fix(imsg-react): tapbacks on a macOS 27 gateway

[ianswope]

On a macOS 27.0.1 gateway every outgoing tapback failed. Four causes, each measured on the Mac:

1. Consent check. system_events_consent reads the per-user TCC.db, which macOS 27 no longer has (the folder is gone), so it always answered "cannot confirm the Automation grant" even though imsg-read's System Events calls succeed. With no file it now asks macOS via AEDeterminePermissionToAutomateTarget(..., askUserIfNeeded=false), which answers and never prompts (System Events is launched hidden through LaunchServices first; the API answers -600 for a target that isn't running). A TCC.db that exists still decides, unchanged.
2. No tapback actions on bubbles. A 27 bubble offers Tapback Details…, Reply…, Attach Sticker…, Copy, Share…, Delete…, none of the six. The bubble search therefore found nothing and could not even page (no-bubbles: it pages with a live bubble's actions). On 27, imsg-react uses Edit ▸ Tapback Last Message… instead, which opens the Tapback Picker on the conversation's newest message with buttons named Heart, Thumbs up, Thumbs down, Ha ha!, Exclamation mark, Question mark. The picker labels its message (Message, <text>); nothing is pressed unless that matches the target. chat.db still decides success.
   Neither an AX press, a CGEvent click nor a System Events click at selects an older bubble (the item stays "Tapback Last Message…"), and arrow keys don't move the picker, so on 27 only the newest message can take a tapback. An older one is refused up front (not-last, exit 65) without touching Messages.
3. Blocking click. A System Events click on that menu item can block until the picker closes (seen by index and by name), so it is sent from AppleScript under ignoring application responses. The button is pressed by its action name, AXPress (press is only the description).
4. Multi-line labels. parseLabel used ., so a bubble whose text spans lines never matched (not-rendered). Also affects 26.

Each run appends <time> <code> exit=<n> to ~/.blip/react.log on the Mac (no GUID or text, rotated at 64 KB), which is how the last two causes were found. README's tapbacks section gains a paragraph on 27.

Verified

• A thumbs-up on the newest message of a 1:1 on a 27.0.1 gateway landed in chat.db (type 2001, 16.9 s), react.log: added exit=0.
• test_react.py: 60 pass (new: no-TCC.db consent answers, picker route, not-last refusal before any UI, picker on another message presses nothing, failed menu click presses nothing, 27 skips the bubble search, 26 path unchanged). bun test: 750 pass.

Not addressed

On the failed runs no error reached Blip's status line, although imsg-react exited non-zero with a reason on stderr. Not investigated here.

[nixfred]

Merged, thank you Ian. This is the one that matters for me personally: my gateway is on macOS 27.2 with no per-user TCC.db, so the old consent check could never pass and tapbacks were dead here. On my Mac the new check now answers macOS 27, granted, and nothing prompted. The safety story is right too: the picker is pressed only when its own "Message, " label matches the target, an older message is refused before Messages is touched, and the probe asks with askUserIfNeeded=false. 750 tests, 60 react tests and every bridge test green; installed on my Mac.

Your "not addressed" note (a non-zero imsg-react with a reason on stderr never reaches Blip's status line) is worth an issue of its own if you have a minute; I have not looked at it yet.

[ianswope]

Thanks Fred, and thanks for Blip itself. I use it all day and really love it, so it's a pleasure to help where I can. Glad this one brought tapbacks back on your 27.2 gateway too.

On the status-line note: I found the cause, so I'll send a PR instead of an issue. The imsg-react run lived in the view that started it, and the popout and the window are both destroyed on close. On a 27 gateway a run takes 15 to 75 seconds, so closing Blip mid-run took the run and its failure with it. The fix moves the run into BarWidget and holds a failure per conversation until a view showing it picks it up. It's deployed here; I'm confirming it on a real failure before I open the PR.