feat(sql): quoted and qualified table names in FROM, JOIN and DELETE (story 21.2)

[fupelaqu]

Closes #252
Closes #85
Closes #292
Closes #293

Story 21.2 — the FROM/JOIN table-name surface. FROM (and JOIN, and DELETE FROM) becomes a
call-site swap onto story 21.1's exported lexing surface plus a leading-qualifier rep, and the
qualifier a statement writes is preserved in the AST instead of being discarded.

The rule

Quoting is the discriminator, and it is #57's shipped decision generalised — not replaced. A
qualifier is the maximal leading run of parts that are each quoted AND each followed by a dot;
the index name is everything after it. A bare dot never separates, which is what has always
protected logs-2025.03.

written	index read	qualifier
FROM elastic.bi_events	elastic.bi_events	—
FROM "elastic".bi_events	bi_events	elastic
FROM `elastic`.`bi_events`	bi_events	elastic
FROM "logs-2025.03"	logs-2025.03	—
FROM "elasticsearch"."prod-cluster"."bi_events"	bi_events	2 levels

The parser preserves, it does not interpret (#85). Table.name stays byte-for-byte its previous
value; Table.parts / StandardJoin.parts record the ordered NameParts. Nothing in sql, core,
bridge or any client reads parts — role assignment (schema? catalog? server alias?) belongs to
the resolver, which knows the venue.

Task 0 baseline — re-measured on the CURRENT origin/main (99d1f438), not the spec's ac54a079

The spec's baseline was four merges stale, and two of its "rejected today" rows had become live
defects, both created by 21.1:

statement	on 99d1f438	after
JOIN "prod_us".customers c	parses, JOIN leg reads index prod_us.customers while the FROM leg reads orders	both read the bare index
FROM elastic."bi_events"	parses, reads index elastic. (the quoted lexeme was stolen as the ALIAS)	elastic.bi_events
FROM bi_events `bi_events` / FROM bi_events "bi_events"	parses (21.1's alias flip)	unchanged
JOIN `customers` c / JOIN "customers" c	parses (21.1's source flip)	unchanged

The six review rows reproduced exactly: the CROSS JOIN message verbatim,
FROM "prod_us".orders o, "prod_eu".orders p ⇒ ListMap(orders -> p),
CREATE TABLE dest (COL INTEGER) parsing, CREATE LOCAL TEMPORARY TABLE rejecting with
regex '(?i)OR\b' expected but 'L' found, FROM elastic."bi_events" and the ON-less/ON-ful CROSS
JOIN pair. Corpus counts re-derived and confirmed (25 backtick + 23 double-quote qualified
statements, zero overlap; 24 DDL probes).

The four Issue-#57 tests are green, byte-unmodified

ParserSpec.scala's // ── Schema-qualified table names (Issue #57) ── block is untouched; this PR
appends two tests to it. If one of the four had gone red, the implementation would have joined
the parts and AD-1 would have been violated.

Design decisions taken at dev time

• AD-9 — tableParts uses 21.1's joined qualifiedName; qualifiedName itself was not
  touched, so 21.1's "one lexing surface" holds with a zero-line diff. Splitting the tail would
  render elastic."bi_events" as "elastic"."bi_events", which re-parses to a different index.
• AD-8 — qualifierPart keeps a separate dot terminal rather than one adjacency-strict
  regex. Measured: the strict form sends FROM "elastic" .bi_events (accepted today, reads
  bi_events) down qualifiedName, whose tail matches .bi_events adjacently, and it would
  silently start reading elastic.bi_events.
• AD-6′ — the alias-map key is the qualified reference only where the bare name is ambiguous
  in its own FROM. The unconditional form was implemented first and measured to break
  Identifier.table and two softclient4es-extensions call sites that look tables up by the bare
  index name, for statements that work today.
• AD-6″ (review follow-up) — Identifier.table is an alias-map key, so anything in sql
  comparing it against an index name must speak the same key language. New From.joinSourceKeys;
  TemporalLiterals' cross-index join-leg guard consumes it.
• AD-7′ (review follow-up) — Join.validate's CrossJoin exemption was unsatisfiable for every
  input before and after AD-7, so it is deleted rather than left as a second, unreachable
  statement of the same rule.

Consumer sweep for tableAliases / aliasesToTable / joinAliases (AC-12)

site	how it reads the map	status
sql/…/package.scala:1266 Identifier.update	reverse lookup → Identifier.table	key language now shared via joinSourceKeys
sql/…/query/TemporalLiterals.scala:304,349	compares Identifier.table against join sources	fixed in this PR
sql/…/query/OrderBy.scala:58	values only	unaffected
sql/…/query/package.scala:119	re-exposes the map	inherits
aliasesToTable	no consumer beyond its definition, in this repo or any sibling	—
extensions graph/JoinDependencyGraph.scala:517-528, graph/FieldAnalyzer.scala:94,97,152,250,296	look tables up in a schemas map keyed by the bare index	why AD-6′ is conditional; unaffected for every unambiguous statement
arrow planner/PredicatePushdown.scala:52-65	a Set of aliases	unrelated

#85 — the consumer note

softclient4es-arrow's JoinPlanner.extractCatalogList / stripAllCatalogPrefixes can retire for
the FROM/JOIN half on the next core bump and read Table.parts / StandardJoin.parts instead,
which also removes nextCatalogFor's positional mis-association. A retiring resolver must treat
an UNREGISTERED qualifier as "no qualifier", not as an error — after story 20.3 a BI tool's
qualifier is routinely the cluster name.

⚠️ The spec's "arrow is unaffected by anything this story does" is refuted:
CatalogTablePattern needs a backtick qualifier + a bare table name, so a fully quoted
FROM `prod_us`.`orders` is never stripped, now parses, and both legs silently run against the
default cluster where they used to fail loudly. Cross-repo, so not fixable here — warned in
joins.md and known_limitations.md, and recorded in the local arrow follow-up record.

Release notes

1. Table.sql / From.sql renderings now carry the qualifier the statement had, canonicalised to
   the ANSI double quote. Anything pinning generated SQL text must look.
2. Table and StandardJoin gained a parts field ⇒ binary-incompatible; downstream rebuild
   required. Shared with 21.1's GenericIdentifier/Alias note — one line for the train.
3. CREATE OR REPLACE MATERIALIZED VIEW on an unchanged definition REBUILDS once after the upgrade
   (the extension string-compares the stored render against a fresh one). Not data loss; an
   expensive surprise on first re-deploy. Shared with 21.1 — one train-level line.
4. CROSS JOIN without ON now validates; the alias maps of a table whose bare name is ambiguous
   in its own FROM are keyed by its qualified name (visible only to callers introspecting
   From.tableAliases / aliasesToTable / Identifier.table).
5. A JOIN leg with a quoted qualifier reads a different index. JOIN "prod_us".customers c
   read prod_us.customers before and reads customers now, matching what the FROM leg of the
   same statement has always done. Write the dotted name unquoted to keep the old reading.
6. A JOIN source can no longer be an expression. JOIN customers::BIGINT c and
   JOIN DISTINCT customers c parsed before — the cast and the DISTINCT reached nothing, while the
   render kept them — and are loud rejections now, matching the FROM side.
7. CREATE WATCHER … FROM a JOIN b ON … parses the JOIN and then silently discards it #191's multi-index watcher guard now fires on a cross-qualifier FROM.
   CREATE OR REPLACE WATCHER … FROM "a".orders o, "b".orders p WHERE o.x = 1 … was accepted and is
   now rejected. A wholly unqualified self-join still defeats it, as before.
8. Federation only: a fully quoted cross-cluster reference now parses and silently runs against
   the default cluster where it used to fail loudly. Quote the prefix, leave the table name bare
   (FROM `prod_us`.orders), until the arrow-side retirement lands.

Story 21.7 owns the rest

Parser.ident — quoted names for INSERT / UPDATE / CREATE / DROP / ALTER, table and
column — is not in this PR and stays rejected, pinned as such. known_limitations.md names the
residual as DML/DDL-only. The #-prefixed temp-table probe outcome and the temp-table product
decision ride 21.7.

Verification

sql 772 · core 887 · softclient4es-sql-bridge 165 · macrosTests 19 ·
+ sql/test and + core/test green on 2.12.20 and 2.13.16 · + sql/compile / + core/compile
· ++ 2.12.20 sql/Test/compile · scalafmtCheckAll. No emitted-JSON change — the index list is
byte-identical by construction. build.sbt's 0.23.0-SNAPSHOT verified and untouched. No core,
bridge or es{N} source was modified. Every SQL example added to the docs was parse-probed
against the built parser.

🤖 Generated with Claude Code