Files
rdbms-playground/docs/handoff/20260529-handoff-51.md
T

16 KiB

Session handoff — 2026-05-29 (51)

Fifty-first handover. Continuation of handoff-50's bug-fix loop. The user steered this session twice with strategic judgement that proved correct both times (see §2). Three issues resolved — two bug fixes and one substantial cross-mode unification — plus two follow-up trackers spawned from #6's earlier /runda pass were tackled (#18 closed benign-but-hardened, #17 fixed). ADR-0022 gained Amendment 3; ADR-0036 gained Amendment 2.

§1. State at handoff

Branch: main. HEAD 10e5197. Tests: 2051 passing, 0 failing, 0 unexpected skips, 1 pre-existing ignored (the same ignored doctest as prior baselines). Clippy: clean. Nothing pushed (per the user's rules — push is the user's step).

Commits since handoff-50's abc9bb6:

10e5197 feat: bring simple-mode insert arity diagnostics to parity with advanced
7cccf4e refactor: make completion-drop gate schema-aware for consistency
fa5d0dc fix: insert VALUES between-values hint points at comma not close-paren

Test count moved 2040 → 2051 (+11 net across the three commits).

§2. The two user steers that shaped the session

This session is worth reading for how the user's judgement redirected the work — both calls were right and both uncovered real problems a naive pass would have missed.

  1. "Most of the open issues were reported at the same time — the fixes we already applied may have changed the outcome of other reports." This re-validation instinct was correct. Acting on it, every issue picked up got a confirm-it-still-reproduces first step. Result: #2's reported root cause was wrong (the report blamed string literals; it was not literal-type-specific at all), and #18 turned out benign (the divergence it described cannot reach the code path it gates). Neither would have surfaced without re-probing against current HEAD first. This is now the standing agreement: the first step of every bug pickup is to reproduce it on current code before designing a fix.

  2. On #17, the user rejected the "easy way out." When implementation of the chosen Approach A revealed a larger blast radius, the lead floated reverting to the lighter Approach B. The user pushed back: "this feels like mistakes were made in the past which resulted in this discrepancy, and approach B is the easy way out … I'd rather stick to A and get this sorted really cleanly — we're still in the original dev phase, it'd be good to create a strong basis for future maintenance, even if it means more work now." This was exactly right: there genuinely was a past discrepancy (issue #1's simple-mode submit teaching note was a workaround for the very gap #17 names), and Approach A — done properly — unified both modes onto one model rather than piling a third bespoke surface on top. Sticking with A also surfaced (and closed) a genuine execution-safety regression that the lighter path would have left buried.

§3. Resolved issues

#2 — VALUES between-values hint (commit fa5d0dc)

Re-characterised on investigation. The report's hypothesis (per- column inference abandoned when the first literal doesn't match the first column's type role) was wrong — probing values ('Oli', values (42, values (3.5 all reproduced identically. Not literal- type-specific.

Real root cause: the ambient-hint ladder's bottom rung (ambient_hint_core_in_mode, src/input_render.rs) parsed schemalessly (parse_command_in_mode) while every higher rung was already schema-aware. Type-blind, the grammar closes an insert … values (…) tuple after one value, so the "Next:" hint pointed at ) when the schema-aware walk (knowing the remaining columns) expects ,. The same schemaless parse also reported wrong-arity closed tuples as complete → the hint read "submit with Enter" for an input the real parse rejects.

Fix: the fallback rung now calls parse_command_with_schema_in_mode(input, cache, mode). One-line change (the cache was already in scope). Verified the friendly arity diagnostic is not masked — in advanced mode it fires at a higher rung and still wins (guard test advanced_mode_wrong_arity_insert_keeps_friendly_diagnostic_over_fallback).

Three typing_surface snapshots re-baselined (form_a one-value Next: ),; form_b too-few and form_c wrong-count: the misleading "Submit with Enter" → the accurate parse error). ADR-0022 Amendment 3 records the schema-aware fallback.

#18 — completion-drop schemaless gate (commit 7cccf4e, closed benign-but-hardened)

The ticket flagged completion.rs:336's input_parses_complete gate using the schemaless parse (sibling of #2). Investigated as the ticket asked: it's benign with the current grammar. The gate drives exactly one thing — dropping a candidate that exactly equals the trailing partial (don't re-suggest pk at the end of create table T with pk). The schemaless/schema-aware parse does diverge (e.g. insert into T values (1, 1, 'x') — int into a bool column), but the divergence is confined to value type/arity, while the drop only ever removes a re-offered keyword == partial — and those positions never coincide. Probed every plausible co-occurrence; candidate output was identical with and without the fix.

Disposition (user chose "keep as hardening"): applied the one-line schema-aware fix anyway (matches #2, strictly safe — can only ever retain a candidate the schemaless gate would have dropped). No behavioural test possible (no observable bug); the suite staying green is the regression cover. Rationale recorded in a code comment.

#17 — simple-mode arity diagnostic parity (commit 10e5197)

The substantial one. Simple-mode wrong-count inserts showed a bare expected ,/)`` while advanced mode showed a friendly "N column(s) … M value(s)" message. Root cause: simple insert and advanced insert are different grammar nodes (ADR-0033 Am1). Advanced routes a wrong-count tuple to a type-blind fallback (tuple_value_list) so it structurally matches and dml_insert_arity_diagnostics fires; the simple DSL value list (column_value_list) built a fixed-length typed Seq, so a wrong-count tuple Mismatched and the diagnostic never ran.

Fix — unified onto the one ADR-0027 model (structural parse + ERROR diagnostic), advanced left byte-for-byte unchanged:

  1. Grammar: new dsl_insert_value_list gate (data.rs) routes a wrong-count DSL insert tuple to the shared FALLBACK_VALUE_LIST so it matches and the diagnostic fires. Gated to Mode::Simple — in advanced the DSL node stays strict (otherwise a non-SQL shape like Form C insert into T (1,2) would spuriously match and be accepted in advanced; sql_insert.rs owns advanced inserts). count_tuple_values (moved to shared.rs) and a new insert_target_columns (extracted from column_value_list) are now shared by both grammars so the gate and the typed slots never disagree.
  2. Diagnostic is mode-aware: dml_insert_arity_diagnostics gained a mode param. Advanced Form B expects all columns (auto-fills nothing); simple Form B/C expects the user-fillable columns (serial/shortid auto-fill, ADR-0018 §3). It now also counts the DSL Form A role (insert_first_item, distinct from SQL insert_column) and scans the keyword-less Form C tuple. New catalog keys diagnostic.insert_arity_mismatch_form_b_simple (names the fillable and the auto-generated columns) and diagnostic.insert_arity_mismatch_all_auto.
  3. Submit safety (the regression that A surfaced): a wrong-count DSL insert now parses Ok + carries the ERROR diagnostic, so it would otherwise have dispatched and executed. A unified Ok-arm pre-flight (dsl_insert_count_mismatch_notes, input_render.rs) now blocks ExecuteDsl and shows the teaching note — mirroring the advanced Ok-arm pre-flight. The issue #1 Err-arm note retires. advanced_alternative_note's gate now reads the validity verdict (not just DefiniteErrorAt) so the cross-mode pointer still fires for the new parse-Ok-with-ERROR shape, but only when the line is genuinely valid in advanced (Form B 3-val → pointer; Form C / too-few → no pointer).

Scope guard (important for future readers): this is arity- diagnostic UX parity only. It does not consolidate value-handling, execution, or serial/shortid auto-fill across modes — ADR-0036's deliberate mode-distinctness stands. The per-mode count difference is a consequence of the auto-fill difference, not a violation of it.

Documented as ADR-0036 Amendment 2 (cross-ref ADR-0033 §8.1) + README index + requirements.md H1a.

§4. ADR / docs work

  • ADR-0022 Amendment 3 (in fa5d0dc): the ambient-hint fallback rung is now schema-aware. Records that Amendment 1's listing of parse_command_in_mode as the fallback was an oversight (every other rung was schema-aware), and that the friendly arity diagnostic still wins at its higher rung. README index updated same-edit.
  • ADR-0036 Amendment 2 (in 10e5197): simple-mode arity-diagnostic parity (full detail in §3 #17). README index updated same-edit.
  • requirements.md H1a gained citations for both #2 and #17 contributions.
  • #18 is a code-comment-only decision (benign hardening); no ADR.

§5. What's open

Bug reports still open (filed handoff-50)

# Title Label
7 Advanced mode: explain not yet supported enhancement
8 Advanced-mode syntax highlighting: identifiers and type keywords share the teal colour bug
9 [ok] explain <Table> is terse; reconsider the [ok]/SQL-echo duplication bug
10 Output panel: [error] tag colour identical to [system] (ADR-0037 gap) enhancement
11 Copy output panel contents to the system clipboard enhancement
12 Long input hints overflow horizontally bug
13 Undo confirmation dialog: too narrow + language polish bug
14 --resume should restore the last-used input mode enhancement
15 Tab completion: offer common SQL function names in expression positions enhancement
16 Restore typing-time column-typo hint for SQL expressions via known-function list enhancement

Closed this session: #2, #17, #18. Spawned + closed within the session: none new (#17/#18 were spawned in handoff-50 from #6's /runda pass).

Categorisation (carried from handoff-50, still current)

  • Bounded code fixes: #8 (identifier vs type colour), #13 (undo dialog).
  • Spec-heavy (discuss first): #9 ([ok] explain redundancy), #10 ([error] tag colour — needs ADR-0037 amendment), #14 (--resume restore mode — ties to ADR-0015 Iter 6).
  • Substantial scope: #7 (advanced explain — feature gap; ADR-0039 design agreed, impl deferred), #11 (clipboard copy), #12 (long-hint overflow — design call: wrap vs cap vs resize). Note: #12 is adjacent to this session's hint work — the field-value hints it cites are exactly the arity / per-column prose touched in #2 and #17, which are now longer (the #17 messages name multiple columns). #12 may be more visible now; worth checking the hint-panel width handling when picked up.
  • Function-list cluster: #15 + #16 — do together (shared curated SQL-function-list infrastructure; both spawned from #6).

Other tracks (unchanged, from requirements.md)

  • Track 2 project storage Iter 5/6 — export/import + --resume + persistent input history + migration framework.
  • C3a (modify relationship), C4 (m:n convenience).
  • H1 friendly DB-error layer (partial); H1a syntax-help (this session added the #2 + #17 contributions).
  • Tutorial / lesson system (needs its own ADR).
  • V4 session log + Markdown export; I1/I1b multi-line + readline; I3 tab-completion polish; I4 syntax highlighting beyond input echo.
  • ADR-0039 "explain over advanced SQL" — design agreed, impl deferred (this is issue #7).

§6. Process pins (carried forward, with this session's additions)

  • NEW — reproduce-first. Per the user's §2.1 steer: the first step of every bug pickup is to confirm it still reproduces on current HEAD before designing a fix. Several handoff-50 reports were filed in one batch; earlier fixes may have changed or eliminated others.
  • NEW — no easy-way-out on architectural discrepancies. Per §2.2: when implementation reveals a fix touches a past discrepancy (a workaround papering over a deeper gap), prefer the unifying fix that removes the discrepancy over a lighter patch that deepens it — the project is still in its original dev phase and a strong basis is worth the extra work. Escalate the trade-off; don't silently take the lighter path.
  • Confirm every commit (propose message, wait). No AI attribution. No issue numbers in commit messages (issue bodies
    • close comments may reference freely).
  • Test-first; green + clippy-clean is the only acceptable end state; current baseline 2051 / 0 / 1.
  • Follow CLAUDE.md solo-mode phases explicitly. /runda is the standard DA tool between/after phases on non-trivial work — this session ran it over the #2 fix and the #17 design (the design pass caught the DSL-Form-A-role mismatch and surfaced the message-wording fork before any code was written). The second-round DA discipline (the /runda re-prompt) caught real issues each time.
  • cargo-insta is NOT installed as a subcommand. Accept pending snapshots by mv -f <file>.snap.new <file>.snap (review the diff first — these are behaviour changes to verify, not rubber-stamp).
  • GitHub issue tracking (gh CLI) is in use; migrating to Gitea (tea) later.

§7. How to take over

  1. Read, in order: this file → requirements.md for open tracks → docs/adr/README.md for the ADR index → the relevant ADR for whatever you pick up. For insert/value-list work, ADR-0036 (esp. Amendment 2) and ADR-0033 §8.1 are now the authority on the arity-diagnostic model — both modes share it.
  2. Baseline: cargo test (2051 / 0 / 1) + cargo clippy --all-targets (clean).
  3. For a bug fix: reproduce it first on current HEAD (process pin above). The panic!() probe-test technique (write a throwaway test that prints the actual behaviour, run, observe, delete) was the fast path all session — see how #2 and #17 were investigated. Then failing test → fix → green → /runda → commit-with-approval → close-with-summary.
  4. Pick the next ticket from §5's categorisation. #8 or #13 are the bounded quick wins. #15 + #16 should be done together. If picking a hint-related ticket (#12), note this session lengthened the field-value hint strings.
  5. Architectural note for future insert work: the insert value-list path now has a clean shared spine — count_tuple_values and insert_target_columns in grammar::shared, consumed by both the DSL gate (data.rs::dsl_insert_value_list, simple-only) and the SQL gate (sql_insert.rs::tuple_value_list). The arity diagnostic (dml_insert_arity_diagnostics) is mode-aware. Keep that symmetry if you extend either side.