feat: ADR-0035 4i(d) — merge shared-entry-word completions

In advanced mode an entry word like `create`/`drop` has several candidate
nodes (the SQL forms + the DSL fallback), but the walker commits to one,
so completion offered only that node's continuations — `drop ` showed
just `table`, and `drop rel` dead-ended at an empty list even though the
DSL drops parse via fallback.

At the entry-word boundary (advanced mode), walk every candidate, keep the
viable (Incomplete) ones, and union their next-keyword continuations:
`drop ` → table·index·column·relationship·constraint; `drop rel` →
relationship; `create ` → table·unique·index. Deeper positions keep the
committed walk untouched (no change to insert/update/delete/select).

Each continuation is classified by producing category (Both/Advanced/
Simple) and block-ordered Both → Advanced → Simple, so they read as
contiguous groups (the foundation for the 4i(e) colour, landing next).
CompletionProbe carries a parallel expected_modes; the parse path is
unchanged (the merge is completion-only).

Tests: completion merge + partial + block-order cases; the two tests that
encoded the old single-node behaviour updated. Full suite 1911 passing /
0 failing / 1 ignored; clippy clean.

docs/plans/20260526-adr-0035-sql-ddl-4i.md

@@ -0,0 +1,157 @@
+# Plan: ADR-0035 Phase 4, sub-phase 4i — verification sweep (completes Phase 4)
+
+The closing sub-phase. Canonical scope: **ADR-0035 §13 4i**. Items:
+
+- **(a)** Refresh the `CREATE TABLE` help/usage skeleton for the 4a.2
+  `DEFAULT`/`CHECK`/composite-`UNIQUE`, 4a.3 table-`CHECK`, and 4b FK
+  forms (the only help/usage debt — 4d/4e/4f/4g/4h carry their own).
+- **(b)** `describe` display of table-level constraints (composite
+  `UNIQUE` + table `CHECK`, incl. **named** CHECKs from 4g).
+- **(c)** 4b self-ref FK pre-submit indicator: stop false-flagging a
+  `references <self>` parent as unknown.
+- **(d)** shared-entry-word completion merge.
+- **(e)** visually distinguish simple- vs advanced-mode completions.
+- **staples:** typing-surface + matrix coverage, engine-neutral error
+  pass, undo-parity (one step per statement).
+
+**User-chosen sequencing (2026-05-26):** the **(d)/(e) design conversation
+first**, then build. (d)/(e) are designed (below); (a)/(b)/(c) + staples
+follow.
+
+## 1. Baseline
+
+- HEAD `ca64434`; **1909 passing / 0 failing / 0 skipped / 1 ignored**;
+  clippy clean.
+
+## 2. (d) + (e) — settled design (user-confirmed 2026-05-26)
+
+### (d) Shared-entry-word completion merge
+
+**Bug:** in advanced mode the walker's `decide()` (`walker/mod.rs`)
+commits **one** candidate node for a shared entry word, so only that
+node's continuations reach completion — `drop ` offers only `table`, and
+`drop rel` returns empty, even though the DSL drops parse via fallback.
+
+**Fix:** in `completion_probe_in_mode`, for an **advanced-mode** shared
+entry word (the entry word resolves to **>1 candidate**), walk **each**
+candidate node speculatively (the existing `scratch`/`walk_one_command`
+machinery on a fresh context), extract each one's expected continuations
+(the same `WalkOutcome`→`expected` match the single-walk path uses), and
+**union** them. Simple mode is unchanged (its single DSL node already
+offers all DSL continuations; SQL isn't offered). A candidate that
+mismatches the current input contributes an empty set, so as the input
+disambiguates (`drop table …`) the union naturally narrows to the
+surviving candidate.
+
+### (e) Visual simple-vs-advanced distinction
+
+Each merged continuation is classified by **which categories produced
+it** — a new `ModeClass`:
+
+- `Both` — produced by ≥1 `Advanced` **and** ≥1 `Simple` candidate (a
+  continuation valid either way, e.g. `drop table`).
+- `Advanced` — only `Advanced` candidates (SQL-only, e.g. `create index`).
+- `Simple` — only `Simple` candidates (DSL-only, e.g. `drop relationship`).
+
+**Colour + order apply only when the candidate list is *mixed*** (>1
+distinct `ModeClass`) — the only place the signal is informative; a
+single-mode list (deep inside any command) keeps today's token-kind
+colours, no mode tint (avoids redundant noise duplicating the mode
+indicator). When mixed:
+
+- **Colour:** `Both` → today's token-kind colour (neutral); `Advanced` →
+  `theme.mode_advanced` (orange); `Simple` → `theme.mode_simple` (cyan).
+  These two mode colours already exist (used by the mode indicator).
+- **Order (user refinement):** group the continuations into contiguous
+  colour **blocks** in the order **`Both` → `Advanced` → `Simple`**, so
+  `Advanced` sits between the other two and each colour reads as one block
+  rather than interleaving.
+
+## 3. Architecture & change list (d/e)
+
+- **`src/dsl/walker/outcome.rs`** (or `CompletionProbe`): add a parallel
+  `expected_modes: Vec<ModeClass>` (same length/order as `expected`),
+  defaulting to all `Both` (so the single-walk path is neutral). New
+  `ModeClass` enum (likely in `completion.rs`, re-exported).
+- **`src/dsl/walker/mod.rs`** `completion_probe_in_mode`: factor the
+  expected-extraction (lines 333-356) into a helper; add the advanced-mode
+  multi-candidate branch that walks each candidate (mode `Advanced`),
+  classifies each expectation by the producing category, and fills
+  `expected` + `expected_modes`.
+- **`src/completion.rs`**: `Candidate` gains `mode: ModeClass`; the
+  keyword-building (`Expectation::Word`) carries the parallel `ModeClass`
+  through prefix filtering; when the keyword set is mixed, **block-order**
+  by `ModeClass` (Both→Advanced→Simple) within the keyword group;
+  non-keyword kinds and single-mode lists are all `Both`.
+- **`src/ui.rs`** `render_candidate_line` (≈907): when the candidate list
+  contains >1 `ModeClass`, colour each by its `ModeClass`
+  (Both=kind-colour, Advanced=`mode_advanced`, Simple=`mode_simple`);
+  otherwise unchanged.
+- **Tests:** `tests/typing_surface/` — `drop ` (advanced) offers
+  `table·index·column·relationship·constraint`; `drop rel` →
+  `relationship`; `create ` → `table` (Both) + `index` (Advanced);
+  block-order assertion (Both then Advanced then Simple); simple mode
+  unchanged. A `ui.rs` snapshot for the mixed-mode coloured hint line.
+
+## 4. (a)/(b)/(c) — outline (build after d/e)
+
+- **(a)** Extend the `sql_create_table` help/usage strings
+  (`friendly/strings/en-US.yaml`) for `DEFAULT`/`CHECK`/composite-`UNIQUE`
+  (4a.2), table-`CHECK` (4a.3), and inline + table-level `FOREIGN KEY`
+  (4b). Catalog-lockstep + vocab-audit guard wording.
+- **(b)** `describe` (`do_describe_table` / the structure renderer): show
+  table-level composite `UNIQUE` and table `CHECK` constraints (named
+  CHECKs show their name). Tier-2 snapshot + Tier-3 describe assertions.
+- **(c)** The pre-submit schema-existence diagnostic: treat a FK parent
+  equal to the in-statement `CREATE TABLE` target as valid (self-ref), so
+  the `[ERR]`/`[WRN]` indicator stops lying. Tier-1/typing-surface test.
+
+## 5. Phase 2/3 — candidates & selection (d/e)
+
+Settled via the design conversation (§2). The merge alternative
+("modify `decide`/`walk` to merge in the core parser") was **rejected** —
+it would change the shared parser used everywhere; doing the merge in the
+completion-only `completion_probe_in_mode` is lower-risk. The colour
+alternatives (always-by-mode; marker tag; defer) were presented to the
+user, who chose **mode-colour-when-mixed + block ordering**.
+
+## 6. Devil's Advocate review of this plan
+
+- **Forks escalated?** The (e) visual treatment was put to the user with
+  four options + previews; they chose option 1 with the block-ordering
+  refinement. (d)'s behaviour is ADR/handoff-specified. ✓
+- **Core-parser risk?** The merge lives in `completion_probe_in_mode`
+  (completion-only), not `walk`/`decide` — the parse path is untouched, so
+  no risk to execution/dispatch. ✓
+- **Simple mode unaffected?** The merge branch is advanced-only;
+  simple-mode completion keeps the single-DSL-node path. A test pins it. ✓
+- **Noise?** Colour/order apply only when the list is genuinely mixed;
+  single-mode lists are visually unchanged. ✓
+- **Perf?** Completion walks each candidate per keystroke — candidates per
+  entry word are few (≤ ~3), and completion is already a per-keystroke
+  walk; negligible. ✓
+- **Ordering vs existing kind-order?** Block-ordering applies *within* the
+  keyword group (where shared-entry continuations live); the
+  identifiers-first kind ordering is preserved. ✓
+
+## 7. Implementation sequence (test-first)
+
+1. **(d) merge** — `ModeClass` + `expected_modes`; the advanced-mode
+   multi-candidate walk + union in `completion_probe_in_mode`; thread
+   through `completion.rs` (functional merge, classes computed but not yet
+   coloured). Typing-surface tests (merge + simple-mode-unchanged) → green.
+2. **(e) order + colour** — block-ordering when mixed; `render_candidate_line`
+   colours by `ModeClass`. Ordering test + a UI snapshot → green.
+3. **(c)** self-ref FK indicator → test-first.
+4. **(b)** describe table-level constraints → snapshot/e2e.
+5. **(a)** CREATE TABLE help/usage skeleton + catalog lockstep.
+6. **Staples** — typing-surface/matrix sweep, engine-neutral error pass,
+   undo-parity spot-check.
+7. **Full sweep** + finished-slice `/runda` → commit proposal(s).
+
+## 8. Exit gate
+
+All §13 4i items done or explicitly deferred-with-user-confirmation; all
+tiers green, zero skips; no regression from 1909; clippy clean; written-DA
+/ `/runda` PASS; ADR-0035 §13 4i + README + requirements lockstep. **4i
+completes ADR-0035 Phase 4** — flip the ADR Status from "4i pending".