linq_fold: array _join splice → m3f join_* parity (chunk N+3)

[borisbat]

Summary

• Closes the m3f-vs-m4 parity gap on the entire join_* bench family. Mirror of PR linq_fold: plan_decs_join + LinqJoin auto-typed result lambda (Session 3) #2848 (plan_decs_join) + PR linq_fold: Theme 3 Phase 1 — cross-arm join+group_by splice (C3) #2861 (_join |> _group_by cross-arm) for arrays. Before this PR, every _fold(arr |> _join(...) ...) chain fell through to tier-2 cascade (join_impl at daslib/linq.das:1674) — 3 intermediate allocations, per-pair invoke overhead, m3f 2-3× slower than m4 across the board. Now m3f beats m4 on all 5 cells in both INTERP and JIT.
• Phase 1 — new register_array_join_rows pattern + emit_array_join produces hashB-collect + srcA-probe inline (preserves the count-no-where bucket-length fast path; mirrors PR linq_fold: plan_decs_join + LinqJoin auto-typed result lambda (Session 3) #2848's D2-A guarantee). 2-source invoke wrap (like Zip); primitive equi-key gate via existing is_primitive_join_key_type; non-primitive keys + decs-bridge srcb cascade silently.
• Phase 2 — new ArrayJoinShape struct + SourceAdapter.ArrayJoin variant. Extends 4 adapter helpers (adapter_bind_name, adapter_element_type, adapter_wrap_source_loop, adapter_wrap_invoke) + allocation prefix in plan_group_by_core. Adds upstream_join slot to group_by_array row + array_join_invariants requires predicate. plan_group_by_core is fully reused — ArrayJoin flows through the adapter abstraction the same way DecsJoin does.

Bench (full INTERP + JIT rerun, n = 100 000)

Cell	m3f INTERP before / after	m3f JIT before / after	m4 INTERP	m4 JIT
join_count	128.2 → 51.0	34.6 → 11.6	63.6	12.7
join_select	148.1 → 72.0	41.6 → 19.7	83.8	22.2
join_where_count	148.5 → 60.1	41.2 → 20.4	74.3	22.5
join_groupby_count	186.2 → 78.3	47.2 → 19.4	90.1	21.6
join_groupby_to_array	217.8 → 88.1	55.5 → 19.5	91.4	21.6

Drift across the non-join matrix is at measurement floor (±5% INTERP, ±10% JIT). The splice is gated on array_source + primitive equi-key + (for cross-arm) upstream_join capture, so non-join chains aren't touched.

Known typer-order subtlety

Bare-_select after _join on array sources (no trailing to_array()) leaves selCall._type.firstType as an unresolved typedecl(result_selector(type<TT>)) because nothing forces resolution in the chain — select's array overload returns array<...> directly. emit_array_join extracts the resolved type from the select lambda's body via peel_lambda_single_return instead of cloning selCall._type.firstType. emit_decs_join doesn't hit this because its bench chains universally end with |> to_array() (decs returns iterator-typed bridge). Test coverage includes both bare-select and iterator-chain-with-to_array forms.

Follow-up (next PR, agreed upfront with the author)

emit_array_join and emit_decs_join are ~70% structural duplicates; same for the DecsJoin / ArrayJoin branches of adapter_wrap_source_loop. Deferred to a follow-up refactor PR (extract emit_join_collect_probe_shape helper across all 4 callsites; ~80-100 LOC dedup) — same staging call PR #2848 made: ship the perf-fix at zero risk to the 4 already-closed decs cells, then refactor with both sides bench-locked.

Test plan

• mcp__daslang__format_file + mcp__daslang__lint clean on every touched .das
• mcp__daslang__run_test tests/linq/test_linq_join.das — 52/52 (10 new splice-exercising tests appended: 7 Phase 1, 4 Phase 2, 1 non-primitive-key cascade)
• mcp__daslang__run_test tests/linq/test_linq_from_decs.das — 198/198 (decs splice not regressed)
• mcp__daslang__run_test tests/linq/test_linq_fold_terminal_select.das — 28/28
• mcp__daslang__run_test tests/linq/test_linq_fold_theme3_decs_join_groupby.das — 14/14 (PR linq_fold: Theme 3 Phase 1 — cross-arm join+group_by splice (C3) #2861 surface intact)
• Bench acceptance: full INTERP+JIT rerun across benchmarks/sql/*.das; all 5 join_* m3f cells beat m4
• mcp__daslang__compile_check benchmarks/sql/join_*.das — all 5 compile under the new splice

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds an array-side _join splice in linq_fold mirroring the existing decs-side emit_decs_join, plus a cross-arm _join |> _group_by path via a new ArrayJoin SourceAdapter variant. Closes the m3f-vs-m4 parity gap across all five join_* SQL benchmarks (m3f now beats m4 in both INTERP and JIT).

Changes:

• New ArrayJoinShape + SourceAdapter.ArrayJoin variant; extends adapter_bind_name, adapter_element_type, adapter_wrap_source_loop, adapter_wrap_invoke, and plan_group_by_core's name prefix.
• New register_array_join_rows / emit_array_join (Phase 1) with array_join_srcb_is_array guard; new array_join_invariants guard added to group_by_array pattern with optional upstream_join capture (Phase 2).
• 11 new tests in tests/linq/test_linq_join.das, docs row in linq_fold_patterns.rst, refreshed bench matrix in results.md.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
daslib/linq_fold.das	Adds ArrayJoinShape/ArrayJoin adapter, array_join_* guard predicates, splice emit + registration, cross-arm group_by wiring.
tests/linq/test_linq_join.das	Adds 11 tests exercising Phase 1 splice arms, Phase 2 join+group_by, and non-primitive-key cascade.
doc/source/reference/linq_fold_patterns.rst	Documents the array-array equi-join patterns and the cross-arm group_by composition.
benchmarks/sql/results.md	Refreshes the INTERP+JIT benchmark matrix to reflect the new m3f wins on join_* rows.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.