linq: table group_by fusion — TableAdapter rides plan_group_by_core

benchmarks/sql/LINQ_TO_TABLE.md

@@ -4,10 +4,22 @@ Sibling of [LINQ.md](LINQ.md) / [LINQ_TO_DECS.md](LINQ_TO_DECS.md). Plan of reco
 `table<K;V>` / `table<K>` as the 6th `_fold` source, plus the `to_table` sink.
 Edited in-place as PRs land.
 
-Status: **stage 6 committed — arc complete** (to_table sink; stage 5 = join probe + table-lead
-joins, 2742f6db2; stage 4 = point-lookup folds, ac441c4a0; stage 3 = `%linq!` table sources,
-29d23baf6; stage 2 = TableAdapter + m7, 571fe879e; stage 1 = `each_kv` builtin, 8751bb9ba;
-master's fixed-array rework merged in after stage 5, 1ab3e6a67).
+Status: **stage 7 committed** (group_by fusion — `can_group_by` + `build_group_by_adapter` on
+`TableAdapter`, riding `plan_group_by_core` with the usage-pruned slot walk as the bucket-fill
+loop; stage 6 = to_table sink; stage 5 = join probe + table-lead joins, 2742f6db2; stage 4 =
+point-lookup folds, ac441c4a0; stage 3 = `%linq!` table sources, 29d23baf6; stage 2 =
+TableAdapter + m7, 571fe879e; stage 1 = `each_kv` builtin, 8751bb9ba; master's fixed-array
+rework merged in after stage 5, 1ab3e6a67; the JIT inline slot walk landed separately, #3100).
+
+Stage 7 findings:
+- **Two overrides were the whole change**: the group_by splice pattern was already adapter-generic
+  (`can_group_by_source` gate → `build_group_by_adapter` → `plan_group_by_core`), so enabling
+  tables = `can_group_by() == true` + a fresh-`TableAdapter` `build_group_by_adapter`. The kv
+  usage-pruner sees the whole accumulation body (key expr + reducer updates + upstream
+  where/select segments), so a group key over `kv.value.brand` walks `values(tab)` alone.
+- m7 `groupby_*` INTERP 144–201 → 30–50 ns/op (count 163→31, ~5×); JIT 44–73 → 8.4–11 (count
+  43.5→8.4, another ~5× — the fused emit rides #3100's inline slot walk). `join_groupby_*` stays
+  on the cascade (deferred edge below).
 
 Stage 6 findings:
 - **Tier-2 surface required for typing**: `_fold`'s argument must fully type before the macro
@@ -212,6 +224,41 @@ PR1 findings:
 
 End of arc: `skills/linq.md` + linq docs mention the table source.
 
+## Late stage (planned) — reducer shapes & general code hygiene
+
+Cross-source cleanups; none are table-specific. Items 1–2 are user-facing reducer-shape fixes,
+items 3–4 are codebase hygiene investigations (the linq_fold surface is workable but "a tad too
+unwieldy" — the table adapter took several stages, and many fuses read as "add this hook, because
+reasons" rather than falling out of the architecture).
+
+1. **Identity-lambda reducers**: `_._1 |> max($(v) => v)` (also `min`/`sum`/`average`) fails with
+   30303 today — the untyped lambda can't infer on the tier-2 lazy-bucket surface, and
+   `recognize_reducer_specs` has no identity arm either. Fix both ends: recognize the identity
+   inner-select and canonicalize to the bare form (`max()`), and make the tier-2 generic accept
+   it so unfused chains agree.
+2. **Untyped inner-select lambda params**: `_._1 |> select($(c) => c.value.price) |> sum()`
+   requires an explicit param type (`$(c : CarKV)`) — the lazy bucket's `select` doesn't flow the
+   element type into the lambda. Thread the type through so the annotation becomes optional;
+   today's explicit-type requirement is a usability trap (the error is an opaque 30303, not
+   "annotate the param").
+3. **match.das adoption survey (linq_fold* + sqlite_* family)**: flatten_opt's move to
+   `daslib/match` bought both fewer lines and more readable matchers; the linq spine-walkers are
+   full of the same hand-rolled `is ExprCall` / `as` / null-guard chains. Prime candidates:
+   `match_key_probe_side` / `extract_key_probe` (manual `ExprRef2Value` peeling + `ExprField`
+   checks), `extract_*_source` gates, and the sqlite_linq chain decomposition. Survey first,
+   convert where the match form is a strict readability win.
+4. **SourceAdapter interface audit**: all initially-wanted adapters now exist (11 classes:
+   Array/Zip/ArrayJoin/Decs/DecsJoin/Xml/XmlJoin/Json/JsonJoin/Table/ProjectedSource) — audit the
+   ~20-method interface against real usage and remove the scaffolding it forces. Known smells:
+   `arrayTop()`/`arraySrcName()` are still marked "transitional … removed once all consumers move
+   into subclass methods" yet remain load-bearing (reserve hint reads them; adapters override
+   them with comments explaining which distant gate reads what); the `build_group_by_adapter`
+   upstream-join arm repeats ~30 lines of keyaLam/keybLam/resultLam validation across
+   Array/Decs/Xml/Json; two overlapping reverse hooks (`emit_reverse_skip_into_tail` vs
+   `emit_reverse_last_backward`); emit fns reconstruct `headCalls` from stringly-keyed captures
+   ("mirrors emit_loop_or_count_lane_decs"). Goal: a new source should not need "several stages"
+   of hook-by-hook enablement for the standard fuse set.
+
 ## Risks / watch items
 
 - **Mangler ICE 50609** (iterator element-const collision) — `each_kv` yields `-const` non-ref
@@ -224,6 +271,10 @@ End of arc: `skills/linq.md` + linq docs mention the table source.
 
 ## Deferred edges (named, not built)
 
+- **`join |> group_by` over a table lead**: `TableAdapter.build_group_by_adapter` declines the
+  upstream-join arm (returns null → tier-2). The fix is a TableJoin analog of `ArrayJoinAdapter`
+  (lead loop from the pruned slot walk, srcB hash/probe from the stage-5 pieces); the
+  `join_groupby_*` m7 cells are the numbers it would improve. Revisit on demand.
 - **Point-lookup conjunct extraction**: `where(kv.key == X && <residual>)` (incl. the collapsed
   multi-where form) could probe and evaluate the residual on the probed element only. The matcher
   currently declines compound predicates; add when a real chain wants it.