[Sskift]

Summary

The public .value view drops earlier array-of-tables elements when unrelated headers split their declaration:

doc = tomlkit.parse("[[a.rows]]\nx = 1\n[other]\ny = 2\n[[a.rows]]\nx = 3\n")
doc["a"]["rows"]  # contains both elements
doc.value["a"]["rows"]  # only x = 3 before this change

This also makes the document compare unequal to its own unwrap() result. Combine the fresh lists produced by AoT.value while merging out-of-order dictionary views. Ordinary formatted arrays retain their existing wrapper and behavior.

Validation on Windows / CPython 3.12.10: all four new split-AoT regression cases fail against the original source; all 1,229 tests pass with this change. Ruff lint/format and git diff --check pass, and mypy reports the same 33 baseline diagnostics. A separate Codex agent checked 97 document tests, 96 additional cases, repeated-read isolation, and ordinary-array identity. Six scalar-table timing comparisons remained within about 4% of the baseline.

Duplicate Check

• I searched the open PRs and found none that already fixes this issue.
• Related open PRs: PR Render each run of a split array of tables where it was written #601 changes where split arrays of tables are rendered; Fix out-of-order AoT element extensions #578 and Merge an out-of-order fragment that extends an AoT element #584 address extending their last element. This change fixes the public value/equality read path for already valid parsed fragments.

Agent Drafting Metadata

• Agent: OpenAI Codex
• Model: Exact model identifier was not exposed to this worker.
• Notes: Codex assisted investigation, implementation, tests, and this description. Independent review was performed by another AI agent; no human review is claimed.