toons

Added

• TOON specification v4.1 support, validated against the conformance fixtures of spec tag v4.1.1:
  • # comment lines, removed in a lexical pre-pass before every other rule, so a comment never ends a scope or counts as a row.
  • Nested field groups in tabular headers (orders[2]{id,customer{name,country},total}:), encoded and decoded at any depth.
  • Keyed tabular form for objects of uniform objects (servers[2:]{host,port}:), including the keyless root form and the header-on-hyphen-line form inside list items.
  • The canonical empty-array forms key: [] and []; the legacy key[0]: and [0]: are still decoded.
  • The normative decoder number grammar: .5, 1., +5, 05, Infinity, NaN, 0x10, and 1_000 decode as strings.
  • \uXXXX escapes in quoted strings and keys, emitted for control characters and rejected for lone surrogates.
  • Byte-order-mark removal, CRLF input, and trailing-space stripping.
• Strict-mode checks from Sections 12 and 14: duplicate sibling keys, indentation depth jumps, over-indented lines, blank lines inside a header span, trailing content after a root array, malformed bracket segments and field lists, header delimiter mismatches, keyless headers outside the document root, and scope width and count mismatches.
• Non-strict mode applies last-write-wins for duplicate keys and never lets a declared [N] truncate a scope.
• Tuples are encoded as TOON arrays, matching the json module.
• Integers outside the 64-bit range keep their exact value in both directions; previously they were silently converted to floats.
• toons.pyi declares __version__.
• toons.__toon_spec__ reports the implemented TOON specification version ("4.1"), as recommended by spec Section 13.

Changed

• The indent option of loads, load, dumps, and dump is now spelled indent_size, matching the spec's indentSize. indent stays accepted as a deprecated alias; passing both with different values raises ValueError. to_json keeps indent for the JSON output, as before, and gains indent_size for the TOON input.
• Breaking: the decoder no longer detects the indentation from the input; it defaults to 2 spaces, as spec Section 13 requires.
• Breaking: the encoder no longer chooses a form by preference. Tabular form is mandatory wherever detection succeeds, keyed tabular form applies in object-field and root positions, a list-item object's first field sits on the hyphen line, and empty arrays emit key: [].
• Breaking: Python 3.8 is now the minimum. PyO3 0.29 no longer offers the abi3-py37 target, so wheels are tagged cp38-abi3. Python 3.7 users stay on 0.7.0.
• Updated PyO3 from 0.28 to 0.29, refreshed the development and documentation dependencies, and bumped the GitHub Actions used by CI.
• Breaking: values with no TOON representation (sets, generators, custom objects, functions) now raise TypeError instead of being encoded as null, and non-string object keys raise TypeError instead of crashing the interpreter.
• Invalid option values raise ValueError instead of being ignored: delimiter must be ",", "\t", or "|"; the encoder indent_size must be at least 2 and the decoder indent_size at least 1.
• Out-of-domain numbers have a documented policy: integers stay exact, a decimal or exponent token beyond double precision is rejected in strict mode and decodes as an infinity when strict=False.
• Strings are quoted when they start with #, when they are numeric-like with a leading +, and no longer merely because they start or end with non-ASCII whitespace.
• Release builds enable link-time optimization and strip symbols.
• Read the Docs installs the documentation dependencies with uv sync against pyproject.toml and uv.lock, so the exported docs/requirements.txt is gone.
• Publishing to PyPI runs in its own pypi environment, so a tagged release shows up under the repository's Deployments and can be gated there, and it authenticates with trusted publishing instead of a stored API token.

Removed

• Breaking: the key_folding and flatten_depth encoder options and the expand_paths decoder option. Key folding and path expansion were removed from the specification in v4.0; dotted keys are single literal keys. Data encoded with key_folding="safe" must be re-hydrated with a v3 decoder using expand_paths="safe" before re-encoding.

Fixed

• Encoding a structure that contains a reference cycle raised a stack overflow that killed the interpreter; it now raises ValueError. Nesting deeper than 1000 containers raises ValueError for the same reason.
• dumps(..., delimiter="") panicked; a panic also escaped for non-string object keys.
• An array nested inside an array under an object key was written with a stray trailing space and a misplaced header, producing output that could not be decoded again.
• Decoding a document nested more than a thousand levels deep overflowed the stack and killed the interpreter; it now raises ToonDecodeError. The same bound applies to the encoder's form detection, which crashed on a deeply nested uniform column.
• Encoding a str that holds an unpaired surrogate raised a misleading TypeError about the type; it now raises ValueError naming the cause.
• The Read the Docs build failed because the crate used let chains, which need a newer compiler than the Rust toolchain pinned in .readthedocs.yaml. The pin now tracks latest, and rust-version in Cargo.toml states the 1.88 the crate actually needs.