[satriadhikara]

Closes #571.

Problem

Validation of out-of-order tables (a table whose parts are split by other tables) had one false rejection and three false acceptances:

1. Valid TOML rejected (Regression in 0.15.1: valid out-of-order child table raises KeyAlreadyPresent #571). _validate_table_candidate looked up the existing value with item(k), which returns an OutOfOrderTableProxy for out-of-order keys. The proxy is not a Table, so the type check raised KeyAlreadyPresent for valid documents such as [a]\n[a.b.c]\n[a.b]\n[[zz]]\n[a.b.d].
2. Table declared twice across fragments accepted. append() compared a concrete [table] only with the last fragment of an out-of-order key, e.g. [a]\n[q]\n[a.b.c.e]\n[a.b]\n[a]\n[a.b.d] ([a] twice).
3. Stale incremental validation cache. Merging a super table in place into the last fragment mutated a fragment that _validation_cache had already copied, so later fragments were validated against an outdated copy: a.x=1\n[a.b.c]\n[a.b.d]\n[q]\n[a.b.d] ([a.b.d] twice).
4. Dotted-key table redefined by a header. A table created by dotted keys is stored as a super table, so the concrete-vs-concrete check never fired: [a]\nb.y=1\n[a.b.c]\n[q]\n[a.b].

Master rejected some invalid documents from cases 2 and 3 only by accident, through bug 1. Fixing bug 1 alone would have made them parse.

Fix

• _validate_table_candidate iterates the raw stored fragments instead of materializing the proxy, applying the existing checks to each one, and rejects a concrete [table] over a dotted-key fragment.
• _validate_earlier_fragments: a concrete [table] may not redefine an earlier concrete or dotted-key fragment. This is a cheap, non-recursive scan; nested collisions are left to the incremental validation.
• The in-place super-table merge also merges the new part into the cached validation container, keeping it in sync.

Verification

• Differential fuzzing against tomllib, about 250k generated documents mixing headers, arrays of tables and dotted keys:

  	master	this PR
  invalid documents accepted	2813	0
  documents rejected that master accepts	–	0

• Parse time stays linear: 0.40s vs master's 0.36s on 2000 split tables. An intermediate version of this branch was quadratic; that is fixed.

• 12 new tests: valid nested out-of-order layouts and each rejection case above.

• pytest: 1109 passed. ruff check and ruff format --check are clean.

• The one failing integration job (Ubuntu / 3.12 / poetry) fails on Poetry's test_extractall_sdist_no_symlink_path_traversal_via_directory_symlink (tarfile behavior of the runner's Python). It passes on macOS and Windows and is unrelated to this change.

Agent Drafting Metadata

• OpenAI Codex (GPT-5.6 Sol): reproduced and minimized Regression in 0.15.1: valid out-of-order child table raises KeyAlreadyPresent #571, drafted the original fix and tests.
• Claude (Opus 5.5): differential fuzzing against tomllib, found and fixed cases 2–4, and drafted this description.

[davidpavlovschi]

Reviewed the validation path against the reported nested out-of-order case and the existing duplicate/redefinition checks. Iterating the raw fragments avoids treating the proxy as a non-table while still validating every fragment against the candidate. I also ran the complete suite locally on this head: 1,051 passed, 1 skipped. The focused regression preserves both the unwrapped structure and exact serialization.

[satriadhikara]

Hi, following up on this PR after reading the discussion in #596.

I’ve updated the branch against current master and added regression tests for collisions across nested out-of-order fragments. All 1,058 tests pass, and lint and formatting checks pass.

Compared with #592, this approach applies the same type-conflict, concrete-table redefinition, and recursive checks to every stored fragment. This keeps validation in one path, avoiding proxy construction and a separate proxy-specific branch that would need to stay consistent with those checks. Both implementations pass the focused tests; I prefer this approach because it keeps the validation logic simpler.

Would you be able to take another look? Happy to address any concerns with this approach.

[satriadhikara]

Updated against current master. While fuzzing against tomllib I found that tables declared twice across out-of-order fragments slipped through, e.g. [a]\n[a.b.c]\n[a.b]\n[q]\n[a.b.d]\n[a.b] and a.x=1\n[a.b.c]\n[a.b.d]\n[q]\n[a.b.d] (master only rejected some of these by accident via #571). Two causes: a concrete [table] was only compared with the last fragment, and merging a super table in place into the last fragment left the incremental validation cache stale. Fixed both plus a dotted-key redefinition case, with regression tests. Over ~250k generated docs this branch accepts no document tomllib rejects (master wrongly accepts 2813) and rejects nothing master accepts; parse time stays linear.

[cig13zs]

Checked bdff635 (after the master merge): the example from #632 now parses, round-trips byte for byte and matches tomllib, while master (eddd27e) still raises Key "b" already exists. Full test suite passes locally (1215 passed).