feat: resolve Oracle join-view INSERT targets by column attribution

[takaebato]

Summary

Implements the remaining inline-view deferral from #58 (design agreed in
discussion): an Oracle join-view INSERT now resolves to its base table by
column attribution, instead of dropping the whole statement.

INSERT INTO (SELECT e.id, e.name FROM emp e JOIN dept d
             ON e.dept_id = d.id WHERE d.active = 1) VALUES (100, 'x')
-- before: everything empty + UnsupportedStatement
-- after:  create=[emp]  read=[emp, dept]
--         writes=[emp.id, emp.name]
--         reads=[emp.dept_id, dept.id, dept.active]  (ON + WHERE, filter)

Attribution rule (composes two existing mechanisms — no new rules)

The row lands in the one relation every projected column agrees on:

projection column	attribution	precedent
qualified (e.id)	its qualifier, resolved against the view's relations — text-only, no catalog	multi-table UPDATE's SET t2.col
unqualified (name)	the catalog-owner rule (unqualified_write_binding, #59) over the view's relations	#59's read-mirrored SET attribution

Any column ambiguous / unresolved / owned by a different relation than its
siblings — or any factor naming a declared CTE (not a base table) — → drop +
flag as before. Real Oracle rejects those shapes too (the INTO columns must
all belong to one key-preserved table); key-preservedness
itself isn't verified — that needs unique-key metadata the catalog doesn't
carry — attribution assumes a valid statement. The classic comma-join
spelling (FROM emp e, dept d WHERE …) resolves identically.

Companion relations: Insert::target_context

The joined tables other than the target (here dept) become scanned context:

• table reads ✓ (children() now includes the context, so the scan
  surfaces; emp itself keeps the sink rule — it reads because its columns
  are referenced, not via a scan, consistent with the single-table view path)
• no data feed ✓ — table lineage's feeding_scans walks input only,
  matching the semantics (the join gates which rows are visible; the value
  path is the source)
• ON + WHERE land in the existing target_predicate (filter reads over the
  full view scope, aliases included)

TableReference's TryFrom<&Insert> stays shape-determined: the gate
admits joins of plain tables, but a join view's base table needs binder
resolution, so the plain identity parse resolves single-table views only
(documented; join → Err).

Edge cases covered on review

• CTE factors: any view factor naming a declared CTE — target (used to
  fabricate a write to the CTE) or companion (used to surface the CTE name as
  a phantom base-table read) — now flags the whole statement, uniformly in
  the factor loop. A WITH inside the view stays gate-rejected; a top-level
  CTE consumed only by the source still traces through its body (pinned).
• USING / NATURAL join views resolve, and the USING clause itself adds no
  reads — consistent with the SELECT path, where reads come from reference
  sites (pinned).
• Self-join views resolve to the one table (identity is right;
  key-preservedness is per-instance and unverifiable without key metadata).
• Table lineage: the companion never feeds (s → emp only, no
  dept → emp) — pinned at table level.
• Gate rejections pinned: FROM-less view, derived factor (single and joined),
  wildcard / expression projections, ClickHouse TABLE FUNCTION targets
  (moved code) incl. an unrepresentable function name.
• For full patch coverage, the gate hands its proof over entirely as
  data (InsertTargetView { factors, projection, join_operators, selection }
  — parse-don't-validate): the unreachable defensive arms are deleted, and
  no consumer re-reads the Query, so factors can't be paired with a
  clause it wasn't derived from.

Tests

Qualified projection (catalog-free), comma form, SELECT-source lineage through
the view (s.x → emp.id), straddling columns → flag, unqualified without a
catalog → flag, unqualified with a catalog attributing by owner
(Oracle-cased catalog), table-level companion read + lineage, USING /
self-join / gate-rejection pins, all CTE placements (factor → flag, inside
the view → flag, top-level source-only → traces through the body), and the
#58 flag test narrowed to an undetermined shape. All gates green (fmt /
clippy / test --all / doc); 788 tests, 100% patch coverage on source
lines (llvm-cov ∩ diff).

🤖 Generated with Claude Code

[github-actions]

✅ PR title follows the Conventional Commits spec.

[codecov]

Codecov Report

❌ Patch coverage is 95.83333% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 94.92%. Comparing base (13f18bd) to head (74bf42d).

Files with missing lines	Patch %	Lines
sql-insight/src/resolver/binder/statement.rs	94.53%	0 Missing and 7 partials ⚠️
sql-insight/src/reference.rs	98.24%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #61      +/-   ##
==========================================
+ Coverage   94.84%   94.92%   +0.07%     
==========================================
  Files          27       27              
  Lines        5747     5893     +146     
  Branches     5747     5893     +146     
==========================================
+ Hits         5451     5594     +143     
+ Misses        216      212       -4     
- Partials       80       87       +7     

☔ 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.