linq_fold: Theme 8 — three specialized fusion arms (closes audit)

[borisbat]

Summary

Closes Theme 8 of benchmarks/sql/linq_fold_chain_audit.md and with it the entire audit (all 8 themes shipped between 2026-05-24 and 2026-05-25 — PRs #2851, #2852, #2857, #2861, #2862, #2865, #2866, #2874, and this one).

Three small splice arms, bundled because they're independent and each is self-contained. Audit explicitly called these "low priority / cheap follow-ups."

2a — each(arr).reverse()._distinct[_by](K).to_array() (plan_reverse)

Array source only. New emission walks source backward via index (arr[len-1-k]), maintains var rev_dset : table<...> and gates push by set-insert on the dedup key. Saves cascade's reverse_to_array allocation AND the distinct_by_inplace second pass. LAST-per-key semantics preserved: backward walk picks first-seen-in-reversed-order = last-in-source occurrence, matching tier-2 reverse.distinct_by. v1 scope tight: array source required (non-array sources can't be backward-indexed); pre-reverse where_/select/take all bail to cascade.

3b — each(arr)._distinct[_by](K1)._order_by[_descending](K2).to_array() without take (plan_order_family)

The bail (distinctName != "" && takeExpr == null) relaxes to permit when firstName == "" (first/first_or_default + distinct still cascades — streaming-min path has no dset hook, deferred). The where_+order fused-loop path generalizes: when distinctName != "", declares var order_dset : table<typedecl(...)> and wraps pushExpr with a set-gated if (!key_exists(...)) block (mirrors Theme 3 Phase 3's bounded-heap distinct gate). Composes with WHERE (filter→distinct→sort) and Theme 1 terminal _select (project at return). Saves cascade's distinct_by_to_array intermediate iterator setup.

C4 — each(a) |> zip(each(b)) [|> chain_ops] |> reverse() |> <terminator> (plan_zip)

Trailing reverse as the last chain op between zip's chain and the terminator. Array lane emits _::reverse_inplace(bufName) before return; counter / accumulator (sum/min/max/avg) / any/all/contains lanes treat reverse as a no-op (mathematical identity). Bails on first / first_or_default (NOT identity under reverse). Constraint: reverse must be the last chain op — anything after would see the reversed stream and change semantics vs cascade.

What this changes

File	Change
daslib/linq_fold.das	+124 / -9 — three planner extensions
tests/linq/test_linq_fold_theme8_fusion_arms.das	+394 — 16 tests / 16 sub-runs (positive + parity + anti-tests)
benchmarks/sql/linq_fold_chain_audit.md	Theme 8 row + Status entry; classification flips on probes 2a / 3b / C4 from FALLS-OFF → SPLICE-FIRES; audit declared fully closed
doc/source/reference/linq_fold_patterns.rst	+3 rows (2a in plan_reverse, 3b in plan_order_family, C4 in zip section)

Notes

• No bench refresh. No existing bench in benchmarks/sql/ covers the 2a / 3b / C4 chain shapes (those were tier-2 cascade until this PR). Per the living-doc policy for results.md, planner-touching PRs only refresh when an existing bench exercises the shape — none does here. Follow-up TODO: add 3 benches for these shapes in a future PR.
• Audit fully closed. This is the last theme. Future linq_fold work should come from new user-observed regressions or new probe shapes, not the existing audit.

Test plan

• mcp__daslang__compile_check daslib/linq_fold.das — clean
• mcp__daslang__lint daslib/linq_fold.das — no issues
• mcp__daslang__lint tests/linq/test_linq_fold_theme8_fusion_arms.das — no issues
• mcp__daslang__format_file tests/linq/test_linq_fold_theme8_fusion_arms.das — formatted
• New Theme 8 test file: 36/36 pass INTERP (including parity tests vs handwritten cascades and 4 anti-tests for out-of-scope shapes)
• All linq_fold tests pass (1144 across test_linq_fold*.das)
• All linq-side tests pass (test_linq_sorting, test_linq_set, test_linq_querying, test_linq_join, test_linq_from_decs, test_linq_aggregation, test_linq_group_by, test_linq_transform, test_linq_bugs, test_linq_concat, test_linq_element, test_linq_generation, test_linq_partition)
• Full tests/linq + tests/decs matrix via CI (local env has pre-existing dasbind.das missing in modules/dasLLVM/bindings/ blocking dastest.exe — unrelated to this PR; verified by reproducing on clean master)
• Full AOT + JIT matrix via CI

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR closes Theme 8 of the linq_fold chain audit by extending _fold’s tier-1 planners with three additional specialized fusion arms, reducing fallbacks to the tier-2 cascade for a few previously uncovered chain shapes.

Changes:

• Extend plan_reverse to splice each(arr).reverse()._distinct[_by](K).to_array() for array sources via a backward index walk and dedup set gate.
• Extend plan_order_family to splice _distinct[_by](K1)._order_by[_descending](K2).to_array() (no take) by adding a dedup set gate to the fused-loop + in-place sort path.
• Extend plan_zip to accept a trailing reverse() as the last chain op before the terminator, emitting reverse_inplace only for the array lane and treating it as identity for count/accumulator/any/all/contains lanes.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
daslib/linq_fold.das	Adds three new splice surfaces across plan_reverse, plan_order_family, and plan_zip to avoid tier-2 cascade allocations/passes for Theme 8 chain shapes.
tests/linq/test_linq_fold_theme8_fusion_arms.das	Adds Theme 8 regression/parity tests plus anti-tests for out-of-scope shapes.
benchmarks/sql/linq_fold_chain_audit.md	Marks Theme 8 probes as splicing and declares the audit fully closed.
doc/source/reference/linq_fold_patterns.rst	Documents the newly supported Theme 8 splice patterns for reverse/distinct, distinct/order (no take), and zip+reverse.

---

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

[Copilot]

Pull request overview

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