Track the deferred field/alias prefix scope as [1.10]

The prefix-truthfulness rule shipped covering TypeRef only. Extending it to
AliasedTypeExpression -- a field type or a type alias -- is a DECISION rather
than effort, and it roughly triples a migration that already touches 1,032
sites in 188 of 188 models, so it is filed with the measurement rather than
assumed.

NOTEBOOK records the rule, that it fires in both directions, and the JSON
fidelity defect it exposed: a keyword field that only ever held its default
value hid a round-trip loss.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: reid-spencer

BACKLOG.md

@@ -237,6 +237,8 @@ task files, so they were NOT physically reshuffled; this index carries the order
 | 11 | Cross-context `tell` isolation seam | 5 | Largest item and the biggest corpus-migration risk. Needs a counting mode built and run under real resolution before the Error ships. Wants its own plan. |
 | ~~12~~ | ~~Two narrow-operand gaps~~ | — | **Part 1 DONE `faf7551c0`** — and it was NOT false-positive-only as filed, it was a missed Error. Part 2 (`emittedMessageTypes`) split out and left last: genuinely advisory, genuinely a restructuring. |
 
+| 13 | **[1.10]** Prefix truthfulness for fields/aliases | — | A DECISION, not effort. Shipped for `TypeRef` 2026-08-24; extending to `AliasedTypeExpression` roughly triples an already 1,032-site migration, so measure first. |
+
 Three real dependency edges, not twelve: **4 ← 3**, **10 ← 9**, **11 ← 5**.
 Everything else is independent and may be reordered by appetite. Items 10 and 11
 each want an approved plan before implementation, per the standing rule.
@@ -898,6 +900,37 @@ each want an approved plan before implementation, per the standing rule.
   replay at all. The cache-restore path is real and is now closed, but the timing observation
   was not the proof it was presented as.
 
+- **[1.10]** **Does prefix truthfulness extend to a field's type and a type alias?**
+  — a DECISION, not effort. The check shipped 2026-08-24 (`cb05e3748`) covering
+  `TypeRef`: portlet types, invariant `requires`, a function's `requires`/`returns`.
+  A field's type (`two: Ctx.Cmd`) and a type alias (`type A is Ctx.Rec`) are
+  **`AliasedTypeExpression`** (`AST.scala:1958`) — a different node that carries its
+  own `keyword` and does NOT route through `checkRef`, so nothing checks them.
+  Verified against the staged binary: `record Holder is { two: Ctx.Cmd }` naming a
+  command draws zero errors.
+
+  **The requesting task asked for it** (`task/done/2026-08-24-reference-prefix-must-
+  match-declared-kind.md`): *"the rule is about `type_ref` generally — a field's
+  type, a `let`'s ascription, an `updates repository` … applying it everywhere at
+  once is the honest reading of the requirement."* It was deliberately NOT done,
+  and the reason is cost, measured rather than guessed: reactive-bbq alone holds
+  **283** portlet references and **542** aliased field references, so extending the
+  rule roughly TRIPLES a migration that already touches 1,032 sites in 188 of 188
+  models. That deserves its own estimate, not a ride-along.
+
+  **Pinned, so it cannot quietly become permanent**: `ReferencePrefixAndRepository
+  PortsTest` has a case named *"NOT yet cover a field's type, which is a different
+  AST node"*. If the scope widens, that test fails — which is the intended
+  direction; update it rather than narrowing the check back.
+
+  **What to decide:** whether a field typed by a record must read `two: record
+  Ctx.Cmd`. The argument for is Reid's own — the prefix exists to make a model
+  readable without chasing the declaration, and a field is where a reader most
+  often wants that. The argument against is that it is the single most common
+  syntactic position in the language, so the cost is not proportional to the
+  ambiguity being removed. Measure the corpus-wide count before choosing; the
+  542 above is one model.
+
 ### 2. Queued, needs a plan
 
 #### Decided in `../RIDDL-Tools-To-Do-List.md` but never built

NOTEBOOK.md

@@ -10,6 +10,41 @@ Orientation for a session with no memory of this work. **Open work is in
 `BACKLOG.md`**; durable facts are in `CLAUDE.md`; what things TAUGHT us is in this
 NOTEBOOK's body. Ask `git` for branch, tree and unpushed span.
 
+**`../bin/riddlc` IS `2.0.0-rc.24-5-cb05e374`, ahead of rc.24, and riddl-models is
+ALREADY RUNNING IT (2026-08-24).** No RC was cut for it, deliberately. Beyond rc.24 it
+carries: product-on-stdout, morph-after-morph, and now **prefix truthfulness**.
+
+**A reference's prefix must name what the target was DECLARED as.** Keyed off the
+DECLARATION, never off what the reference carries — an alternation declared
+`type XEvent is one of { … }` IS a type, so `is type XEvent` stays legal, and keying
+off the carried kind would have reddened all 230 such references in reactive-bbq and
+been wrong about every one. **A BARE reference is held to the same standard** (Reid's
+ruling): `TypeRef.keyword` defaults to `"type"`, so the AST cannot tell an omitted
+prefix from a written one, and the prefix exists to remove exactly that ambiguity.
+**Corpus cost: 1,032 sites in 188 of 188 models** — no model is clean.
+**The rule fires in BOTH directions**, which the requesting task did not anticipate:
+riddl's own `dokn.riddl` had `is event CompanyEvent` where `CompanyEvent` is a declared
+*alternation*, so that site LOST a keyword rather than gaining one.
+
+**SCOPE IS PARTIAL AND TRACKED AS `[1.10]`.** Only `TypeRef` is covered — portlets,
+invariant `requires`, function `requires`/`returns`. A field's type and a type alias
+are `AliasedTypeExpression`, a different node, and are NOT checked. Deliberate: 283
+portlet vs **542** aliased field references in reactive-bbq alone. A test named *"NOT
+yet cover a field's type"* pins the boundary; if scope widens that test fails, which is
+the intended direction.
+
+**It also exposed a REFLECTIVITY defect worth remembering the shape of:** the JSON
+surface dropped a `TypeRef`'s keyword on schema fields, so a round trip rewrote
+`as record X` to `as type X`. BAST was already correct. It was invisible for as long as
+`type` was the only keyword anyone wrote — **a fidelity bug can hide behind a field
+that only ever holds its default value.**
+
+**`riddlc find -replace` did 13 of the 24 in-repo migration sites**, at Reid's
+suggestion, and found sites in an INCLUDED file that a grep on the entry file would
+have missed. The script reads `source` + `type.carries` from the projection and returns
+the source unchanged when `carries` is absent (an alternation), so those are left alone.
+
+
 **`../bin/riddlc` IS AHEAD OF rc.24 — it is `2.0.0-rc.24-3-40c0574f`, staged with
 `publishLocal`, and that is deliberate (2026-08-24).** Reid's call: *"don't build an RC,
 just stage riddlc to ../bin and publishLocal — I want to avoid this round-trip between