fix(grammar): stop flagging valid SQLite migrations as partially parsed - #2957

What

Valid SQLite migrations were reported as partially parsed. The vendored SQL
grammar (DerekStride/tree-sitter-sql @ 851e9cb257ba) rejects SQLite syntax
that real migration files use, so each use became an ERROR node and the file
was flagged parse_partial:

• PRAGMA foreign_keys = ON;
• id INTEGER PRIMARY KEY AUTOINCREMENT
• a trigger body: CREATE TRIGGER ... BEGIN ... END;, also after an
  unparenthesized WHEN condition, with INSERT OR IGNORE inside
• a column named key after the first column
• ) WITHOUT ROWID;
• CREATE VIRTUAL TABLE messages_fts USING fts5(...)

IF NOT EXISTS on a trigger or an index was already accepted upstream; in the
#2246 reproduction the failing lines were the PRAGMA, the AUTOINCREMENT and the
trigger body.

The key case is lexical: after a comma inside a column list the grammar also
accepts MySQL's KEY name (cols) index, so key lexes as that keyword there
and key TEXT NOT NULL became a malformed index.

How

sql becomes a self-maintained fork of the pin we already vendored, on the
same model as tools/tree-sitter-pascal/:

• tools/tree-sitter-sql/ holds upstream's grammar.js + grammar/
  (blob-identical to the pin and to upstream main @ 3c99273f9b32, where its
  corpus lives), upstream's MIT LICENSE, and one patch,
  patches/sqlite-dialect.patch, which touches grammar/ only. Its index
  lines carry upstream's pre-image blob hashes.
• AUTOINCREMENT is upstream's own change (c042cde1e2a6, 2026-11-08, after
  our pin), taken verbatim, so that hunk disappears on the next refresh.
• The trigger body is the existing block node. Precedence keeps Postgres
  WHEN (...) EXECUTE FUNCTION trees unchanged.
• The virtual table is aliased to create_table, so it is extracted as a
  Table like any other.
• Upstream node types keep their shape. Node types only gain entries
  (pragma_statement, module_argument, four keywords; create_table gains a
  module field).
• The vendored scanner.c is unchanged. It carries our start_tag leak fix
  (0e61867), which is now listed in the MANIFEST's "Local source patches"
  table; upstream fixed the same leak after our pin (ee2f84f9b857).
• One trade-off, documented in the MANIFEST: since key can now start a
  column, the type keywords that take no arguments (name, date, text,
  uuid, ...) are accepted as MySQL index names so KEY name (name) stays an
  index. An unquoted MySQL index named after a type that takes arguments
  (KEY timestamp (ts)) is now an ERROR; backtick-quoted names, which
  mysqldump writes, are unaffected.

Registered like pascal: MANIFEST.md fork rows and notes (ABI tally recounted,
unchanged), the provenance audit's FORKS map, THIRD_PARTY.md, and the
vendored checksums.

Provenance

• Without the patch, tree-sitter-cli 0.25.10 regenerates upstream's shipped
  parser.c byte-for-byte except line 1 (upstream's generator stamp carries no
  version); grammar.json, node-types.json and tree_sitter/*.h are
  byte-identical. The previously vendored files were upstream's minus the final
  newline.
• Patched: three generations are byte-identical (parser.c SHA-256
  3c64e5c415d00f7d935b7f9f0516f7be04801b241bc5da2e5e71ca31cf337887), ABI 15,
  states 30,622 -> 31,081, symbols 770 -> 782, 0 non-ASCII bytes, no string
  literal containing a space, one comment (the generator stamp), and the same
  identifiers in call position as before.
• tree-sitter test: upstream's 511 corpus cases pass unchanged.
  test/corpus/sqlite.txt adds 11 cases; the 9 SQLite cases fail on the
  unpatched grammar and the 2 controls pass on both.

Verification

Real migration sets, each indexed from scratch with an isolated cache, main
6d34ca15 vs this branch:

• All 15 valid SQLite files are clean.
• The 7 atuin files still flagged are the invalid ALTER TABLE ... DROP INDEX
  down migration (SQLite has no such statement) and the 2 MySQL and 4 Postgres
  migrations, with the same ranges as before. No file is newly flagged.
• The new Table nodes are the fts5 table and a table that the old trigger error
  region swallowed.

Tests in tests/test_parse_coverage.c:

• The two SQLite tests from fix: clear false partial-parse warnings for SQLite and TSX #2246, carried over as written:
  sqlite_valid_migration_not_flagged (red before this change) and the
  sqlite_malformed_still_flagged control.
• New cases: key columns, WITHOUT ROWID, a WHEN trigger with
  INSERT OR IGNORE plus an fts5 table (asserted as a Table), PRAGMA forms.
• Two new controls: MySQL's KEY index still clean, and
  ALTER TABLE ... DROP INDEX still flagged.
• On main, 5 of these tests fail and the 3 controls pass. With this change,
  all pass.

Narrow suites under ASan/UBSan: parse_coverage, grammar_regression,
extraction, language, lang_contract, registry, grammar_labels, complexity,
grammar_probe_g and pipeline all pass (1,569 tests), with no sanitizer report.

Credit

Distilled from #2246 by @kerfern (Co-authored-by on the commit); the original
could not merge because its regenerated parser had no reproducible grammar
source and its commit has no DCO sign-off. Kerf's report and tests defined the
problem precisely, and the malformed-input control is theirs.

Follow-up

Offer the remaining SQLite forms to DerekStride/tree-sitter-sql so the local
patch can shrink on a later refresh. Upstream's AUTOINCREMENT request
(DerekStride/tree-sitter-sql#299) has since been closed by c042cde1e2a6,
which this fork already mirrors. Not opened yet.