max / makenotwork

server: describe the media picker, and delete media-picker.js Shape 4 step 1 (6fb46f7c), which has been stopped at the landing since 2026-08-18: the plan's own text says "STOP HERE and report: insertAtCursor and _mediaPickerSelect have no expression. See blockers, this is the finding." That finding was quasicoherent f35aafee, Max ruled (a) on 2026-08-21, and Act::fills shipped in quasi 0.54.0 this morning. So the rest of the step goes in one pass. src/quasi/media_picker.rs is the glue module, three routes on the media router serve it, and four Insert Image buttons across three templates call it. A card is Act::new(filename, Action::local()).filling(into, markdown_ref): the destination is a field name, never a caret, and the press makes no request -- the old one made none either, it just did 155 lines of DOM building instead. FIVE THINGS THE PLAN DID NOT ANTICIPATE. The picker's region is per destination, not per page. item_details.html carries three of these buttons, and one shared `media-picker` id would be three elements answering to it. `trigger` ships the button and its own empty region together, and `into` is gated to a plain handle before an id is built from it -- it arrives on a query string, and what the gate stops is not an injection (the emitter escapes) but a well-formed selector meaning something other than the element it names. A filter answer is aimed at the grid, one region further in than the plan said. Aiming it at the modal replaces the box being typed into and takes the caret with it on every keystroke. The name box also asks through a Consult with a 150ms wait rather than through `changes`, which is discover_typeahead's call for discover_typeahead's reason: a media library is a Postgres round trip and `changes` fires per keystroke. The folder select keeps `changes`, since a select settles rather than being typed into. A modal region needs two boxes and MNW had no rule for either. layout.css has no .region rules at all, so the first described modal on this host would have rendered in flow while claiming role="dialog" aria-modal="true" -- an attribute that would have been a lie. The scrim is the modal region and the panel is a region inside it, because one element cannot be both without its contents being laid out across the viewport. Both rules are in style.css now and every described modal after this one gets them. Three dead one-off button rules go in the same pass: .insert-image-btn, .blog-editor-media-btn and .text-editor-insert-btn were three spellings of the same two declarations. The destination has to be unique in the document, which is new. Act::fills names a Field::name, and a name identifies a field within one description; Filling::id_prefix exists because that stops being true when a document holds two copies of one form, and rich_field leans on it -- five editors, all named `body`, told apart by their ids. So an editor a picker can write into has to be the only `body` in its document. The two surfaces that had a working button already are. item_details.html's two section bodies were raw <textarea> elements with an id and no name at all, addressable by nothing, so they are described now and carry names of their own through the new rich_field::named. Their ids are unchanged, which is what item-details.js reads them by. The shipped picker had three defects and one of them wrote to the database. The JSON field is `filename` and the script read f.file_name, which is undefined: every tile shipped unlabelled and the name filter matched nothing, so typing one character hid the whole library. And markdown_ref is already ``, which _mediaPickerSelect wrapped a second time -- so every insert wrote `)` into somebody's document. The first two die with the file. The third is filed as 2554d643, because no redeploy repairs a written document. THE GAP IT FOUND, in the other direction. Act is a label, and nothing says a control reads as a picture. Cell::new and Row::new both take a string, so a picture cannot sit beside a control in a table or a list either: Node::Image goes in a region body and nowhere else. A media tile is therefore a region holding two members and only the file name answers a press, which on a picture picker is the affordance. Filed as quasicoherent db998898 with three candidate spellings and the measuring to do first. frontend_globals HIGH_WATER 113 -> 111: window.mediaPickerOpen and window._mediaPickerSelect go with the file.