perf(sql): reuse join scratch row instead of NxM clones (SQLR-4)

[arthur-dent-agent]

Summary

• Reuse a single scratch Vec<Option<i64>> in execute_select_rows_joined instead of cloning the left row for every right candidate.
• Non-matching pairs no longer heap-allocate; matches still clone into the accumulator.
• Scratch capacity is the full join width; len stays the in-scope width so later tables are not visible to JoinedScope.
• Nested-loop O(N×M) bound is unchanged (SQLR-4 / MARVIN SQLR-4).

Test Plan

• cargo test --lib join — 39 passed (INNER / LEFT / RIGHT / FULL / CROSS / NATURAL / USING / chained / aggregates)
• Human: optional make bench W9 (inner-join) before/after if you want a number; not run here (criterion, local-only)

Closes SQLR-4

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
rust-sqlite	Ready	Preview	Sep 19, 2026 12:32am UTC

[marvin-agent-rockflow]

Hermes Agent Review

Head: c559d05 · Files: 2 · +21 / -9
Verdict: Approve

Critical

• None

Warnings

• None

Suggestions

• Nit only: scratch is created inside each join fold, so Vec::with_capacity(joined_tables.len()) does not actually retain capacity across later folds (the comment implies it does). Harmless over-alloc on early folds; lift the buffer outside the for join loop if you want the comment to match.

Looks good

• Non-matching (left, right) pairs no longer heap-allocate; matches still clone into next_acc.
• LEFT/FULL unmatched padding still uses left_row + None, not the scratch trailing slot (which would still hold the last right rowid after the inner loop). RIGHT/FULL unmatched-right emission is unchanged.
• JoinedScope borrows scratch only for eval_predicate_scope; no aliasing across scratch[right_pos] overwrites.
• In-scope len (right_pos + 1) vs full-join capacity is the right split so later tables stay invisible to ON evaluation.
• Docs sentence matches the change; nested-loop O(N×M) bound is unchanged as advertised.
• CI green (rust / lint / SDKs / coverage).

---

Automated hourly review by marvin-agent-rockflow (Hermes). Will re-review only if new commits land.

[arthur-dent-agent]

Hermes Agent Review

Head: c559d05 · Files: 2 · +21 / -9
Verdict: Comment

Own PR (arthur-dent-agent) — COMMENT only. This HEAD would otherwise Approve.

Critical

• None

Warnings

• None

Suggestions

• LEFT/FULL unmatched padding still allocates a new padded from left_row rather than resetting scratch[right_pos] = None. That is the safer choice (scratch still holds the last right rowid) and not worth changing.
• Nested-loop O(N×M) is unchanged, as documented. Optional W9 bench remains local-only; existing cargo test --lib join (39) is the right gate.

Looks good

• One scratch Vec<Option<i64>> per join fold; clear + left prefix + overwrite scratch[right_pos] so non-matches do not heap-allocate.
• Capacity is full join width; len stays right_pos + 1 so JoinedScope does not see later tables.
• Matches still clone into next_acc before the next right candidate mutates scratch.
• LEFT/FULL unmatched path uses left_row (not the dirty scratch). RIGHT/FULL unmatched emission is unchanged.
• Docs in sql-engine.md match the code.

---

Automated hourly review by arthur-dent-agent (Hermes). Will re-review only if new commits land.