[milindsrivastava1997]

Part 2 of #1046; part 1 was #1069. With this PR, #1046 is done except for the cross-crate test-only items, which moved to #1057.

Why

• The row-at-a-time reference operators exist only to check that the columnar operators match them. That check was there for the columnar conversion, which is finished. The oracle never says what an operator should return; it only compares against the old code.
• Four devtools tools each kept their own copy of the corpus helpers, and the bgp_catalog copies had drifted apart.

What

• executor:
  • Removes operators/row_reference/ (11 files), operators/aggregate/row_reference.rs, and the #[cfg(test)] hook in Operator::start that switched to them.
  • Removes the oracle parity tests (*_matches_the_row_reference). As decided on Remove dead and test-only code: row-reference oracle, old_docs, test-only demand model, batch bridge, memo snapshot #1046, they are not replaced with fixed expected-row tests.
  • Removes temporal::reduce, the test-only row reducer only the oracle used.
  • Removes five #[cfg(test)] helpers that only the oracle called: collect_rows, row_bytes, key_bytes, compare_rows, group_key.
• devtools:
  • Adds a new asap_devtools::corpora module holding sql_statements, promql_lines, dqc_catalog, netflow_catalog, bgp_catalog and parse_sample_interval.
  • sketch_coverage, variant_coverage, analyze_corpora and examples/canonical_examples use the module instead of their own copies.
  • sketch_coverage stays a separate binary; it is not merged into variant_coverage.

How

bgp_catalog is the union of the four old copies:

	sketch_/variant_coverage	analyze_corpora	canonical_examples	corpora::bgp_catalog
bgp_updates columns	7	16	3	16
bgp_rib_state	5 columns	missing	3 columns	5 columns
bgp.-prefixed names	both	updates only	none	both

bgp_analytics.sql reads bgp.bgp_rib_state and origin, and the jan2024 workload reads all 16 update columns. None of the old copies covered all of these.

Key code interfaces added / modified

// crates/devtools/src/corpora.rs (new module, pub in asap-devtools)
/// Statements of a SQL corpus: `#`/`--` comment lines dropped, split on `;`.
pub fn sql_statements(corpus: &str) -> Vec<String>;                 // new
/// Queries of a PromQL corpus: one per non-empty, non-`#` line.
pub fn promql_lines(corpus: &str) -> impl Iterator<Item = &str>;    // new
pub fn dqc_catalog() -> SqlCatalog;                                  // new
pub fn netflow_catalog() -> SqlCatalog;                              // new
/// Union of the former per-tool copies (table above).
pub fn bgp_catalog() -> SqlCatalog;                                  // new
/// The required `--sample-interval-ms` CLI argument.
pub fn parse_sample_interval() -> u64;                               // new

All the executor items removed here are pub(crate) or narrower and #[cfg(test)].

Major algorithm changes

None.

Before this PR

• Every converted operator ran a second time under with_row_reference in tests.
• Each corpus tool lowered BGP queries against its own catalog.

After this PR

• Operator tests run only the columnar operators.
• All corpus tools share one catalog per corpus.

Evidence

On variant_coverage, sketch_coverage (--sample-interval-ms 1000) and analyze_corpora --sql-corpora, I compared the outputs before and after this PR:

• The coverage reports are byte-identical.
• The analyze_corpora dumps are identical, apart from one DataFusion "did you mean" hint in an error message.

The union catalog changes no output today. The BGP queries that read origin or the RIB table already fail earlier, on ClickHouse functions DataFusion does not support (arrayFilter, arrayJoin, isIPAddressInRange, toIntervalMinute) or on parse errors. For bgp_analytics in variant_coverage, 5 queries lower and 10 fail, both before and after.

Verification

• Unit and end-to-end tests: not run locally. The commit was made with --no-verify, so the workspace test suite has not run on this branch. CI runs it.
• Other checks: cargo check -p asap-executor --lib --tests and cargo clippy -p asap-devtools --all-targets report no warnings, and the before/after tool comparison is above.

No new tests.

Snapshot changes

None.

Architectural decisions

• The oracle is deleted rather than frozen into expected-row tests. Freezing it would turn random 150-row outputs into golden files that nobody can review.
• bgp_catalog is the union of the old copies, not any single one of them. This was decided on Remove dead and test-only code: row-reference oracle, old_docs, test-only demand model, batch bridge, memo snapshot #1046.

Limitations and follow-up

• Apart from series_labels, the columnar operators now have no unit tests. Coverage comes from crates/executor/tests/ and the corpus/integration tests.
• Decide how to hide test-only public items used across crates #1057 covers the remaining test-only public items.

Human review — do not complete with an agent

• The MVP boundary is correct.
• New conceptual layers or public interfaces are necessary.
• The before/after description matches the intended product behavior.
• Human reviewer:
• Decision and rationale:

🤖 Generated with Claude Code