perf(runtime): add typed columnar aggregate reducers

[ashione]

Summary

• add state-columnar typed reducer paths for single-INT64 and two-INT64 SUM/COUNT/AVG aggregate shapes
• extend source pushdown shape classification to predicate single-key COUNT/SUM/AVG and move CSV predicate aggregate execution onto typed reducer paths
• add Phase 1 internal columnar execution substrate skeleton: ColumnarExecBatch, ColumnarExecColumn, and ColumnarExecView
• preserve Arrow-backed lazy input inside the execution batch and make row materialization an explicit boundary
• add a local columnar kernel benchmark gate covering aggregate typed-shape selection, file-source pushdown ratios, and string builtin plan reuse
• document measured proof points and rejected mixed-key attempts in the plan/design docs
• address review-found edge cases: safe int64 dense-domain estimation, predicate AVG empty-numeric semantics, and execution-batch backing row-count validation

Performance Evidence

• single-int64 SUM: 52.0 ms -> 38.5 ms (~1.35x)
• single-int64 COUNT: 41.0 ms -> 28.0 ms (~1.46x)
• single-int64 AVG: 53.5 ms -> 38.0 ms (~1.41x)
• two-int64 SUM: 163.0 ms -> 121.5 ms (~1.34x)
• two-int64 COUNT: 159.0 ms -> 108.0 ms (~1.47x)
• two-int64 AVG: 166.0 ms -> 122.5 ms (~1.36x)
• CSV mixed predicate source count: 267,744 us -> 235,904 us (~1.13x)
• rejected mixed string/int64 dictionary-id reducer attempt after mixed-string-int64 regressed to 634 ms

Validation

• bazel test //:planner_v03_test //:file_source_test //:columnar_batch_test --test_output=errors
• bazel test //:core_regression --test_output=errors
• bazel test //:experimental_regression --test_output=errors
• bazel build //:sql_demo //:df_demo //:stream_demo
• bazel run //:actor_rpc_smoke
• ./scripts/run_columnar_kernel_benchmark_gate.sh
• bazel run //:batch_aggregate_benchmark -- 1048576 5
• bazel run //:file_source_benchmark -- 200000 3
• bazel run //:string_builtin_benchmark -- 100000 5
• git diff --check
• targeted secret-pattern scan over changed files

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d99a12df4a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Advance SIMD tail after bounded vector hit

On AVX2 builds, if a bounded selection (for example WHERE ... LIMIT n via combinedFilterSelection(..., max_selected)) reaches max_selected inside this vector loop, goto scalar_tail leaves i at the start of the same vector block. The scalar tail then reprocesses lanes that were already emitted and appends duplicate indices (e.g. LIMIT 1 can produce [0,0]), so filterTable can duplicate rows and exceed the requested limit; return immediately or advance the tail past the processed lane/block.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Keep Float32 values serializable

When a SQL FLOAT/REAL table receives rows, this rewrites cells to the new DataType::Float32, but the existing table codecs were not updated for that type: BinaryRowBatchCodec falls through and deserializes it as a string, ProtoLikeSerializer decodes type 6 as null, and nanoarrow materialization infers string then calls asString(). Any inserted float column crossing RPC/plan serialization/export boundaries is therefore corrupted or throws, so either handle Float32 in those codecs or avoid emitting it here.

Useful? React with 👍 / 👎.