Render an emptied array of tables as an empty array

[dchaudhari7177]

Summary

Addresses the array-of-tables half of #553.

When every element is removed from an array of tables, there is no [[key]] header left to render, and _render_aot looped over an empty body and emitted nothing — so the key vanished from the output entirely:

doc = tomlkit.loads("[[a]]\nx = 1\n")
doc["a"].pop()
tomlkit.dumps(doc)   # '' — the 'a' key is gone

This renders the inline a = [] form instead, which is what tomlkit already produces for an empty array built through the API (doc["a"] = []).

The ordering part

Rendering a = [] in body order is not enough. TOML only reads bare key/value pairs before the first table header, so an emptied array of tables sitting after one would be read back as a key of that table:

[t]
q = 2

a = []      # parses as t.a, not a

So empty arrays of tables are hoisted to just before the first table header. Non-empty ones are untouched and still render in place, byte for byte.

Scope: the other half of the issue

The issue also covers doc["x"]["y"].pop("z") leaving an empty super table that renders as nothing. I have not fixed that here, because the current behaviour is pinned by two existing tests — test_delete_out_of_order_table_key and test_overwriting_out_of_order_table both assert that a super table emptied by del disappears from the output rather than rendering its own header.

Making emptied super tables render [x.y] breaks both. That looked like a deliberate call rather than an oversight, so I left it alone rather than rewriting your tests to suit. Happy to follow up if you'd like that changed — it is a small diff once the intended behaviour is settled.

Tests

Four tests added, covering the bare case, the hoisting case, an emptied array preceded by a scalar, and a control asserting non-empty arrays of tables still round-trip unchanged.

Revert-verified: with container.py reverted, the first three fail and the control still passes.

Full suite on Windows / Python 3.12: 1055 passed (1051 on a clean tree, plus these 4). ruff check and ruff format --check are clean on both changed files.

[davidpavlovschi]

The empty fallback breaks nested arrays of tables because it keeps the full header prefix even though the fallback now renders as a key/value inside the current table scope. Minimal case: parse [t] followed by [[t.a]], pop the only t.a element, then serialize. This head emits [t] / t.a = [], which reparses as t.t.a instead of t.a. The same duplication occurs at deeper tables and inside an AoT element. Please add a nested regression that asserts doc.unwrap() equals parse(doc.as_string()).unwrap(). For an empty AoT rendered inside [t], the key/value fallback needs to be a = [] in that scope, not t.a = []. I ran the full suite on this head as well: 1,054 passed, 1 skipped; the current tests cover only root-level empty AoTs.

[dchaudhari7177]

Thanks — you were right, and the case was wider than the one you posted.

Your minimal case ([t] then [[t.a]], pop the only element) was fixed in 84ad727. But writing the parametrized round-trip you asked for turned up two more shapes that were still wrong, both where the array's parent is an implicit super table:

input (pop the only element)	before	reparsed as
[[t.a]]	a = []	{'a': []} — the t is gone
[t] + [[t.u.a]]	[t] / t.u.a = []	{'t': {'t': {'u': {'a': []}}}}

The first is your bug in the other direction: dropping the prefix is only right when the parent actually emitted a [t] to be bare inside. The second shows the prefix has to be relative to the nearest emitted header, not the absolute path.

So the fallbacks of a scope are now collected up front — walking through the super tables that emit no header of their own — and each is keyed relative to the scope that will read it. [[t.a]] renders t.a = []; [[t.u.a]] under [t] renders u.a = [].

Verification on this head:

• doc.unwrap() == parse(doc.as_string()).unwrap() over 32 generated shapes (explicit/implicit parent × one/two levels × sibling sub-table × preceding header × trailing header): all pass. On 84ad727 six of those failed.
• The parametrized regression now covers 12 shapes, plus two exact-output tests pinning t.a = [] and u.a = [].
• Full suite: 1069 passed.
• Every tests/examples/*.toml still reparses byte-identical, so untouched documents are unaffected.