max / goingson

Picking a file is the host's call, so the imports work again THE BUG. quasi::data's three imports and the project dashboard's attach could not receive a file, and had not been able to since the described screens started serving. A FieldKind::File renders <input type="file">; htmx submits it urlencoded, where a browser reports a masked filename rather than a path; and quasi_http::is_form refuses multipart outright, on the grounds that a description has no word for a byte stream. So the field delivered neither bytes nor a path, while every handler read payload["file"] as a filesystem path. The tests could not see it. They hand the handler a real path as a param, which is a value the renderer would never send, so they passed against a screen that did not work. WHY IT LOOKED FINE. quasicoherent 3b830122 measured this in August and recorded "Tauri can hand over an absolute path, which is what POST /projects/{id}/attachments reads today", concluding one site, working, Tauri-only. What handed over a path was the JavaScript: the old screens called window.__TAURI__.dialog.open() and passed the command the result. Tauri's webview is still a webview, and the described field never went through the dialog at all. That premise is corrected on the task. THE FIX is the mechanism already ruled for it. Action::by_host (a81384d4, "that the host makes this call, and the renderer does not") emits data-sends with no transport: the address is the half the description knows, and how the file gets there is the host's. The controls are acts carrying it now, with the accept list travelling as a param so a native dialog knows what it is asking for. frontend/js/host.js is the host half, and the first hand-written script since the swap deleted 84 of them. It is allowed to exist for the reason the ruling gives: picking a file is not describable and was never going to be. It opens the dialog and hands the path back through htmx.ajax rather than fetch, because the answer is an ordinary fragment with ordinary swap headers and doing the request by hand would mean reimplementing retarget and reswap and getting one subtly wrong. Its header says what it is not for. The routes are unchanged: a path under `file` is still exactly what arrives. The document shell is lifted out of protocol() as document_shell(), because what the document loads is now a fact worth asserting: data-sends carries no transport, so a document without host.js draws buttons that look pressable and are not, which is this bug again in a new place. There is a test. Three tests rewritten to assert the working shape rather than the broken one, and two added: what the control emits, and that the shell loads the host half. 794 total. NOT IN THIS COMMIT. Compose still has no Attach. Its attachments want an attachments row against the draft and a drainer that resolves blobs to paths at send time, which is the right shape anyway: a queued message may go hours later and the file it was written from can be gone, while a blob cannot.