[AlexanderWalls]

Closes #1.

The problem

Every consumer of the dialect catalog treated "not in the 42-entry ATTRS list"
as "invalid", so th:content, th:for, th:open — and ~160 other perfectly
legal attributes — produced Unknown Thymeleaf attribute warnings.

What the Standard Dialect actually accepts

I went to the tutorial
rather than adding the three reported names, because the gap is much wider than
three, and it has a documented shape:

• §5.2 "Setting value to specific attributes" — ~140 th:<html-attr>
  processors (th:content, th:for, th:cite, th:colspan, every th:on*
  handler, …).
• §5.5 "Fixed-value boolean attributes" — 24 boolean processors
  (th:open, th:required, th:async, th:hidden, …).

Beyond both lists, Thymeleaf 3's StandardDefaultAttributesTagProcessor handles
any remaining th:* name generically — that is how th:data-* custom
attributes work.

The change

Two-tier recognition in src/thymeleaf.rs. ATTRS keeps the curated,
hand-documented processors; the two new lists are transcribed verbatim from the
tutorial and get synthesized descriptions. classify() -> AttrKind resolves the
curated tier first, so hand-written docs always win for names that appear in
both (th:class, th:href, th:fragment, …). th:data-* / th:aria-* count
as known too.

The diagnostic is re-aimed at typos. Since the default attribute processor
accepts any name, "not in a list" was never a sound basis for a warning — but
catching th:txt genuinely is useful. So the warning now fires only when a name
is unrecognized and within MAX_SUGGESTION_DISTANCE of a real one, and it
names the intended attribute:

Unknown Thymeleaf attribute 'th:txt'. Did you mean 'th:text'?

closest_known + levenshtein moved from code_actions into thymeleaf, so
the diagnostic and the quick fix share one search. That is now an invariant:
every warning has a matching Change to ... action.

The suggestion threshold drops 3 → 2. With the candidate pool grown from 42
to ~200 names, distance 3 matches almost any short word by chance — th:bogus
would have "corrected" to th:cols. At distance 2 the separation is clean:

input	nearest	distance
th:txt	th:text	1	warns
th:contnt	th:content	1	warns
th:fragmnet	th:fragment	2	warns
th:opne	th:open	2	warns
th:bogus	th:cols	3	silent
th:wibble	th:title	3	silent

The other consumers follow the same classifier, so generic attributes are no
longer second-class: hover documents them (via a new owned AttrInfo so the
generic tier can synthesize text), semantic_tokens highlights them, and
completion offers both tiers with sort_text ranking curated names first.
The bracket-balance ERROR now applies to every th:* value rather than only
catalog ones — the value of any Thymeleaf attribute is an expression.

Verification

make check is clean (fmt, clippy -D warnings, 85 tests — 25 new).

Two existing tests asserted a warning on th:bogus, which is silent by design
now; both were retargeted to th:txt and additionally assert the message names
its fix. New coverage includes the issue's exact repro producing zero
diagnostics, th:data-*, the curated-wins-over-generic precedence, and the
threshold boundary cases above.

Also checked against the real index.html (with the three repro snippets added
to it): the only diagnostic left is the intentional th:txt typo.

Notes for review

• The two lists are transcribed from the tutorial rather than derived from an
  HTML attribute index, so they track what Thymeleaf documents rather than what
  the HTML spec happens to contain.
• MAX_SUGGESTION_DISTANCE is the one judgement call here. I documented the
  reasoning at the constant; happy to move it if you'd rather be more or less
  aggressive.
• Completion now offers ~160 more items. They are ranked below the curated ones
  and clients filter as you type, but say the word if you'd rather keep the menu
  curated — it is a one-line change to drop the generic tier from completion
  only.

[dozken]

thanks for PR

[AlexanderWalls]

thanks for PR

No worries - I'll keep using and contrib things when I see fit