@ai-matrx/records-ui 0.62.0 → 0.65.0

package/CHANGELOG.md

@@ -1,5 +1,138 @@
 # Changelog — @ai-matrx/records-ui
 
+## 0.65.0
+
+**A condition about a relation column stops asking a person to type a uuid.**
+(lane RELATION-DISPLAY-2.)
+
+* **`ConditionField` carried no type, so the ONE condition builder could not know a
+  column points at another record.** It drew the same plain text box every other
+  clause gets — and a relation's stored value is the other record's id (FLD-6 /
+  REL-1), so the only way to write a true clause was to find and TYPE a uuid. The
+  dispatcher's actual house rule, *"a job for Harborview Property Group cannot
+  leave Scheduled without a purchase order"*, was unwritable by exactly the person
+  the builder exists for. A box that accepts anything and can only produce a rule
+  matching nothing is law 4's dead control.
+* `ConditionField` now carries an OPTIONAL `field`. When it is a relation, the value
+  control is **`RelationPicker`** — the same picker the cell uses, on the same
+  `relation_words_many` the cell reads — so a filter offers the column's own joined
+  words ("Harborview Property Group, Carpinteria") and writes the ID. A filter and
+  the column it filters can no longer disagree about what a record is called.
+* It is asked for ONE record even when the column holds several: a clause compares
+  against one value, and a picker that let somebody choose three would write an
+  expression this row cannot draw back.
+* Both callers in this package already held the whole `Field` and were throwing it
+  away one line before handing it over — `StageRules` and `FormBuilder`. In the form
+  builder this is the OWNER's screen; a stranger is still never asked a relation
+  question (`publicQuestions.ts` is untouched).
+
+**Consumer action:** none — `field` is optional and a caller passing only
+`{id, key, label}` gets exactly the old behaviour, which has its own clause. A host
+that wants the picker passes the `Field` it already has.
+
+*Numbered 0.65.0, not 0.64.0: a peer lane claimed 0.64.0 for the coercion primitive
+while this work was in flight and tagged it first. Two lanes bumping one package in
+one afternoon is the cost of a shared checkout; the fix is to move, never to take a
+version somebody else's tag already names.*
+
+## 0.64.0
+
+**The paste and the editor stop owning their own opinion of a column, and the capture sheet
+stops minting NaN.**
+
+- **`editorKindFor` is `fieldKindFor`** (`@ai-matrx/records` 0.45.0) with this package's
+  kernel File/Person ids bound. The classifier body lived here and was the only one.
+- **`valueFromPastedText` is one call to `coerceTypedAnswer`**, keeping this package's label
+  so a refusal names the column the way the grid's header does. Its 150-line body WAS the
+  coercion — and while it sat here, the public booking page was unbookable over any table
+  with a number field.
+- **`ChoiceOptions`** is re-exported from the core instead of declared twice.
+- **`CaptureRun`** used `Number(raw)`, which answers NaN for "about twelve" — and a NaN
+  becomes null on the wire, so a reading taken in a yard could silently record nothing. It
+  coerces through the one body now and KEEPS the typed text when the value is refused.
+- **`ViewSwitcher`**'s calendar offer tested `kind === "datetime" || kind === "date"`. FLD-11
+  has no "date" kind; it compiled only because the old `EditorKind` union carried two words
+  nothing produced.
+- **`BuildOrAsk`** (landed 0.63.x→here): "Ask an agent" is ABSENT when `mayBuild` is false —
+  asking an agent is the same act as building by hand, through the same doors.
+- **One label map per enum** (`enumLabels.ts`): no panel keeps its own copy, and the
+  "Done"/"Finished" split between `ChecklistRunner` and `ActionInbox` is gone.
+
+*Consumer action:* none. `EditorKind` is now an alias of `FieldKind`; a host that compared a
+kind against `"date"` or `"boolean"` was comparing against a value the classifier never
+returned.
+
+## 0.63.0
+
+**A relation cell finally SHOWS what its column was configured to show, and the
+six hand-rolled copies of "what is this id called" are gone — all six, by moving
+the code, none by allowlisting.** (lane RELATION-DISPLAY-2.)
+
+* **The display spec reached the store in 0.62.0 and no screen could read it.**
+  REL-DISP gave a relation column a `display` — which of the target's columns it
+  shows, in what order, joined by what — and shipped `ReferenceBuilder` to
+  configure it. But `labels.tsx`, which every grid, board, gallery, calendar and
+  peek cell reads through, resolved names ITSELF: 200 rows of the target Table
+  plus `tableList()` for its `title_field`. So a person could set Rincon
+  Plumbing's `jobs.customer` to read "first name + last name", press Save, and
+  the board would go on drawing "Maria Chen" for both of the Maria Chens. The
+  builder wrote a setting that nothing displayed.
+  `labels.tsx` now fetches NOTHING up front and asks `relation_words_many` for
+  the ids the page actually draws — one call per Field per tick, the store's one
+  resolver, this column's own spec, the visibility ladder asked BEFORE the name
+  is read, and no 200-row prefetch ceiling. Its own comment recorded the batch
+  door as missing ("there is still no batch door on the store"); REL-DISP built
+  it, so the per-id `record_read` lane is gone entirely.
+* **`RelationPicker` offers the same words the cell shows.** It had the third
+  copy of the derivation, which is why a column configured to read "Maria Chen —
+  Rincon" still asked a person to choose between two identical "Maria Chen"s.
+  One `relation_words_many` now covers the offered rows AND the already-picked
+  ids, so what you pick from and what you read back can never disagree.
+* **`recordNameIn` / `rowNameIn` — the one naming call for a record whose Table
+  you already hold.** `Grid` (the Talk action), `Peek` (the peek title),
+  `PortalsPanel`'s client picker and the host test bench were each writing
+  `recordName(row.document, table.title_field)` out by hand. That is not the
+  store's question — there is no id to resolve — but it is the same question
+  answered five times in four files, which is how a Table titled by `cname` read
+  "Untitled" on one screen and correctly on the next. It lives once now.
+  NOT a second resolver: hold an id, and the answer is still the store's.
+* **`check:one-reference-builder` goes from a baseline of SIX to ZERO.** The
+  guard shipped saying out loud that allowlisting the six "would be the same lie
+  with extra steps". None was allowlisted. Proven red on a planted copy at
+  baseline 0, then green.
+
+**Consumer action:** none for a host that only mounts these components — the
+change is behind them. A host resolving a relation label itself should delete
+that code and call `client.relationWordsMany`; one holding a Table and a
+document should call `recordNameIn` / `rowNameIn`.
+
+## 0.62.0
+
+**THE ONE REFERENCE BUILDER.** (lane RELATION-DISPLAY.)
+
+*Written after the fact by lane RELATION-DISPLAY-2 — 0.62.0 published with no
+entry here, and a released version with no changelog row is a version nobody
+downstream can read. Sourced from that lane's own progress note, not invented.*
+
+* **`ReferenceBuilder` + `previewWords`** — the one place a reference to another
+  table is built. It asks, in the order a person thinks of it: which table →
+  what a record of it is called TODAY (its own reference column, SHOWN rather
+  than hidden) → only if they want it, which columns instead, in order, with a
+  separator → a preview on that table's REAL first rows. `FieldEditor`'s private
+  `TableToPointAt` is deleted and this stands in its place.
+* **A relation is asked its questions when EDITING too.** That branch skipped
+  every existing column that was not a choice list, so there was nowhere to say
+  which columns a link should show — the one thing about an existing relation a
+  person most wants to change. The target stays locked there, because the panel
+  sends no `relation_target` in a patch and a picker that changed it would be a
+  control that silently does nothing.
+* **`looksLikeId` and `WITHHELD_RECORD_LABEL` are exported at last** — they were
+  defined and unreachable, so a host could not agree with the package about what
+  a name is.
+
+**Consumer action:** a surface that lets a person choose a column's target table
+mounts `ReferenceBuilder` instead of building its own dropdown.
+
 ## 0.61.0
 
 **A link to a board opens a board, and a cell that is still resolving says what