Does anyone want to take over maintenance of the Kowainik ecosystem?

Hi! I really like tomland, especially for its bidirectional codecs, and it made …me a bit sad to see it fall behind the spec. Now that I have Claude at my side, I thought I'd put in the effort to push through some changes that were planned years ago (#248, #264, #373). There are a few of them and they are necessarily API-breaking, so I made sure to split them into individual commits that add up to full TOML 1.1.0 support. The short version: against the official toml-test suite (the files-toml-1.1.0 list), the parser now accepts all 218 valid documents and rejects all 494 invalid ones. Closes #373, closes #248, closes #264.

What's in here

The commits fall roughly into three groups. Parser (dc09f7d.. dac66c7). Mostly about the TOML 1.0 rules that weren't enforced, plus the 1.1 additions:

• whitespace, newlines and comments are only consumed where the spec allows them. sc now consumes only spaces and tabs and a new scn handles newlines/comments. A key/value pair or header has to be followed by a newline;
• whitespace around dots in keys, ASCII-only bare keys;
• string rules: \e and \xHH (1.1), tabs and non-ASCII in basic strings, control characters rejected in literal strings, "" before the closing """, etc.;
• times without seconds (1.1), lowercase t/ z, range-checked offsets, no leading zeros in floats or 0_1;
• inline tables can span lines and have a trailing comma (1.1), and can be nested in inline tables and arrays. There are two printer fixes in this group too, because the round-trip tests started failing on them. MultiLine mode printed key = followed by a newline, which isn't valid TOML. Control characters were printed with show, so you'd get \ESC, which no TOML parser accepts. Mixed arrays and unquoted keys (c4d29d5, 20fd317).
• Array holds [AnyValue] instead of [Value t]. There's a new array helper for the homogeneous case and an _Either BiMap to decode into a sum type.
• Piece used to keep the quotes, so key, "key" and 'key' were three different keys. The spec says they're the same key, and a codec had to know how the file spelled it. Now the parser strips the quotes and the printer adds them back when a piece isn't a bare key. New AST and validator (a9b79c2). This one is the big one and it's basically #264. The old TOML had two ways to encode a nested table, either as a prefix-tree sibling or as a nested sub-document, and which one you got depended on the order of headers in the file. As a result:
• lookups and Eq depended on header order;
• reopening a parent after a subtable duplicated the subtree;
• int "a.b" and table (int "b") "a" read different places;
• the validator had no idea which tables could still be extended, so it accepted 43 invalid documents from toml-test. Now TOML is a map from a single Piece to an Entry:

data Entry
    = EValue AnyValue
    | ETable TableKind TOML
    | ETableArray (NonEmpty TOML)
data TableKind = HeaderTable | ImplicitTable | DottedTable | InlineTable

Every level of nesting is exactly one piece, so a.b.c = 1, [a.b] + c = 1 and a = { b = { c = 1 } } all produce the same document. TableKind exists only so the validator can apply the 1.1 rules (inline tables are closed, dotted tables can't be reopened with a header, and so on). Eq ignores it. Value and TOML are now mutually recursive (Table :: TOML -> Value 'TTable for inline tables inside arrays), so they both live in Toml.Type.Value. The validator is now a single sequential pass that tracks scopes, with a new ExtendClosedTable error.

Compatibility

This is a breaking change, so I've bumped the version to 1.4.0.0. I tried to keep the damage limited to people who work with the AST directly:

• the codec API is unchanged, and existing codecs should keep working. list/ nonEmpty read both [[key]] and key = [ {..}, {..} ];
• tomlPairs, tomlTables and tomlTableArrays are still there as deprecated read-only functions with their old types, and Toml.Type.PrefixTree is kept as a deprecated module. My suggestion would be to drop them in 1.5;
• two things can't be shimmed: Array holding AnyValue, and Piece being unquoted. The second one is the nasty one, because Piece "\"a.b\"" still compiles but now means something different. I wrote a "Migrating from 1.3" section in the changelog that covers both.

Testing

• cabal test: 328 examples, 0 failures (GHC 9.12.4 locally; I haven't tried the rest of the CI matrix yet, so let's see what CI says).
• toml-test, 1.1.0 list: 218/218 valid accepted, 494/494 invalid rejected. tomland doesn't ship a toml-test decoder, so this is an accept/reject sweep with a small runghc script that calls Toml.Parser.parse. It doesn't check the decoded values against the expected JSON. Happy to add the script (or a proper decoder executable) to the repo if you think it's worth it.
• For what it's worth, the 1.0.0 list gives 208/208 and 492/501. The 9 misses are exactly the things 1.1 relaxed: times without seconds, multi-line inline tables and trailing commas in them, and \x escapes. There's no "1.0 mode", and I don't think one is needed.

Open questions

• Even split into commits, it's still a lot to review. The parser commits come before the AST change and stand on their own (the test suite passes at dac66c7), so I can split them into a separate PR if that's easier for you.
• I haven't measured performance. The new representation should be in the same ballpark (HashMap per level instead of a prefix tree), but I haven't checked.