wapdown/tests/test_golden.py

70 lines
2.7 KiB
Python
Raw Permalink Normal View History

Extract the WML exporter from md2txt into wapdown md2txt is a suite of minimal text-based markup exporters, and every other renderer in it (text, micron, ama, gemini, nex) is a straight transform of a document into a flat text format. The WML renderer is neither: it emits an XML tree, and it restructures the document rather than transforming it -- splitting one file into many <card>s, inventing navigation between them, building a menu card that exists in no source document, and paginating on a byte budget. That is why it needed levers no sibling renderer needed, and why it alone forced a CARD_BREAK block kind and a {.card} directive into the shared parser that every other renderer ignores. wapdown is that renderer given a repository of its own, where restructuring a document into a navigable deck is the point. Ported: the Markdown parser, block event models, and pipeline core, each trimmed to what a deck compiler actually uses (no FIGlet, no hyphenation, no ASCII-art includes, no text-layout frontmatter), plus the eight inline regexes the renderer had been borrowing from a 1039-line text renderer. The plugin registry comes along so other WAP-era targets can register alongside wml later. Zero runtime dependencies. Config is now first-class: every lever is an argparse flag with validation and a frontmatter equivalent, layered CLI > frontmatter > default, replacing the generic --renderer-option KEY=VALUE passthrough. Five defects found while reading the code, each with a regression test: - Link and image URLs were shredded by the emphasis patterns, which ran before LINK_RE: a path like /v1_2_3/ became /v1<i>2</i>3/ inside the href. Links and images are now resolved and stashed first. - <table> was emitted as a direct child of <card>. The WML 1.3 content model is (onevent*, timer?, (do | p | pre)*), so it must sit in a <p>. - Link labels carried emphasis, but <a> is declared (#PCDATA | br | img)* -- there is nowhere valid for a <b> inside one. Markers are now stripped. - The menu card was built directly and never byte-packed, so the one card most likely to be large was the one card exempt from the budget. - Output was written with CRLF line endings. WML is XML served over HTTP. New, WAP-idiomatic rather than Markdown-idiomatic: - deck-level <template> hoisting nav that would otherwise repeat per card - <select> menus, picked with a keypad digit instead of scrolled to - --deck-per-card, one file per section cross-linked by filename, which is the real answer to per-deck cache limits - <head> metadata: Cache-Control max-age and <access> - a Home options softkey, which <prev/> alone cannot provide - an image policy (keep/alt/drop) for the WBMP-only reality of WAP 1.x browsers -- no conversion is performed 186 tests, where neither repo had any. Golden files pin whole-document output; every deck is validated against the real WML 1.3 DTD, vendored at tests/dtd/wml13.dtd. Rendering examples/trail.md still matches md2txt's committed output byte for byte, modulo the CRLF fix. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-23 09:27:49 +03:00
"""Byte-for-byte comparison against committed reference output.
These catch unintended drift in whole-document output, which unit tests on
individual constructs cannot. When a change is deliberate, regenerate with
`uv run python tests/regenerate_golden.py` and review the resulting diff.
"""
from __future__ import annotations
from xml.etree import ElementTree
import pytest
from conftest import EXAMPLES, GOLDEN, render_file
2026-09-26 12:41:09 +03:00
from golden_cases import CASES, MULTI_CASES, XHTMLMP_CASES
Extract the WML exporter from md2txt into wapdown md2txt is a suite of minimal text-based markup exporters, and every other renderer in it (text, micron, ama, gemini, nex) is a straight transform of a document into a flat text format. The WML renderer is neither: it emits an XML tree, and it restructures the document rather than transforming it -- splitting one file into many <card>s, inventing navigation between them, building a menu card that exists in no source document, and paginating on a byte budget. That is why it needed levers no sibling renderer needed, and why it alone forced a CARD_BREAK block kind and a {.card} directive into the shared parser that every other renderer ignores. wapdown is that renderer given a repository of its own, where restructuring a document into a navigable deck is the point. Ported: the Markdown parser, block event models, and pipeline core, each trimmed to what a deck compiler actually uses (no FIGlet, no hyphenation, no ASCII-art includes, no text-layout frontmatter), plus the eight inline regexes the renderer had been borrowing from a 1039-line text renderer. The plugin registry comes along so other WAP-era targets can register alongside wml later. Zero runtime dependencies. Config is now first-class: every lever is an argparse flag with validation and a frontmatter equivalent, layered CLI > frontmatter > default, replacing the generic --renderer-option KEY=VALUE passthrough. Five defects found while reading the code, each with a regression test: - Link and image URLs were shredded by the emphasis patterns, which ran before LINK_RE: a path like /v1_2_3/ became /v1<i>2</i>3/ inside the href. Links and images are now resolved and stashed first. - <table> was emitted as a direct child of <card>. The WML 1.3 content model is (onevent*, timer?, (do | p | pre)*), so it must sit in a <p>. - Link labels carried emphasis, but <a> is declared (#PCDATA | br | img)* -- there is nowhere valid for a <b> inside one. Markers are now stripped. - The menu card was built directly and never byte-packed, so the one card most likely to be large was the one card exempt from the budget. - Output was written with CRLF line endings. WML is XML served over HTTP. New, WAP-idiomatic rather than Markdown-idiomatic: - deck-level <template> hoisting nav that would otherwise repeat per card - <select> menus, picked with a keypad digit instead of scrolled to - --deck-per-card, one file per section cross-linked by filename, which is the real answer to per-deck cache limits - <head> metadata: Cache-Control max-age and <access> - a Home options softkey, which <prev/> alone cannot provide - an image policy (keep/alt/drop) for the WBMP-only reality of WAP 1.x browsers -- no conversion is performed 186 tests, where neither repo had any. Golden files pin whole-document output; every deck is validated against the real WML 1.3 DTD, vendored at tests/dtd/wml13.dtd. Rendering examples/trail.md still matches md2txt's committed output byte for byte, modulo the CRLF fix. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-23 09:27:49 +03:00
REGENERATE_HINT = (
"Output changed. If that was intended, run "
"`uv run python tests/regenerate_golden.py` and review the diff."
)
@pytest.mark.parametrize("name", sorted(CASES))
def test_matches_golden(name):
source, options = CASES[name]
golden = GOLDEN / f"{name}.wml"
assert golden.exists(), f"missing golden for '{name}'; {REGENERATE_HINT}"
actual = render_file(EXAMPLES / source, **options).text
assert actual == golden.read_text(encoding="utf-8"), REGENERATE_HINT
@pytest.mark.parametrize("name", sorted(MULTI_CASES))
def test_matches_golden_directory(name):
source, options = MULTI_CASES[name]
directory = GOLDEN / name
assert directory.is_dir(), f"missing golden dir for '{name}'; {REGENERATE_HINT}"
documents = {
doc.name or "index.wml": doc.text
for doc in render_file(EXAMPLES / source, **options).documents
}
expected = {path.name: path.read_text(encoding="utf-8") for path in directory.iterdir()}
assert sorted(documents) == sorted(expected), REGENERATE_HINT
for filename, text in documents.items():
assert text == expected[filename], f"{filename}: {REGENERATE_HINT}"
@pytest.mark.parametrize("path", sorted(GOLDEN.glob("**/*.wml")), ids=lambda p: p.name)
def test_golden_files_are_well_formed(path):
ElementTree.fromstring(path.read_text(encoding="utf-8"))
2026-09-26 12:41:09 +03:00
@pytest.mark.parametrize("name", sorted(XHTMLMP_CASES))
def test_matches_golden_xhtmlmp(name):
source, options = XHTMLMP_CASES[name]
golden = GOLDEN / f"{name}.xhtml"
assert golden.exists(), f"missing golden for '{name}'; {REGENERATE_HINT}"
actual = render_file(EXAMPLES / source, renderer_name="xhtmlmp", **options)
assert actual == golden.read_text(encoding="utf-8"), REGENERATE_HINT
@pytest.mark.parametrize("path", sorted(GOLDEN.glob("**/*.xhtml")), ids=lambda p: p.name)
def test_xhtmlmp_golden_files_are_well_formed(path):
ElementTree.fromstring(path.read_text(encoding="utf-8"))
Extract the WML exporter from md2txt into wapdown md2txt is a suite of minimal text-based markup exporters, and every other renderer in it (text, micron, ama, gemini, nex) is a straight transform of a document into a flat text format. The WML renderer is neither: it emits an XML tree, and it restructures the document rather than transforming it -- splitting one file into many <card>s, inventing navigation between them, building a menu card that exists in no source document, and paginating on a byte budget. That is why it needed levers no sibling renderer needed, and why it alone forced a CARD_BREAK block kind and a {.card} directive into the shared parser that every other renderer ignores. wapdown is that renderer given a repository of its own, where restructuring a document into a navigable deck is the point. Ported: the Markdown parser, block event models, and pipeline core, each trimmed to what a deck compiler actually uses (no FIGlet, no hyphenation, no ASCII-art includes, no text-layout frontmatter), plus the eight inline regexes the renderer had been borrowing from a 1039-line text renderer. The plugin registry comes along so other WAP-era targets can register alongside wml later. Zero runtime dependencies. Config is now first-class: every lever is an argparse flag with validation and a frontmatter equivalent, layered CLI > frontmatter > default, replacing the generic --renderer-option KEY=VALUE passthrough. Five defects found while reading the code, each with a regression test: - Link and image URLs were shredded by the emphasis patterns, which ran before LINK_RE: a path like /v1_2_3/ became /v1<i>2</i>3/ inside the href. Links and images are now resolved and stashed first. - <table> was emitted as a direct child of <card>. The WML 1.3 content model is (onevent*, timer?, (do | p | pre)*), so it must sit in a <p>. - Link labels carried emphasis, but <a> is declared (#PCDATA | br | img)* -- there is nowhere valid for a <b> inside one. Markers are now stripped. - The menu card was built directly and never byte-packed, so the one card most likely to be large was the one card exempt from the budget. - Output was written with CRLF line endings. WML is XML served over HTTP. New, WAP-idiomatic rather than Markdown-idiomatic: - deck-level <template> hoisting nav that would otherwise repeat per card - <select> menus, picked with a keypad digit instead of scrolled to - --deck-per-card, one file per section cross-linked by filename, which is the real answer to per-deck cache limits - <head> metadata: Cache-Control max-age and <access> - a Home options softkey, which <prev/> alone cannot provide - an image policy (keep/alt/drop) for the WBMP-only reality of WAP 1.x browsers -- no conversion is performed 186 tests, where neither repo had any. Golden files pin whole-document output; every deck is validated against the real WML 1.3 DTD, vendored at tests/dtd/wml13.dtd. Rendering examples/trail.md still matches md2txt's committed output byte for byte, modulo the CRLF fix. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-23 09:27:49 +03:00
def test_every_example_has_a_case():
"""A new examples/*.md without a golden case would go unverified."""
covered = {source for source, _ in list(CASES.values()) + list(MULTI_CASES.values())}
assert {path.name for path in EXAMPLES.glob("*.md")} == covered