feat(tpcds): picodata schema + data-load (query port pending)

[Cianidos]

Summary

Adds Picodata as a load target for TPC-DS. Schema + 24-table bulk load verified end-to-end. The 99-query suite is out of scope for this PR — sbroad (picodata's SQL engine) cannot parse it yet (blocker documented below with examples). pico.sql ships as the mechanical-transform skeleton and the port starting point.

Also ships a cross-driver fix found along the way: the bulk-insert bind-parameter clamp moved centrally so Picodata (and a latent YDB path) stop overflowing on wide tables.

What works (picodata)

./build/stroppy run tpcds/tpcds -d pico \
    -D url=postgres://admin:T0psecret@localhost:1336/admin \
    -e SCALE_FACTOR=1 --no-steps workload

• drop_schema / create_schema / create_indexes: all 24 tables + indexes, green.
• load_data: 0/24 table errors, all 24 tables populated.
• Schema (schema.pico.sql): char(N)→varchar(N), date→datetime, a PRIMARY KEY per Tarantool space (engine-required), no FK. Plain DROP TABLE IF EXISTS (picodata has no CASCADE). ANALYZE skipped (sbroad rejects it).
• tpcds.ts: picodata wired into the dialect maps; drop/analyze branches handled.

What does NOT work — query execution (documented blocker)

sbroad (picodata 26.3) accepts only explicit INNER JOIN ... ON. It has no comma join, no CROSS JOIN, no PRAGMA (so no YDB-style AnsiImplicitCrossJoin), and no session setting for implicit cross joins. TPC-DS query templates use comma joins pervasively, so ~88 of 95 queries fail at parse time.

Concrete examples, run against a live picodata 26.3 container:

SELECT count(*) FROM store_sales, date_dim WHERE ss_sold_date_sk = d_date_sk;
-- ERROR: sbroad: rule parsing error: --> 1:33
--   expected EOI, IndexedByExpr, or DqlOption

SELECT count(*) FROM store_sales CROSS JOIN date_dim WHERE ss_sold_date_sk = d_date_sk;
-- ERROR: sbroad: rule parsing error: --> 1:54
--   expected Identifier

PRAGMA AnsiImplicitCrossJoin;
-- ERROR: sbroad: rule parsing error: --> 1:1
--   expected EOI, ... SetParam, SetTransaction, ...

Dominant query failures (all the parser hitting , in a FROM list): expected EOI ×47, expected OrderBy ×21, expected IndexedByExpr ×21.

Secondary gaps behind the comma-join wall: rollup/grouping sets (7 q), INTERSECT (4 q; EXCEPT works), FULL OUTER JOIN (2 q), rank()/dense_rank()/lag()/lead() (8 q, engine-missing — already skipped+logged in tpcds.ts), correlated subqueries, datetime+int, round(). pico.sql's header lists them.

pico.sql is kept (not stubbed) as the port starting point: date/char/cast fixes and date+N offsets are already baked. Run the workload step only after that port lands; today use --no-steps workload.

Included fix — central batch clamp (independent value)

The 65535 bound-parameter cap (pgwire extended protocol = MySQL prepared Error 1390) was enforced only in the MySQL driver (capBatchByColumns). Moved it into the shared sqldriver.RunBulkInsert so every sql.DB-backed dialect is protected:

• picodata: date_dim (28 cols), catalog_sales/web_sales (34 cols) previously aborted with extended protocol limited to 65535 parameters and loaded 0 rows. Fixed.
• ydb plain_bulk: was unclamped (latent; tpcds uses native BulkUpsert so never triggered). Now protected.
• mysql: behavior identical (central clamp subsumes the per-driver helper, which is removed).

One constant (maxBoundParameters = 65535), one place. Net: less code, wider coverage.

Verification

• make linter_fix → 0 issues.
• make tests → all packages ok (race + coverage).
• Live picodata 26.3 docker, TPC-DS load 0/24 table errors at both SCALE_FACTOR=0.01 and SCALE_FACTOR=1 (SF=1 → 20.45M rows; all 24 tables populated, fact tables at correct SF1 magnitude: inventory 11.7M, store_sales 2.88M, catalog_sales 1.44M, web_sales 719K, customer_demographics 1.92M).

Follow-ups (not in this PR)

• TPC-DS picodata query port — separate epic. Preferred path: AST comma-join rewrite (parse → lift join predicates from WHERE into ON → nested INNER JOIN → deparse), then targeted rollup/intersect/full-outer rewrites; or upstream sbroad implicit-cross-join support.
• File picodata upstream issue for comma-join / AnsiImplicitCrossJoin parity with YDB (would also fix tpch-pico queries, which were never actually validated — TestTpchLoadOnPicodata only asserts queries are logged, not that they pass).

---

Refs #86 (partial: schema + load + wiring landed; query dialect still open — blocked on sbroad comma-join gap).