Reference picker: only the first --ref <uuid> is selectable; later UUIDs time out (virtualised grid never scrolled) #282

Summary

In a multi-reference gflow image i2i --ref <uuid1> --ref <uuid2> call, the first UUID selects fine (image_ref_selected_existing) but every subsequent UUID fails: TransportTimeoutError: image ref '<uuid2>' could not be selected in the picker and has no local file to upload. A single --ref <uuid> can also fail when the asset never made it into the picker library (see #281 — history-not-updated).

Live evidence (2026-07-10, v0.30.0, agentic cohort)

• gflow image i2i "<scene prompt>" --ref 87f1f771-… --ref 372afcd4-… --model nano-pro --aspect 9:16 --project f6caf027-…
• Log: first ref → ui_automation_video.image_ref_selected_existing; second ref → the timeout above (include_selector_tier surface=image_ref_include tier=text preceding).
• Same call with local file paths for both refs succeeds — confirming the picker, not the composer, is the failure point.

Root cause (code reading)

ui_automation_video.py: _select_existing_asset (~1493) locates the tile via [role='option']:has(img[src*='<media_id>']) and waits 4s, but never scrolls the virtualised picker grid — unlike the entity picker's _find_picker_entity_tile (~1612), which loops _scroll_picker_grid (~1599) to force off-screen react-virtuoso tiles into the DOM. The first UUID is typically in the initial viewport; the second (older) asset is off-screen and never renders. Possible aggravator: stale search-input text carried across --ref iterations in _attach_image_uuid_refs (~1560).

Proposed fix

• Mirror the _scroll_picker_grid loop inside _select_existing_asset before declaring a miss.
• Reset/clear the picker search input at the start of each ref iteration; re-resolve the dialog locator freshly.