fix: don't reject an out-of-order super-table extended by a dotted key

[chuenchen309]

Summary

tomlkit rejects valid TOML on the parse() path when an out-of-order table reopens an implicit super-table with a dotted key.

[a.b.c] makes a.b an implicit super-table. Reopening [a] and extending a.b with a dotted key (b.z = 1) only adds a key to that super-table — it is not a redefinition — but Container._validate_table_candidate raises Redefinition of an existing table:

import tomlkit, tomllib
doc = "[a.b.c]\n[a]\nb.z = 1\n"
tomllib.loads(doc)   # -> {'a': {'b': {'c': {}, 'z': 1}}}   (accepted)
tomlkit.parse(doc)   # -> ParseError: Redefinition of an existing table at line 3 col 0

stdlib tomllib — the validity oracle the existing tests already cite (test_reject_out_of_order_dotted_* reference it) — accepts it.

Cause: the if k.is_dotted(): raise guard in _validate_table_candidate fires before the super-table recursion right below it, so a dotted key whose existing side is a super-table never gets to the recursion that would check the subtree for a real conflict. The dotted rejection is only correct when the existing side is a concrete table.

Fix: only take the dotted-key rejection when the existing entry is not a super-table; otherwise fall through to the existing recursion (which raises only on an actual redefinition). Two lines.

The concrete-table case ([a.b] then [a] b.z = 1) stays rejected, matching tomllib and the redefinition rejections added in #516 / #530 — all 27 out-of-order / redefinition tests still pass, and the full suite passes (372, including the new regression test). The target case now parses, round-trips byte-for-byte, and unwrap() matches tomllib.

Regression test added next to the sibling out-of-order tests in tests/test_toml_document.py; it is red before the fix (ParseError) and green after.

Note: I also noticed a separate pre-existing defect where parse() succeeds but .unwrap() raises KeyAlreadyPresent on some valid AoT-based out-of-order docs (e.g. [[y.a]]\n[b.d.x]\n[[y.a.c]]). That's a different code path (AoT merge) and is not touched here — flagging only in case it's of interest.

Agent Drafting Metadata

• Agent: Claude Code
• Model: Claude Opus 4.8
• Notes: An AI-assisted review pass surfaced the over-rejection; the repro against tomllib, the fix, the regression test, and the full-suite run were verified before opening. I read the _validate_table_candidate logic and confirmed the fix preserves every intended rejection (reject a table redefined after its parent table header #516/fix: reject out-of-order concrete+super table redefinitions at parse time #530) myself.