fix(lsp): position inlay hints from the document being edited

[tripleaceme]

Description

Closes #5630.

Inlay hints render inside a column name while it is being typed — zip_code shows as zip::TEXT_code — and correct themselves on save, re-render or reopening the file.

The cause is not a race, debouncing, or a document-version mismatch. sqlmesh/lsp/main.py registers did_open and did_save and no textDocument/didChange handler, so the server's parsed model is frozen until the file is saved. get_hints derives every hint's character from sqlglot token metadata on lsp_context.context.get_model(...).query, which is the model parsed from disk at the last load. VS Code invalidates and re-requests hints on every change, so it is asking correctly; the server answers with last-saved offsets.

saved buffer:    1 AS big_boy              hint at character 14
edited buffer:   1 AS big_boy_column       server still returns character 14
VS Code renders: 1 AS big_boy::INT_column

Version-stamping the responses would not have helped — the server has no correct answer at any version, it would simply withhold a wrong one.

get_hints now takes the document text and reads positions from parsing that, using sqlmesh.core.dialect.parse plus _split_sql_model_statements, which is the loader's own query-extraction rule, so macro-union and jinja models behave as before. Types are still resolved by name against the loaded model. main.py passes ls.workspace.get_text_document(...).source, which pygls keeps in sync with didChange.

Behaviour, all covered by tests:

buffer state	before	after
unchanged since save	correct	byte-identical
column renamed / added	hint lands inside the name	no hint until the next reload
line inserted above	all hints off by that shift	hints shift correctly
mid-edit, unparseable	positions from the stale parse	no hints

A second symptom of the same cause, not reported in the issue: inserting a line above shifted all the text but not the hints, so they attached to the wrong columns with the wrong types.

Three things for your judgement

1. Hints disappear while a column is being renamed, rather than appearing inside the word. I think that is the better of the two, but it is a UX call. The alternative is holding the last good hint and suppressing only the ones that would split a token.
2. This parses the model file per request, and hints are requested on roughly every keystroke. It is one small file and comparable to the normalize_identifiers(query.copy()) already done per request, but I have not benchmarked it against a large model.
3. _split_sql_model_statements is private. I imported it rather than re-implementing the rule, so the two cannot drift. sqlmesh/lsp/ already reaches into internals such as context._models and target._path, so this seemed in keeping, but you may prefer it promoted or duplicated.

Out of scope here, but worth flagging: diagnostics are published from the same reload-on-save context, so they are likely stale between saves in the same way. I have not investigated.

Test Plan

Two tests in tests/lsp/test_hints.py:

• test_hints_are_positioned_from_the_edited_document — a column renamed in the buffer but not saved produces no hint inside the name, and the remaining hints stay correct.
• test_hints_for_unparseable_document — mid-edit text that does not parse yields no hints rather than stale ones.

Both fail without the change, and the failure prints the bug verbatim:

assert 'SELECT customer_id::INT, zip_code' in "...SELECT customer_id::INT, zip::TEXT_code..."

pytest tests/lsp
63 passed

ruff check, ruff format --check and mypy clean on the changed files. No TypeScript touched — the extension has no inlay-hint code, it uses the stock vscode-languageclient provider.

Checklist

• I have run make style and fixed any issues
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test)
• My commits are signed off (git commit -s) per the DCO

[tripleaceme]

@cmgoffena13 — closes #5630.

The cause turned out to be simpler than the symptom suggests: the LSP has no textDocument/didChange handler, so its parsed model is frozen until save, and the hints carry offsets from that stale parse. VS Code is requesting them correctly.

Three judgement calls are listed at the end of the description — mainly that hints now disappear while a column is being renamed rather than rendering inside the word. Also worth knowing: diagnostics come from the same reload-on-save context, so they are probably stale in the same way. I have not investigated that and it is out of scope here.