rdbms-playground/docs/handoff/20260515-handoff-14.md

Session handoff — 2026-05-15 (14)

Fourteenth handover. This session cleared the remaining handoff-12 backlog: every §2.1 carry-forward and the §2.2 deferred items. Eight focused commits, each a deliberate decision the user signed off on.

Headline: the handoff-12 §2 catalogue is now empty of actionable items. What remains are two flagged observations (§3) the user should rule on, plus the standing ADR roadmap.

State at handoff

Branch: main. Working tree clean. origin/main is at 42cf851 (handoff-13); local HEAD is 8 commits ahead — this session's work, unpushed (the user pushes asynchronously).

Commits since handoff-13:

50b7825 Remove dead parse.token.* catalog entries
bcc5ad2 Matrix: pin natural candidate ordering
f1ff597 Hint: pedagogical Form-A pointer at Form B's first value slot
911a537 Walker: node-attached HintMode via Node::Hinted
9bbb96e Walker: memoize DynamicSubgrammar resolution to bound the Box::leak
90e3f5d Insert grammar: Form C type-awareness via lookahead
f46606b Runtime: schema-aware replay parsing
03dd900 Help: consume CommandNode.help_id — REGISTRY-driven in-app help

Tests: 1006 passing, 0 failing, 1 ignored (up from 989). The ignored test is the long-standing ```ignore doc-test in src/friendly/mod.rs.

Clippy: clean with nursery lints + -D warnings.

§1. What shipped — handoff-12 backlog cleared

Dead parse.token.* catalog entries removed (50b7825)

The 5 structural-class + 3 lex-error entries handoff-12 §2.1 listed as unreachable are gone (catalog YAML + keys.rs).

Ranker / natural candidate ordering (bcc5ad2)

The user's actual ranker need — to before table so add column to table T reads in order; keywords before schema identifiers — already worked via declaration-order preservation + keywords-first sectioning in candidates_at_cursor. Nothing pinned it; 8 matrix tests in tests/typing_surface/candidate_ordering.rs now do. See §3 for the Ranker type itself.

serial/shortid pedagogical Form-A hint (f1ff597)

handoff-12 §2.2: at the first value slot of insert into T values (…) for a table with auto-generated columns, the hint now appends "(id auto-generated — skipped here; list columns explicitly … to set it)". hint_resolution_at_input derives the skipped columns from the post-walk WalkContext (Form B = no user_listed_columns + table has serial/shortid columns); the note fires only at the first slot. New HintResolution::form_b_autogen_skipped, catalog key hint.value_slot_autogen_skipped.

Node-attached HintMode (911a537)

handoff-12 §2.1: the hint resolver's signature-matching (does the expected set contain all five literal forms? an Ident{NewName}?) is replaced by a grammar-declared annotation. New Node::Hinted { mode, inner } wrapper; the walker records the mode in WalkContext::pending_hint_mode on entry and clears it on any successful match (the cursor moved past the slot — this also undoes the leak where a failed Hinted branch of a Choice would strand a stale mode). The resolver reads pending_hint_mode directly.

Mechanism note: handoff-12 sketched threading HintMode through the Expectation enum. ADR-0024 §HintMode only says "nodes carry HintMode, the walker propagates it" — mechanism- agnostic. The WalkContext::pending_hint_mode route (mirroring the existing pending_value_type) was chosen as lower-risk; the user was told and did not object.

DynamicSubgrammar memoization (9bbb96e)

handoff-12 §2.1's Box::leak-per-walk. The handoff's arena sketch was unworkable (it needs a lifetime-generic Node — a major refactor). Instead resolve_dynamic memoizes factory output on the schema state the factory reads (keyed by factory fn-pointer + ctx fields). Each distinct value-list shape leaks once — total leak bounded by distinct (schema × form) combinations, not keystroke count. TableColumn gained Hash.

Form C type-awareness (90e3f5d)

handoff-12 §2.2. Form C (insert into T (vals)) shared the ( opener with Form A, so its values weren't typed. The explicit-Choice-branch split is impossible (committed-choice semantics commit after ( matches), so a new Node::Lookahead(fn(&WalkContext, &str, usize) -> Node) variant peeks the source: a value-literal first token routes the paren through the typed column_value_list (Form B dispatch contract); an identifier or empty paren routes to a Form A column-name list. Form C values are now type- and count-checked at parse time. insert into T ( cleanly shows Form A column candidates instead of mixed Form-A/C suggestions.

Schema-aware replay (f46606b)

handoff-12 §2.1: run_replay parsed schemalessly. It now re-snapshots the schema per line (extracted build_schema_cache, shared with the interactive path) and parses with parse_command_with_schema — typed-slot rejections fire at replay parse time, matching interactive. New integration test replay_rejects_typed_slot_violation_at_parse_time.

help_id consumption (03dd900)

handoff-12 §2.1: every CommandNode declared an unused help_id. note_help now iterates the command REGISTRY and translates each help_id — a new command appears in help automatically. 20 per-command catalog entries + 3 framing entries; help.in_app_body removed. CommandNode.help_id lost its #[allow(dead_code)].

§2. Bug found this session

libyml 0.0.5 scanner panic on long space runs in double-quoted YAML scalars. While authoring the help entries, a space-aligned double-quoted catalog string ("quit — exit") panicked the YAML scanner with "String join would overflow memory bounds". Block scalars (|-) are unaffected — that's why the old block-scalar help worked. Bisected and worked around: all per-command help entries use |-. If you author new catalog entries, avoid long internal space runs in double-quoted ("…") values — use a block scalar or keep runs short. A catalog comment in en-US.yaml records this.

§3. CRITICAL: two flagged items needing a user decision

3.1 The Ranker type is vestigial. completion::Ranker / candidates_at_cursor_with have no production caller passing a non-identity ranker. The user's stated ranker need (candidate ordering) is met by declaration-order preservation, not the ranker layer. So Ranker, identity_ranker, and the candidates_at_cursor_with variant are unused scaffolding. Per CLAUDE.md "don't remove without confirmation" they were left in. Decide: remove them, or keep for a future frequency- ranking feature? (handoff-12 §2.1 listed the ranker as "scaffolding-only … future work" — this is the same item, now confirmed genuinely unused.)

3.2 CommandNode.hint_mode is now genuinely dead. The per-command hint_mode: Option<HintMode> field predates the node-attached HintMode work; HintMode is now per-node (Node::Hinted), never per-command. The field is still #[allow(dead_code)] and read by nothing. Removing it is a safe mechanical edit across the 20 CommandNode declarations. Decide: remove it, or keep? Not done this session (20-site edit, separate from the HintMode mechanism change).

§4. Open items — standing roadmap (unchanged)

handoff-12 §2's actionable backlog is cleared. What remains is the ADR roadmap in CLAUDE.md "Things deliberately deferred" (complex WHERE expressions, SQL advanced mode, indexes, m:n convenience, snapshot/replay/undo, tutorial system, etc.) and handoff-13's two accepted items:

• Partial entry words classify as DefiniteErrorAt (handoff-13 §3) — the user accepted this; the matrix test app_commands::partial_entry_word_classifies_as_definite_error_but_completes documents it.
• Matrix scope (handoff-13 §4) — cursor coverage is "meaningful transitions" not every byte offset; assertion (5) is parse-layer not a live dispatch differential. User accepted both.

§5. Architectural delta (vs. handoff-13)

New Node variants

• Node::Hinted { mode: HintMode, inner: &'static Node } — node-attached hint-mode annotation.
• Node::Lookahead(fn(&WalkContext, &str, usize) -> Node) — source-aware dynamic subgrammar (Form A/C discrimination).

New WalkContext field

• pending_hint_mode: Option<HintMode> — set on Hinted entry, cleared on any match.

Walker driver

• walk_node split into a wrapper (clears pending_hint_mode on match) + walk_node_inner (the dispatch).
• resolve_dynamic + DYNAMIC_CACHE — memoized DynamicSubgrammar resolution.
• Node::Lookahead arm — not memoized (source-dependent), returns a small node.

New API surface

• input_render::classify_input_with_schema (added handoff-13, noted here for completeness).
• HintResolution::form_b_autogen_skipped: Vec<String>.
• runtime::build_schema_cache (extracted from refresh_schema_cache).

Catalog

• Removed: parse.token.* (×8), help.in_app_body.
• Added: hint.value_slot_autogen_skipped, parse.custom.insert_form_a_missing_values (handoff-13), help.intro / help.dsl_section / help.types_reference, help.{app,ddl,data}.* (×20).

§6. How to take over

1. Read this file, then handoff-13, then 12 for the chain.
2. Read CLAUDE.md — the working-style rules. This session escalated every ambiguous fork (HintMode mechanism, the Box::leak arena's true cost, the Form C restructure twice) rather than deciding unilaterally.
3. Run cargo test — 1006 passing, 0 failing, 1 ignored.
4. Run cargo clippy --all-targets -- -D warnings — clean.
5. Resolve §3 — the two flagged dead-code items — with the user.
6. Then the standing ADR roadmap (§4) is the next structural work; pick per the user's priorities.

Note on the typing-surface matrix

tests/typing_surface/ (now 144 cells) is the regression net for everything walker/hint/completion. After any grammar or walker change: a failing matrix cell with correct new behaviour → update its snapshot (INSTA_UPDATE=always cargo test --test typing_surface_matrix <family>); a failing cell with wrong behaviour → the cell earned its keep. The Form C type-awareness work this session was guarded entirely by it.