@ai-matrx/records 0.54.0 → 0.56.0

package/CHANGELOG.md

@@ -1,5 +1,83 @@
 # Changelog — @ai-matrx/records
 
+## 0.56.0
+
+**A Person or File column takes an id** (lane RELATION-TARGETS): `coerceTypedAnswer`'s member and
+attachment arms pass a member's user id, an uploaded file's id or the kernel record's own id
+through as the value, because the store resolves it to the Person / File record at every write
+door; a typed NAME is still refused, with the control that holds it.
+
+**And the grid's store primitives** (lane GRID-PRIMITIVES), below. Published by PUBLISH-SWEEP-4.
+
+### The grid's store primitives
+
+**Everything the /data grid rebuild needs from the store, as typed calls** (`src/grid.ts`):
+`tableDecorate` / `tableDecorations` (row color by a choice column, ordered color rules, hand
+highlights on rows, cells and columns — Fields and records by id), `gridLayout` (platform →
+organization knob `custom/grid_layout` → the saved view's own `definition.layout`, each choice
+saying where it came from), `formulaParse` (the older grid's formula text → the store's
+expression), `actionDeclare` / `rowActions` / `actionRun` (a row action on a whole selection in
+ONE transaction; `actionRefusals(error)` lists every record and field that refused),
+`tableWebhookDeclare` / `tableWebhooks` / `tableWebhookArchive` (every record change out to a
+signed webhook), `tableFromExample` (one real use case of `./use-cases`, built whole),
+`autonumberBackfill`, `ioExport` and `exportXlsx` (the export door's own page as an .xlsx, a
+withheld column named on its own sheet). `AggregateOperation` gains `median`, `filled`, `empty`,
+`unique`.
+
+**Consumer action:** none until the chair applies the five `gridprim_*` migrations to the main
+database (01:00–04:00 PT) and `pnpm records:generate` lists the doors in `STORE_DOORS`. Until
+then every new method answers `door_absent` naming GRID-PRIMITIVES — never a guess, never an
+empty answer.
+
+## 0.55.1
+
+**0.55.0, actually published.** The tag `npm/records/v0.55.0` never reached the registry: its
+publish gate failed on `pnpm typecheck` (OLD-TABLES-2's relation-words suite had four type
+errors vitest never reads) and then on two `real-doors` clauses that still asserted the record
+store OFF after STORE-ON's owner ruling made it ON by default. Both are fixed in the suites; no
+runtime byte differs from what 0.55.0 describes below. The version moved rather than the tag
+because a published tag is never rewritten. (lane PUBLISH-SWEEP-4.)
+
+## 0.55.0
+
+**Three things that were already in the source and had never reached a consumer.** (published
+by lane PUBLISH-SWEEP-4; built by STORE-TXN, OLD-TABLES-2 and ORG-CLEANUP-2, each of which
+landed its work marked "source only, not published".)
+
+* **`recordWriteGraph`** — the typed client over `custom.record_write_graph` (lane STORE-TXN).
+  A parent, the edges it stands on and the rows that belong to it in ONE transaction, with one
+  op id for the whole graph. `recordWrite` and `recordWriteMany` are each one statement, so a
+  caller needing a parent and its children to land together had no door: a loop is N
+  transactions and a failure on child nineteen leaves a parent standing without the lines it
+  promised. Not a second write path — the door calls `custom.record_write_many`,
+  `public.assoc_add` and `platform.relation_set`, so every guard and trigger runs once as
+  before. `RecordGraphWritten` counts the two edge writers apart (`edge_ids` / `relations`)
+  instead of making a caller subtract.
+* **`relationWordsMany` dispatches by store** (lane OLD-TABLES-2). A relation cell holds an id
+  and a screen shows words, and there are two stores holding such cells: `custom.record` and
+  the pre-doctrine `workbench.udt_*` tables, where most people's data actually sits. ONE client
+  primitive now reads both, discriminated by an optional `store` — and **a caller that names
+  none reaches exactly the door and the arguments it reached before**, so the thirteen live
+  screens on `custom._words_for` change not one line. The older arm goes to
+  `workbench.udt_row_words_many` through `WORKBENCH_DOORS` / `WORKBENCH_SCHEMA`, named by hand
+  for the same reason `TRUST_DOORS` is (`records:generate` reads `custom` and only `custom`).
+  The arms are asymmetric on purpose: the unified arm takes the FIELD because that is where the
+  display spec is declared and judged; the older store has no declare-time judge, so its arm
+  carries the spec. An id the door answered nothing for stays ABSENT from the map either way —
+  inventing "Untitled" would erase the difference between a cell pointing at nothing and a row
+  that was refused. New types: `RelationStore`, `RelationDisplaySpec`, `RelationWordsManyArgs`.
+* **The `throwaway` placeholder pattern** (lane ORG-CLEANUP-2) — after 48 organizations named
+  "Throwaway Dash 3fee58c2" walked back onto the owner's organization picker overnight and
+  `PLACEHOLDER_PATTERNS` could not see one of them: 28 of the 48 carried no `zz` at all. It
+  now catches "throwaway", "disposable org", "safe to delete" and "test only", and mirrors
+  `iam.placeholder_in_a_name` on the platform database — a pattern added to one belongs in the
+  other.
+
+Also regenerated `store.generated.ts`, which was stale (`real-doors.test.ts` was red on main
+claiming `record_write_graph` absent while it exists). Proof: `src/__tests__/relation-words-by-store.test.ts`
+goes RED on 3 of 5 clauses with one line of the dispatch reversed. Suite at publish: 105/105
+across 9 files.
+
 ## 0.54.0
 
 **A formula can be re-aimed through a typed door.** (lane STORE-TAILS-2's ask, published by