The top-layer + cascading-close behavior is exactly why we finally dropped a bunch of custom overlay managers.
On positioning: we got farther with CSS anchor positioning once we stopped treating it like absolute offsets and started from "this is the preferred edge, then let the browser flip." Still awkward for nested context menus, but far less brittle than measuring getBoundingClientRect in a resize observer.
The watercolor framing is closer to my experience than the director one.
A director can assume the camera person will push back if a shot is unusable. An LLM will cheerfully hand you something that compiles, looks polished in the demo path, and quietly fails on the cases you didn't spell out. Retakes help, but only if you already know which frames are wrong.
That's the part of the steak essay that lands for me: the machine will follow a recipe at scale. Judging "this is charcoal dressed as medium-rare" is still on you.
Yeah — "the eMarker said 40Gbps" is basically the same problem as trusting the print on the jacket.
For home use I've ended up treating real capability as an empirical property of the specific cable+port+device combo: does this exact cable drive this monitor at the mode I want, under load, for an hour? If yes, it gets a colored heat-shrink band and never leaves that use case. Everything else is a phone/charge cable until proven otherwise.
On positioning: we got farther with CSS anchor positioning once we stopped treating it like absolute offsets and started from "this is the preferred edge, then let the browser flip." Still awkward for nested context menus, but far less brittle than measuring getBoundingClientRect in a resize observer.