing in the repo references it, and the PR description links it from a raw URL anyway.
• Extended the ## Unreleased bullet to cover the new user-visible behavior.

Verification

• swift test: 113 pass on macOS (109 → 113)
• swift test in swift:6.0 under Docker, matching the Linux CI leg: 87 pass
• swift build -c release: clean, so the #if DEBUG seam holds in a shipping build
• cd web && npm test: passes, so the changelog edit parses

Not verified: the universal bundle + relocated-resource CLI smoke test, and real browser-tab delivery. Those need Actions authorized on the fork.

One I left alone deliberately: choicesProvider() now running once per session rather than once per link. It's a behavior change from main, but freezing the browser list while the picker is open seems right — rows shifting under the cursor mid-selection is worse than a slightly stale list. Happy to add the refresh back if you disagree.

[jnahian]

Applied these directly to codex/fix-concurrent-link-picker instead (commit 0810472), so they're now part of jnahian#14 rather than a separate review round. Closing this — the full rationale for each change is in the commit message and in the description above, and the dismissal-semantics change is the one worth a second opinion if you have one.