refactor(tui): drive cursor context from the token stream (T9)

[TimelordUK]

Completion decided where the cursor was by scanning strings. Two near-identical scanners did it — analyze_statement when the partial query happened to parse, analyze_partial when it did not — using rfind('.'), split_whitespace().last(), ends_with(" AND") and query_upper.contains("SELECT"). analyze_partial even built a token stream first and then ignored it, and a third heuristic (determine_context) ran when both returned Unknown.

Measured on data/countries.csv:

WHERE region = ' WhereClause, 76 column names inside the
string literal
WHERE region = 'Am WhereClause, nothing
WHERE region IN ('Asia', ' WhereClause, 76 column names
WHERE "name.common" = no comparison operator found at all
WHERE region= no comparison operator found at all
GROUP BY suggested WHERE and ORDER BY
LIMIT suggested a table name

Replaced by src/sql/cursor_context.rs, the one owner of where is the cursor: truncate at the cursor, tokenize once, and match on the tail of the stream. The last token says what kind of position this is; the token before it disambiguates a bare word. Keywords are Token variants, so a keyword the lexer learns the completer learns with it — the completer knew six, Token::from_keyword knows fifty-five.

Three fixes fall out as consequences rather than as separate changes: quoted and dotted columns reach their operator, because columns come from Identifier/QuotedIdentifier joined across Dot instead of chars().all(|c| c.is_alphanumeric() || c == '_'); operators no longer need surrounding spaces; and keywords inside a literal stay inside it, because the unterminated-StringLiteral check runs before anything else looks at the stream.

The AST the old code went to the trouble of building was consulted only for stmt.where_clause.is_some()-shaped questions the tokens answer directly, so detect_cursor_context no longer parses the query — one less full parse per keystroke. 539 lines out of recursive_parser.rs.

Settles char-vs-byte offsets rather than deferring them: the lexer gained tokenize_all_with_byte_positions, and the analyzer speaks bytes throughout. tokenize_all_with_positions indexes the internal Vec, while cursor_pos and replace_start (T1) are byte offsets; they agree only on ASCII, and a wrong offset would have silently corrupted a spliced value.

T4 is now "put values in the empty vector":
CursorContext::InStringLiteral { column, in_list, value_start } is produced, threaded through ParseResult::replace_start and exempted from identifier filtering. The completer deliberately offers nothing there, which is still an improvement on offering column names inside quotes.

determine_context / ParseState survive as the Unknown fallback. Removing them is T12, kept separate because ParseState also has users in src/completer.rs and main.rs validation.

One test changed rather than added: test_order_by_quoted_partial_completion asserted partial_word == Some(""Customer"), opening quote included, which was an artefact of the deleted scanner. The partial now comes from the lexer, which has consumed the quote. Nothing filters on partial_word — that is find_completion_token's job and it still sees the quote — so the test's two substantive assertions are untouched.

Tests: 20 unit tests in cursor_context.rs, 8 in tests/completion_schema.rs driving the real get_completions entry point against countries.csv. Full suite green: 800 lib, 482 integration, 534 Python, 152/153 examples (the one failure is a pre-existing smoke test with no expectations, on a path completion cannot reach).

Claude-Session: https://claude.ai/code/session_01JkqLCLn1Usij41xUjMUoPz