@yuneta/gobj-ui 7.25.13 → 7.25.15

package/README.md

@@ -819,24 +819,29 @@ A write is known by its topic AND its serial (`<topic>^<form_write>`): each
 topic's form counts its serials from 1, and keyed by the serial alone two
 forms saving at once shared one entry.
 
-**A write cut by the drop is not a refusal** (7.25.13). A refused write
-reads its topic again at once, so a cell edited in place goes back to what the
-store has. A write that fails while the transport is NOT in session is the
-drop: a routing adapter (gui_agent's `C_AGENT_TREEDB_LINK`) settles what it
-had in flight when its session closes. The topic cannot be read then — the
-transport refuses the read (*"cannot route 'nodes' -- not in session"*) — so
-the view logs a warning and owes the read. Writes in flight on the
-disconnect edge are owed the same way. The first edge that finds the
-transport in session again pays it: each owed topic whose table is open is
-read once. An `EV_CONNECTION_STATE` "up" that comes before the view's own
-transport is in session leaves the read owed. A form is answered once, also
-when the edge answered it before the failure arrived.
+**A drop is followed by a read of every open table** (7.25.13, 7.25.14). A
+refused write reads its topic again at once, so a cell edited in place goes
+back to what the store has. A write that fails while the transport is NOT in
+session is the drop, not a refusal: a routing adapter (gui_agent's
+`C_AGENT_TREEDB_LINK`) settles what it had in flight when its session closes.
+The topic cannot be read then — the transport refuses the read (*"cannot route
+'nodes' -- not in session"*) — so the view logs a warning and waits.
+
+A drop also hides the node events published while it lasts: another writer's
+create, update or delete reaches no table. So the disconnect edge marks the
+view, and the first edge that finds the transport in session again reads
+EVERY open table once — as the schema editor reloads its model on the
+reconnect. An `EV_CONNECTION_STATE` "up" that comes before the view's own
+transport is in session does not read; the read waits for an edge that finds
+the transport in session. An "up" with no drop before it reads nothing. A form
+is answered once, also when the edge answered it before the failure arrived.
 
 ```js
-// Session up: a cell edit goes out as update-node.
-// Session drops; the adapter answers it {result: -1}   -> no read, a warning
+// Session up, tables `users` and `roles` open.
+// Session drops                       -> writes in flight answered refused
+// (another writer creates a user; its EV_TREEDB_NODE_CREATED is lost)
 // EV_TRANSPORT_STATE {connected: true}, transport in session
-//                                                      -> `nodes` of that topic, once
+//                                     -> `nodes` of users and of roles, once each
 ```
 
 **The pkey2 of an update goes back as the record had it** (7.25.5). It names
@@ -1785,6 +1790,25 @@ size. A card on its tier's default now saves its position and nothing else;
 a closed node saves no size at all (a square is not the card's size), its
 entry keeps what it had.
 
+**A Save that does not land is written again** (since `7.25.15`). `Save`
+writes one `__graphs__` record per topic that changed, and nothing answers such
+a write when it lands, so the engine takes it for granted when it leaves. A
+refused one was recorded as saved all the same: the next Save found nothing to
+write, and the arrangement was never stored. Now the host (`C_YUI_TREEDB_GRAPH`)
+tells the engine every way the write can fail -- the backend answers an error
+(shown as before), the transport refuses the command, or there is no session
+(the write is not sent) -- with `EV_GRAPHS_WRITE_REFUSED {topic}`. The engine
+forgets what it believed the backend holds for that topic
+(`forget_refused_graphs_write()` in `graph_save_plan.js`), logs a warning and
+lights Save again, so the next Save writes the topic. A host of its own that
+hosts `C_G6_NODES_TREE` sends the same event when its write of `__graphs__`
+fails:
+
+```js
+// in the host, when the answer to its update-node of __graphs__ is an error
+gobj_send_event(engine, "EV_GRAPHS_WRITE_REFUSED", {topic: record.topic}, gobj);
+```
+
 **Every saved look has a way back, in the context menu** (since `7.23.134`,
 replacing `reset sizes` / `reset topic sizes`). In edition, the node, the port
 and the edge menus each offer three resets — this one, its kind (`reset topic
@@ -2240,6 +2264,28 @@ else.
   gobj_send_event(editor, "EV_REFRESH", {}, host);                 // Save -> "wait for them"
   // ... the load lands (model_gen N+1): the form closes, the operator reopens it
   ```
+- **A move sent by the host closes the dialog of the screen it left**
+  (7.25.15). `C_YUI_SHELL` keeps the overlays open on a move that changes only
+  the subpath (a url typed in, a link), so a dialog outlived its screen: its
+  Save, its import Preview or an orphan delete sent an event only that screen
+  declares, answered *"Event NOT DEFINED"*, and the edit was lost with no word.
+  Now an `EV_SHOW` that moves the view to another position -- at once, or when
+  a load or a write in flight ends -- closes the column form, the topic form,
+  the import (its plan goes with it) and the orphans, and tells the operator
+  `the view moved: open the dialog again` (an i18n key the consumer's locales
+  carry). The export and the check stay up: they show the texts they were
+  opened with and send nothing a screen declares. The same position sent again
+  closes nothing. A **confirmation** is a shell modal, which the move does not
+  close; it carries the position it was asked on (`seg`), and a Yes answered
+  on another position is refused with the same words. Before, a Yes to "delete
+  this column?" asked on `db/users` and answered after a move to `db2/users`
+  deleted `db2`'s column of the same name.
+
+  ```js
+  gobj_send_event(editor, "EV_EDIT_COLUMN", {col: "id"}, host);   // on db/users
+  gobj_send_event(editor, "EV_SHOW", {subpath: "db"}, host);      // form closed, said
+  gobj_send_event(editor, "EV_SHOW", {subpath: "db"}, host);      // same position: nothing
+  ```
 - **A load refused IN session keeps the model it was replacing** (7.25.9): a
   routing adapter's deadline on a slow `nodes` answers -1 while the session
   is up. The records are put back, the screen and any open dialog stay, and