Prune NOTEBOOK: graduate 12 facts to CLAUDE.md, file 4 backlog items

NOTEBOOK.md had reached 4,876 lines, roughly 85 dated entries. A notebook
entry mixes a story with a fact, and the two have different lifespans: the
story is scaffolding, valuable while the work is fresh; the fact is a
constraint someone needs again. This separates them.

NOTEBOOK 4,876 -> 547 lines. Removed: ~60 entries covering 2026-07-24 ->
2026-08-20, plus the JSON-input roadmap (all 9 phases done), the stale
"Open Tasks in task/" list, and four Design Nuances subsections already
duplicated in CLAUDE.md -- one of which, AIHelperPass, described a pass
DELETED in 1.24.0. Entries were kept whole or removed whole; nothing was
summarised or rewritten in place.

Graduated to CLAUDE.md, each verified live in the tree rather than taken
from the narrative:

- GBNF is GONE (deleted 2026-08-20 for measured performance, not quality);
  a grammar change touches two artifacts, not three
- TatSu's nameguard refuses a bare letter token touching a digit -- why
  `("e"|"E")` fails on `1e3` and an inline regex does not
- a new .riddl fixture is a grammar-surface change: INCLUDE_FRAGMENTS, and
  the Scala suites cannot speak for the Python validators
- validateType skips its walk for a TOP-LEVEL aggregate, so a family of
  checks fires only on nested inline ones (ValidationPass.scala:2752)
- Finder.fieldChildren: 27 field-held sites were unreachable, and the arm
  it still ends in means arm 12 of Value will be invisible
- BAST writeContents writes a count and trusts a distant caller for the
  items -- a contract that has failed four times; plus positionsKnown
- Keywords.keyword ends in a cut, so factor the prefix, never reorder;
  and test a fastparse alternation rather than reading it
- never @JSExport an overridden toString (breaks JS ToPrimitive)
- three JSON codec traps: upickle $type, hand-written codecs, $kind
- ThirdPartyNotices.scala is a hand-maintained constant, stale in silence
- run sbt --server when you need to read its output
- a case class reaching a document has an O(document) hashCode, and only
  Scala.js notices

Two corrections, both places CLAUDE.md was wrong today:

- it said scala.yml caches target/out. It must NOT be cached and is not;
  scala.yml:165 carries the ban and its reason.
- the predefined-module inventory omitted GeneratorError, which is live at
  PredefinedModule.scala:83 and is half of what "both records are unused"
  refers to.

Three NOTEBOOK sections carried open work that existed nowhere else. Per
CLAUDE.md, if it is not in BACKLOG.md it is not tracked -- so they were
filed rather than deleted: [2.7] put/get UI-boundary statements, [2.8]
single failure point per saga do-block, [2.9] TypeScript AST declarations,
[2.10] repo housekeeping (three items, each verified still outstanding; a
fourth was dropped as already resolved).

HANDOFF was stale in two ways and is corrected: the RC6 upgrade is
committed, not uncommitted, and task/ is not empty -- riddl-models filed
the saga/tell boundary request today.

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

Author: reid-spencer

BACKLOG.md

@@ -1691,6 +1691,71 @@ that needs a ruling before either can be fixed.
   this branch, and the second where the stale claim was a promissory note about
   work that had since landed.
 
+- **[2.7]** `put`/`get` UI-boundary statements. **Graduated out of NOTEBOOK
+  2026-08-26, where it had sat as a "Deferred — do NOT start yet" section since
+  before 2.0; filing it here so it is tracked rather than remembered.** Design as
+  discussed: `put <value> to output <outputRef>` pushes to a UI Output from an
+  on-clause; `get` is `let <id> = get from input <inputRef>` — its OWN statement
+  kind, not an extension of `LetStatement`. Both keywords existed long ago as
+  storage statements and were removed when `repository` sufficed. This completes
+  the boundary-statement census: send/tell = streaming, call = function, put/get
+  = UI. **Both CAN FAIL** (put when the UI is absent/headless, get when the input
+  is unset), so model it as `def canFail: Boolean` on `Statement` — one source of
+  truth shared with [2.8]. Witnessing is greenfield: a `ShowOutputInteraction` is
+  witnessed by a `put` to that output, a `TakeInputInteraction` by a handler on
+  the input's message type OR a `get` naming it; an unwitnessed step draws a
+  CompletenessWarning. **Prerequisite: the UI / application
+  (output-input-triplet) model must land first** — that is why it was deferred,
+  and the prerequisite has not moved. Open question for when it unblocks: which
+  handler statement-sets may contain put/get. Note `get from input` is ALREADY
+  partly present (`GetValue.source` is `InputRef | StateRef`) — check what exists
+  before designing, per the standing "a plan cannot notice the work happening"
+  rule.
+
+- **[2.8]** Single failure point per saga do-block. **Graduated out of NOTEBOOK
+  2026-08-26 with [2.7]; same provenance.** A saga step's do-block is
+  all-or-nothing (undo assumes all-or-none happened), so warn when it contains
+  more than one potential failure point. **Blocked on the complete can-fail
+  census**, via the shared `Statement.canFail` in [2.7]. The census as filed:
+  send, tell, call, yield, put, get CAN fail; let, set, when, match, foreach
+  CANNOT. **Deferred rather than shipped-partial deliberately** — with only
+  send/tell counted it would give false "single failure point" passes that flip
+  to warnings later, which is worse than waiting. **The census itself needs
+  re-deriving before use**: it predates `forward`, `initiate`, `terminate` and
+  `error`-is-terminal, and this repo's standing lesson is that a predicate
+  borrowed or inherited wholesale stops answering its own question (see
+  CLAUDE.md on A23 vs A26).
+
+- **[2.9]** TypeScript AST declarations for the JS side. Low priority. The full
+  AST hierarchy (Domain, Context, Entity, …) is still OPAQUE to TypeScript; only
+  the public `RiddlAPI` methods are declared. JS consumers are expected to use
+  the facade rather than the raw AST, which is why this has never blocked
+  anyone — file it, do not schedule it. Graduated out of NOTEBOOK's "Active Work
+  Queue" 2026-08-26.
+
+- **[2.10]** Repo housekeeping — three items, all VERIFIED still outstanding
+  2026-08-26, all independent and each a few minutes:
+  1. **Delete the `development` and `old-development` branches** (local +
+     remote). `development` is fully contained in `main` — 0 commits ahead as of
+     1.31.0 — so nothing is lost. `git branch -a` confirms `development` and
+     `origin/development` still exist. Deferred pending an explicit go-ahead,
+     which is the only thing missing.
+  2. **Fix `.claude/skills/ship/SKILL.md`** — it still prescribes the GitFlow
+     pre-flight (fast-forward `main` from `development`, `SKILL.md:21-30`) and
+     the post-release merge-back. Both contradict current policy (CLAUDE.md
+     § Branch Strategy: `main` IS the working and release branch) and were
+     skipped for 1.30.0 and 1.31.0. A skill that documents a process nobody
+     follows is a trap for the next session that trusts it.
+  3. **Delete the stray `help` git tag** (`git tag --list help` confirms it).
+     Almost certainly a typo'd `git tag help`. Harmless, except that it sorts to
+     the top of `git tag --sort=-v:refname` and so LEADS the tag list when
+     working out the latest release.
+
+  Dropped from this list as already resolved: *"verify the `unset GITHUB_TOKEN`
+  guidance for `gh`"* — settled 2026-07-29 and recorded in `../CLAUDE.md`
+  § GitHub CLI Integration (both credentials work; prefer plain `gh`, and never
+  unset the token in a shell that also runs sbt).
+
 ### 3. Owed to other repos
 
 - ~~**[3.1]** riddl-generator's `Finder` incompleteness.~~ — **CLOSED by THEM