fix: stop DML binder paths from silently losing writes and reads

[takaebato]

Summary

Six related binder bugs, all making UPDATE / INSERT / SELECT INTO surfaces silently incomplete, plus the first dialect-semantics mechanism.

• A qualified SET target whose qualifier named no writable relation dropped the whole assignment. UPDATE t SET address.city = old_city || '-x' lost the RHS reads, the lineage, and — as the sole assignment — the entire UPDATE from writes and the CRUD buckets, with no diagnostic. The dotted path now resolves by dialect: in struct-capable dialects (PostgreSQL family — whose SET grammar forbids relation qualifiers outright, so the leading segment is always a column, verified on PostgreSQL 18 — plus BigQuery / Hive / DuckDB / Generic) it reads as a struct subfield path on the single sink, at any depth (SET address.city and SET address.city.zip both write t.address; a leading segment that addresses the root in scope — the alias when aliased, the bare name otherwise — is a stripped prefix, and a shadowed name is a column: UPDATE t AS x SET t.a writes column t, exactly as PostgreSQL 18 reads it); in table-qualifier-only dialects (MySQL / MSSQL / SQLite / ClickHouse, where such a qualifier can only be a mistyped table path), among several writable relations, and for unrepresentable paths it surfaces unattributed (table: None, Unresolved). The assignment is never discarded. The tuple form inherits the fix, which also restores its reads/lineage symmetry.
• Dialect-semantics mechanism: the binder now holds the parsing dialect, and each dialect-dependent resolution rule is a question method in the new binder/dialect.rs concern file (supports_struct_set_targets() is the first). Bind code never downcasts the dialect ad hoc; a future dialect-dependent rule is one more method there.
• T-SQL UPDATE <alias> ... FROM t AS <alias> fabricated a phantom write target. The bare target name bound as a table a, so writes pinned a nonexistent relation while a.id became ambiguous between the phantom and the real one. When the join-less bare target names a FROM alias, the FROM relations now bind first (mirroring DELETE's USING order) and the target resolves through them: writes: t.x, a.id reads t.id.
• SELECT INTO buried a data-modifying CTE's write. WITH c AS (INSERT INTO t1 ... RETURNING a) SELECT a INTO t2 FROM c wrapped the CreateTableAs around the statement's WITH, hiding the CTE's INSERT inside the root's input where the write-root collection never descends — t1's write vanished from writes, CRUD, and lineage. The WITH now stays outermost, as PostgreSQL's top-level-only rule for data-modifying CTEs implies.
• PG's INSERT INTO counters AS c upsert alias was ignored, leaving the conflict action's c.n unresolved (no sink read, no lineage source). The alias now rides the conflict and RETURNING scopes.
• UPDATE ... FROM / DELETE ... USING loops dropped merge columns. The joined relations were kept but their USING / NATURAL merge columns were not, so an unqualified reference to one stayed Ambiguous instead of fanning in to the joined sides as the same join does in a SELECT A parenthesized-join MERGE source had the same drop, one site over — fixed alike.
• MySQL UPDATE ... ORDER BY keys were never bound (DELETE's were), losing their reads.

Tests

• New pins for each: struct subfield targets (plain, root-prefixed, nested-depth, aliased-target shadowing, tuple, and the MySQL unattributed counterpart), the unattributed multi-writable case, the T-SQL alias form, the SELECT INTO CTE write, the upsert target alias, the UPDATE FROM and MERGE-source fan-ins, and the ORDER BY reads. One existing pin updated: a 5-segment SET target now attributes to the leading column instead of vanishing. Full workspace: 866 tests, clippy and rustdoc clean.

🤖 Generated with Claude Code

[github-actions]

✅ PR title follows the Conventional Commits spec.

[codecov]

Codecov Report

❌ Patch coverage is 94.83871% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.21%. Comparing base (97132af) to head (5246254).

Files with missing lines	Patch %	Lines
sql-insight/src/resolver/binder/statement.rs	94.69%	4 Missing and 3 partials ⚠️
sql-insight/src/extractor/crud_table_extractor.rs	50.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #72      +/-   ##
==========================================
+ Coverage   95.19%   95.21%   +0.01%     
==========================================
  Files          27       28       +1     
  Lines        6179     6307     +128     
  Branches     6179     6307     +128     
==========================================
+ Hits         5882     6005     +123     
- Misses        206      209       +3     
- Partials       91       93       +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.