Story 19.4 — BI-tool emitted-SQL capture: CapturedSqlProbe (layer-2 parser verdicts for the R2a MINIMUM set)

[fupelaqu]

Story: _bmad-output/implementation-artifacts/19-4-bi-emitted-sql-capture.md (Epic 19 — R2a function-library audit, Layer 1).

Closes #254

What this PR adds (one tracked source, by design)

sql/src/test/scala/app/softnetwork/elastic/sql/census/CapturedSqlProbe.scala — layer 2 of the BI emitted-SQL extraction: a main-bearing Test/runMain object, deliberately NOT a ScalaTest suite (its input is a gitignored capture artefact that does not exist on a clean checkout — a suite would either pass vacuously via assume or leave sql/test permanently red). It reads the corpus CSV (RFC-4180, explicit UTF-8), runs the real Parser on every captured statement (Throwable-guarded — #250 can throw instead of Left), and emits per-row verdicts, debound re-parses for ?-bearing rows, census-id resolution (form-level where a measured discriminator decides; ALL candidates of the token otherwise — the inflating, safe direction), and an unmatched_spellings scan backed by the real SQLKeywords registry. A resolver/splitter/codec self-check runs before every probe.

Verified: cross-compiles on 2.12.20 + 2.13.16; sbt "sql/test" stays 532/532 and does not collect the probe; scalafmt clean; bare headerCheck unaffected (Apache-2.0 header copied from SQLKeywordsSpec).

Investigation results (artefacts are local-only under gitignored _bmad-output/, per the epic's design)

• Corpus: 99 captured statements — Superset 6.0.1 over Arrow Flight SQL (agent-driven) + Tableau Desktop 2026.2.2 Professional-trial over JDBC (lead-driven, dialect=MySQL W1–W8 and Generic SQL-92 W2/3/4/5/8). Every row round-trips byte-for-byte to its raw seam line.
• MINIMUM function set: 49 census rows (17 form-confidence / 32 token-only), joined on spelling ∪ aliases, keyed on 19.1's census id (Authoritative source-derived dialect census with anti-drift guards (R2a function-library audit) #248).
• OQ-4: the Tableau dialect dropdown moves the MINIMUM set by 25 census ids (MySQL pass matches 41 ids, SQL-92 pass 30, intersection 23).
• All 89 Tableau statements are parse errors — both dialects fully qualify + quote every identifier (backticks / "schema"."table"), see Identifier quoting coverage: backticks rejected in every position; double-quoted table names unparseable in FROM #252; SELECT 1 health checks abort flows, see Support FROM-less SELECT of literal expressions — SELECT 1 is the connection/health idiom of Tableau and Superset and is rejected #251.

Findings — 10 product issues filed during capture (lead-approved) + 1 amendment

Repo	Issues
softclient4es-jdbc	#32 (context-classloader bootstrap — driver unusable under Tableau), #33 (log4j-core without log4j-api in es8/es9 fat jars), #34 (JDBC namespace dead-ends BI browse silently; amended: catalog = constant elasticsearch, schema = cluster name — homogeneous with story 1.2's pinned CatalogName), #35 (misleading Error parsing schema DDL statement: prefix)
softclient4es-arrow	#166 (ADBC shares the classloader defect), #167 (sidecar image + ADBC jars share the log4j hole), #168 (JOIN schema probe dies on date columns), #169 (image reads ELASTIC_HOST, docs say ES_HOST)
SoftClient4ES	#251 (FROM-less SELECT — both tools' connection idiom), #252 (identifier-quoting coverage, 10-shape acceptance table), #253 (aggregate-free GROUP BY returns per-document rows)

🤖 Generated with Claude Code