feat(validation): rewrite the suite's dialect-sensitive SQL via RewriteSql — 30 more cases gate-enforced

[fornwall]

What

Now that the DriverQuirks::RewriteSql(query_id, default_sql) quirk has landed upstream (apache/arrow-adbc#4496), this uses it to enable every validation case whose only blocker was hardcoded non-GoogleSQL SQL.

Empirical triage first (full suite run against the emulator at the old pin): of the 42 EXCLUDED cases, the dominant failures were (1) the emulator rejecting NULLS FIRST/NULLS LAST, (2) SELECT * ingest readbacks surfacing the synthetic adbc_ingest_key column, (3) non-Spanner CREATE TABLE DDL (INT/INTEGER/TEXT, no PRIMARY KEY), (4) INSERT without the column list GoogleSQL requires, and (5) SELECT @p0, @p1 defeating Spanner's parameter-type inference. All five are pure SQL problems.

Upstream side — apache/arrow-adbc#4514 (merged)

Upstream previously routed only the two CAST(1.5 AS FLOAT) queries through RewriteSql. apache/arrow-adbc#4514 routes the remaining hardcoded statement-test SQL through the hook under stable query ids; the TestSqlIngestType id carries the ingested Arrow type as a suffix so quirks can rewrite per type. Default behavior is unchanged for drivers that don't override the hook.

This PR was originally developed against a fork while #4514 was in review. #4514 is now merged, so ARROW_ADBC_TAG points back at an apache/arrow-adbc main revision and GIT_REPOSITORY is upstream again — no fork anywhere in the build.

Driver-repo side (this PR)

• ARROW_ADBC_TAG → apache/arrow-adbc main 1ba7fe66 (contains #4514); GIT_REPOSITORY → apache/arrow-adbc.
• SpannerQuirks::RewriteSql grows into a table of GoogleSQL substitutions (each pinning the upstream default so a tag bump that changes a query fails loudly): Spanner-valid DDL, column-listed INSERTs, readbacks that select the ingested columns explicitly and drop NULLS FIRST/LAST (GoogleSQL's ASC/DESC defaults match exactly), ORDER BY col[SAFE_OFFSET(0)] for the list cases (GoogleSQL can't ORDER BY an ARRAY), and CASTs for the bare-parameter select.
• IngestSelectRoundTripType declares Spanner's readback widenings (small ints → INT64, large/view strings → STRING, binary variants → BYTES).
• ValidateIngestedTemporalData (fixture override; the base class FAILs) checks timestamp readbacks: any-unit Arrow timestamps → Spanner TIMESTAMP → Timestamp(Nanosecond, "UTC").
• PrimaryKeyIngestTableDdl → nullopt: SqlIngestPrimaryKey append-ingests rows omitting the PK and expects auto-assigned ascending keys — Spanner has no ordered auto-increment (and its SEQUENCEs are bit-reversed, so even a DEFAULT key can't satisfy the insertion-order readback). The quirk-sanctioned self-skip states that truth.

Result

EXCLUDED shrinks 42 → 11; the suite is now 82 pass + 7 self-skip + 11 excluded = 100 cases, all three checks green locally (gate, expected-failure guard, stale guard).

Newly gate-enforced (30): the whole ingest type family — bool/ints/floats/string/binary/date32/timestamps including large/view layouts, dictionary-encoded strings and both List cases — plus SqlIngestAppend/Replace/CreateAppend/MultipleConnections/Sample, SqlBind, SqlQueryEmpty, SqlQueryInsertRollback, SqlQueryRowsAffectedDelete{,Stream}, and SqlPrepareSelectParams.

Still excluded, none fixable by SQL rewriting:

Bucket	Cases	Why
Insertion-order readbacks	SqlPrepareUpdate{,Stream}	assert rows come back in insertion order; Spanner returns PK (random-UUID) order
Unmapped Arrow types	SqlIngestUInt8/16/32/64, Duration, Interval, FixedSizeBinary	"cannot create a Spanner column for Arrow type …" (UInt8/16/32 could widen losslessly to INT64 — future driver work)
Empty-stream ingest	TestSqlIngestStreamZeroArrays	driver rejects an ingest with no bound data
ECANCELED errno	SqlQueryCancel	arrow-rs FFI stream exporter can't emit errno 125

Testing

• Rebased on main (5537486, v0.7.0) — clean, no conflicts.
• Repoint verified to be behavior-neutral: all 23 query ids the Spanner quirks handle exist upstream (exact match), and the RewriteSql call sites — ids and pinned default SQL — are identical between the old fork rev and upstream 1ba7fe66; the only file diff is clang-format reflow. Upstream's suite differs from the fork's base by #4514 alone (the sole c/validation change in the 11 commits since), and all 81 StatementTest methods are unchanged, so the EXCLUDED list and the stale/expected-failure guards stay valid.
• shellcheck scripts/run-adbc-validation.sh clean.
• No Rust changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TXzBG1e8nCAUsTarm8sruY