Adopt legaldown-validator 0.4.0's public API; remove validator_bridge.py

[dvejsada]

The renderer now uses only legaldown-validator 0.4.0's public, semver-covered API, so validator_bridge.py and its private imports are gone. Rendered output is unchanged.

Merge only after legaldown-validator 0.4.0 is on PyPI. The dependency is now legaldown-validator>=0.4.0,<0.5, so the jobs that install from PyPI fail until that release is published. The "Against legaldown-validator main" job should pass straight away.

Changes

Was	Now
legaldown.parser.quote_content, MAX_QUOTE_DEPTH, and render's own _without_drafting_marker / DRAFTING_MARKER / past-depth branch	legaldown.syntax.quote_blocks / drafting_note_blocks (ForLegalAI/legaldown-validator#88, #93)
legaldown.markdown fence helpers and build.py _code_block	legaldown.syntax.code_content (#93)
legaldown.cli._read_answers	legaldown.load_answers, with the CLI keeping the same messages
parse_document / validate_document	parse / validate
assemble(source, answers)	parse_template(source).form(answers).assemble()
result.sections, placed_markers, is_template, *_lookup	result.index.*
legaldown.validator constants and conditions; top-level slugify_identifier	legaldown.grammar
top-level lex, Directive, find_definition_anchors, is_drafting_note, list_items, render_block	legaldown.syntax
resolver _VALUE_TYPES / _DECISION_TYPES	grammar.VALID_PLACEHOLDER_TYPES / DECISION_QUESTION_TYPES

• Deprecation guard: pytest now turns any legaldown DeprecationWarning into an error, so a call that 0.5.0 will remove cannot slip in.
• Docs: architecture, roadmap, CONFORMANCE (now "legaldown-validator 0.4"), and short notes in ADR 0002 and ADR 0007 saying the bridge is gone.

Behaviour

• Unit, golden and conformance tests: 192 unit tests pass (2 new) and 197 spec conformance tests pass. The goldens are unchanged, and ruff is clean.
• Output comparison: I rendered 1438 outputs before (with validator 0.3.0) and after, and they are byte-identical, diagnostics included. They cover every spec example and fixture plus the test documents, in 3 styles × HTML/text, in final mode, and through the CLI with --answers.
• One edge-case difference: a quote nested past the validator's quote depth (17+ levels) that holds only an HTML comment used to render an empty <p class="ld-p"></p>. It now renders nothing, as at every shallower depth. A new test pins this.

Left as is, on purpose

These are places where the 0.4.0 answer is not exactly what the renderer needs, so the renderer keeps its own logic.

• The placeholder survey vs result.index.blanks: the renderer needs one decision per occurrence, and Blank.consistent has a different scope.
• _choose vs grammar.choose_problem: for a choice question with malformed choices, choose_problem passes the directive, but the renderer must still show it as invalid.
• Answers are not coerced with Template.coerce: the renderer still reports answer-invalid for inputs like 5000 EUR instead of reshaping them. Coercing would be a behaviour change, so it is left for a separate decision.
• Directive lines: the renderer's own diagnostics still have no line number, because 0.4.0 gives the line of a block or list item but not of a directive inside one.

Release

PUBLISHING.md says a new validator minor version needs a renderer release. The renderer's version is not bumped here; bump it when cutting that release.

🤖 Generated with Claude Code

https://claude.ai/code/session_013WmBAc5T7UCKVdpUmg9qxz

---

Generated by Claude Code

[dvejsada]

CI status on this PR:

• Tests (Python 3.11/3.12/3.13) and "Specification examples and fixtures" are failing. They fail while installing, before any test runs: Could not find a version that satisfies the requirement legaldown-validator<0.5,>=0.4.0. This PR requires 0.4.0, which is not on PyPI yet. These jobs go green once legaldown-validator 0.4.0 is published (its release commit, ForLegalAI/legaldown-validator@d8b611d, passes CI). There is nothing to fix in this PR for them.
• "Against legaldown-validator main" failed for a reason that was this PR's. The job installed the renderer from PyPI before swapping in the validator from main, so the new version range could not be met. 54596c3 installs the validator from main first. In a fresh environment set up the same way, that gives legaldown-validator 0.4.0 and 192 passing tests.

---

Generated by Claude Code

[dvejsada]

Review: adoption of legaldown-validator 0.4.0

The PR moves the renderer onto validator 0.4.0's public API correctly, and rendered output is unchanged apart from the one edge case already described. I found one CI problem and fixed it in 90fccc0. The rest are small comments.

What was checked

• CI on this head, with 0.4.0 from PyPI (0.4.0 is published now, so the earlier red runs are out of date):
  • ruff is clean.
  • pytest: 192 passed and 1 skipped (the spec corpus) on Python 3.11, 3.12 and 3.13.
  • Conformance: 197 passed.
  • The job against validator main passes.
• Public API only: every legaldown import is in the __all__ of legaldown, legaldown.syntax or legaldown.grammar, which the validator README says are semver-covered. There are no private names, no internal modules and no 0.4.0-deprecated calls. Nothing refers to validator_bridge except the dated notes in the ADRs and the roadmap.
• Deprecation guard: LegaldownDeprecationWarning is exported at the top level of legaldown. All 8 warnings.warn calls in 0.4.0 use it. A test calling parse_document fails under the new filterwarnings.
• Old and new calls behave the same. Each was checked against the 0.3.0 and 0.4.0 sources and by running both:
  • VALID_PLACEHOLDER_TYPES and DECISION_QUESTION_TYPES hold the same values as the old tuples. They are frozensets, so the new isinstance(declared_type, str) guard is needed: checking an unhashable value against a frozenset raises a TypeError.
  • load_answers gives the same CLI messages and exit codes. A BOM is now accepted, and a very deeply nested YAML file now gives a clean error where it used to raise a traceback.
  • parse_template(...).form(...).assemble() returns the same AssemblyResult as assemble().
  • code_content gives the same result as the old _code_block on 30k random fence and indent inputs.
• Quote depth: passing depth=self.depth before the increment is what the depth parameter of quote_blocks and drafting_note_blocks means: how many list items and quotes enclose the block. The validator counts list items and quotes together, as the builder does, so there is no off-by-one.
• Rendered output, base (main + 0.3.0) against this PR (0.4.0): about 38k comparisons.
  • Inputs: the spec corpus, every document the test suite renders, sweeps nesting quotes and lists to depths 1–74, code fences, drafting notes, templates with good and bad answers, and the CLI's error paths.
  • Each comparison covers 3 styles, HTML and text, draft and final, strict, the API and the CLI.
  • The only difference is the documented one: an empty <p class="ld-p"></p> no longer appears.

Fixed in 90fccc0

The validator-main CI job would quietly stop testing main. It installed validator main first and the renderer second. That works only while main's version is inside >=0.4.0,<0.5. Once main becomes 0.5.0.dev, pip replaces it with 0.4.x from PyPI. The job would then pass against the released validator just when 0.5 starts removing the deprecated names. Now that 0.4.0 is on PyPI, the commit puts the original order back: the editable install first, then pip install --force-reinstall --no-deps from main. It also updates the matching sentence in PUBLISHING.md. Checked: ruff is clean, pytest has 192 passed and 1 skipped, and after the CI steps pip freeze shows the validator installed from git main.

Nits (no change needed to merge)

• PR description: the empty-paragraph difference is triggered at 17+ levels of quotes and list items combined, not quotes alone. For example, 10 list levels around 6 quotes holding a comment-only drafting note also trigger it.
• cli.py comment "Read as legaldown-validator reads an answers set": the validator's own CLI also runs Template.coerce, so 5000 EUR is accepted there and refused here (answer-invalid). The description already says this is deliberate. The comment could say the same.
• ADR 0007 lines 32–34 still name ValidationResult.placed_markers / is_template, and ADR 0002 line 27 still names legaldown.directives. The dated notes below them correct this, so it reads fine as history.
• Version: __version__ is still 0.2.0 while the package now needs validator 0.4. Bump it to 0.3.0 when cutting the release, as the description says.

Possible follow-ups using 0.4.0 features

• result.index.blanks (Blank.type / .consistent) could replace part of the placeholder survey.
• Document.line_of could give the renderer's own diagnostics block-level lines.

---

Generated by Claude Code