Real-world programs reference object indexes well past the coordinator's
discovery range — typical example from a live OmniPro II: a program
that's "Turn ON Unit 33025" where the unit number is a raw byte value
from undecoded extended-output addressing. The discovered units bucket
only covers slots up to ~511, so 33025 doesn't match any entry.
Before this commit the dropdown silently fell through to the first
known unit (e.g. ROOM ONE), making it look like the user had selected
that unit. The underlying draft.pr2 stayed at 33025, but a user who
glanced at the form and clicked Save would either preserve the original
(if they didn't touch the select) or accidentally clobber it with the
first list item (if they did).
Fix: _bucketWithPreserve prepends a synthesized option
"(undiscovered <kind> <idx> — preserve original)" when the current
value isn't represented. Applies in all four picker sites:
* Action object picker (Unit / Zone / Area / Button for action commands)
* EVENT trigger Button picker
* EVENT trigger Zone picker
* EVENT trigger Unit picker
The synthesized entry sits at the top of the list (visually distinct)
and is the selected default. Picking any other entry from the dropdown
then becomes an explicit choice — no more silent coercion.
Smoke-tested against the real panel: slot #1 (WHEN OPEN BIG GAR →
Turn ON Unit 33025) now shows "#33025 (undiscovered unit 33025 —
preserve original)" as the selected Unit option. Screenshots updated.
Two bugs surfaced when smoke-testing against a real OmniPro II:
1. Empty list after page load. _discoverViaList ran fire-and-forget;
connectedCallback then synchronously checked _entryId (still null
because await hadn't resolved) and skipped _loadList. The panel
rendered "No programs match the current filters" forever — until
the next 5-second poll tick, which never fires because
_startRefreshTimer was also gated on the same null check.
Fix: have _discoverViaList itself trigger _loadList and
_startRefreshTimer after _entryId lands. The connectedCallback /
updated paths can stay gated on _entryId; the discover path now
takes ownership of "do the initial load too."
2. Dev installs with both a working entry and a setup_retry entry
(mock container down, real panel up) had the panel pick the
setup_retry one first and surface "panel not configured" on every
call. Fix: prefer entries with state === "loaded" in the discover
step, falling back to first entry only when none are loaded.
Also: screenshot.py drops the seed-via-WS step (was unsafe — would
write Programs to whatever entry is loaded, including real panels).
Updates the in-page click helpers to walk the shadow DOM recursively
instead of hardcoding HA's host-element path, so detail/editor
screenshots work on the actual depth-8 element location.
Smoke test against real panel: 154 programs render correctly with
structured English, BEDTIME / OPEN BIG GAR / Zone 133 events all
decoded, B. GAR MAN DOOR [SECURE] live-state badge visible.
Detail panel + editor mode both function end-to-end.
