ADFA-5039: Convert kotlin-web-site docs to JSON

[alexmmiller]

Summary

Adds md_to_json.py, which converts a kotlin-web-site/docs checkout
(JetBrains Writerside-flavored Markdown) into the JSON block schema this
project's templating engine renders.

Scope

This is split out of the larger end-to-end Kotlin-docs pipeline in
#21,
which is being separated into two ticket-scoped PRs:

• ADFA-5039 (this PR): producing the JSON data for the Kotlin website.
• ADFA-4739 (follow-up PR): build_nav.py, find_missing_assets.py,
  populate_db.py, media insertion, and the GitHub Action that loads
  everything into documentation.db. That PR depends on this one merging
  first, since populate_db.py/build_nav.py import md_to_json.py.

Changes

• ProcessDocs/ProcessKotlinDocs/ProcessKotlinWebsiteJSON/md_to_json.py - the converter.
• config.json - theming config it takes as input.
• markdown-it-py added to root requirements.txt.
• README.md scoped to this script's usage and output schema.
• review_build_json.sh - a throwaway reviewer helper (installs
  requirements, clones kotlin-web-site, runs the converter) so you can see
  real JSON output with no other setup. Not part of the actual pipeline.

Test plan

• Ran review_build_json.sh locally end-to-end against a real
  kotlin-web-site clone - converted 304/304 files, output matches the
  documented schema (topics/**/*.json, theme.json, images/).
• Reviewer confirms md_to_json.py/config.json match their originals
  in PR ADFA-4739: Pipeline for producing template-based Kotlin documentation #21 (no edits made during the split).

[luisguzman-adfa]

PR #23 — review comments

Ran md_to_json.py against a small fixture with intentionally broken references (and cross-checked with review_build_json.sh): it converts as expected and the output matches the documented block schema:
title + %var% substitution, cross-page link resolution (b.md#first-heading → /b.html#first-heading), external/broken-link coloring, image rewrite with {width=...} folding, and %var% inside code all come through, with the correct "unknown topic" warning for the deliberately-broken link.

Docstring and README are accurate and the "known limitations" are well-scoped.

Test block (Ubuntu)

# 1) get the repo on the PR branch
git clone https://github.com/appdevforall/OfflineDocumentationTools.git
cd OfflineDocumentationTools
git switch fix/ADFA-5039
cd ProcessDocs/ProcessKotlinDocs/ProcessKotlinWebsiteJSON

# 2) run md_to_json.py against a tiny fixture with intentionally broken refs
python3 -m venv /tmp/mdvenv && source /tmp/mdvenv/bin/activate
pip install markdown-it-py --quiet
D=$(mktemp -d); mkdir -p "$D/docs/topics" "$D/docs/images"
printf '<vars><var name="v" value="2.0.0"/></vars>' > "$D/docs/v.list"
printf 'PNG' > "$D/docs/images/m.png"
cat > "$D/docs/topics/a.md" <<'MD'
[//]: # (title: Sample %v%)
## First heading
Para with [x-link](b.md#first-heading), [external](https://x.com), [broken](missing.md).
![alt](m.png){width="200"}
```kotlin
fun main() { println("%v%") }
```
MD
printf '[//]: # (title: B)\n## First heading\nhi\n' > "$D/docs/topics/b.md"
python3 md_to_json.py "$D/docs" "$D/out" config.json && python3 -m json.tool "$D/out/topics/a.json"

Decision points

1. Heading ids vs. the anchors links point at. slugify() derives ids from heading text while resolve_href keeps the source #anchor verbatim. Simple headings line up (First heading → first-heading matched its link), but if Writerside's anchor algorithm diverges from slugify on headings with inline code/punctuation, or on duplicate headings, a link resolves to the page but lands on no anchor, silently. Worth spot-checking those cases against a real cross-reference.

2. Tests. No automated test for md_to_json.py (only review_build_json.sh). The block schema is the contract the templating step depends on, so a small fixture + a check on the emitted blocks would guard it, or, if that's landing with ADFA-4739, a one-line note saying so.

Suggestion

3. parse_attrs blanks a bare attribute value when the tag also has a quoted one: parse_attrs('group-key=gradle title="Gradle"') → {'group-key': '', 'title': 'Gradle'}. The '="' in attr_str check is whole-string, so any quoted attribute empties every bare one. Only fires for a block-level <tab>/<tabs> with a bare attr (Writerside usually quotes everything), so low-trigger, but real, deciding per-pair (use the bare group when the quoted one didn't match) fixes it.

Minor suggestions

4. slugify doesn't de-dup ids — two headings with the same text share an id, so an anchor to the second lands on the first. The usual -2/-3 suffixing handles it.

5. Broken/external coloring appends style="color: …"; on an <a> that already has a style that's a duplicate attribute (first wins in HTML5), so those links won't be colored — and the color is injected unescaped, so a hex check in load_config would fail a typo loudly.

6. page_id isn't posix-normalized (str(rel.with_suffix(""))), while build_topic_index normalizes backslashes, so the id (and the resolved /….html path) come out with backslashes on Windows. .as_posix() keeps them consistent; no-op on Linux/CI.

7. README usage shows --topics-subdir but not --images-subdir (both exist).

[hal-eisen-adfa]

Deep review of md_to_json.py and friends. 15 findings, left inline.

These aren't read-the-code guesses — I sparse-cloned JetBrains/kotlin-web-site (304 topics + v.list) and ran this converter over the whole corpus, so most findings below come with real file:line citations and occurrence counts from actual Kotlin docs.

Three separate content-loss bugs, all confirmed against the live corpus:

1. TAG_RE eats <table> (line 84) — <table> appears 38 times in the source and 0 times in the generated JSON. 76 table lines across 18 files are silently consumed.
2. merge_attr_lines deletes any {...} paragraph (line 427) even when it recovers zero attributes — the content is dropped and nothing is gained.
3. <tabs> discards all non-tab children (line 469) — intro and trailing prose inside a tabs block vanish with no warning.

Findings 2 and 5 compound badly: { style = "note" } (spaces around =) parses to no attributes and gets swallowed, so tour/kotlin-tour-welcome.md — the first page of the Kotlin tour — renders its note as a plain blockquote. 7 files hit this.

Also flagging line 278 as a markup-injection hole: broken-ext-link-color is interpolated into an HTML attribute unvalidated. Config is repo-controlled so it isn't remotely exploitable, but it's a one-line fix.

Please note this file is byte-identical to the md_to_json.py in #21 — git diff between the two branches is empty for it. Findings on lines 84, 142 and 278 duplicate comments I left there. Fixing once fixes both, but whichever PR merges second will carry them if only one gets patched. Worth deciding how #21 and #23 relate before either lands.

In fairness, a number of things I went looking for turned out to be fine, and I want to be explicit so this doesn't read as a wall of doom: no _raw leakage into the output JSON (0 corpus-wide), no tag_marker leakage (0), no empty-string heading ids (0), no IndexError on empty table cells, no blank-line stripping in raw <pre> blocks, and the unpinned markdown-it-py is genuinely safe (1.1.0 already exposes Token.attrs as a dict). The overall shape of the converter is sound — the damage is concentrated in the Writerside tag/attribute parsing.

[alexmmiller]

Pushed 6b25e26 addressing both reviews. Replied inline to each of the 15 line comments individually; summary + the two review-level threads here.

Fixed (all 12 correctness findings)

Every "blocking" item from @hal-eisen-adfa's review is fixed, plus @luisguzman-adfa's suggestion (#3, same bug as the line-142 finding) and minor points (#4 slugify de-dup, #5 duplicate style attr, #6 path separator, #7 README --images-subdir). Each is verified with a direct unit test in the new tests/test_md_to_json.py (30 cases), and re-confirmed against a real kotlin-web-site corpus run via review_build_json.sh where the original finding cited corpus numbers.

One more bug found during verification (not flagged by either review)

While re-running the corpus to confirm the fold_image_attrs fix, I noticed heading lines with a trailing Writerside attribute suffix - most commonly ## Title {id="custom-anchor"}, also seen as {completion-point=...} - were rendering that suffix as literal visible garbage text, with zero attribute handling. Same bug family as the image-attrs findings, just a different injection point the corpus checks happened not to exercise. Affected ~39 pages. Fixed the same way: Converter.extract_trailing_attrs strips it and, when an explicit id is given, uses it to override the auto-slug (which is the actual point of that syntax - a stable custom anchor) rather than just discarding it.

Decision point #1 (heading id vs. anchor mismatch)

Didn't fix - it's not a demonstrated bug, no concrete repro was given, and reproducing Writerside's own anchor-slugging algorithm to compare against is a bigger investigation than this PR's scope. Documented as a known limitation in both the module docstring and README instead, so it's tracked rather than silently dropped.

Decision point #2 (tests)

Added - tests/test_md_to_json.py, one case per fix above plus a few backward-compat checks, several reusing your own repro snippets directly (the getting-started.md:89 two-brace-group case, the missing-</tab> case, the {a.length}/{it.length} case). Run with:

uv run --with pytest --with-requirements ../../../requirements.txt python3 -m pytest tests/

On #21/#23/#24

@hal-eisen-adfa flagged that md_to_json.py is byte-identical between this PR and #21, so the same findings apply to both. That's expected - #21 is the original, unsplit branch (fix/ADFA-4737) that #23 and #24 (fix/ADFA-4739) were carved out of; #21 stays open for now purely as a reference for that split and will be closed as superseded once both land, never merged itself. So fixing it here is the right (and only) place - #21 inherits nothing further from this point on.

Worth flagging for @hal-eisen-adfa and @luisguzman-adfa specifically: build_nav.py, find_missing_assets.py, and populate_db.py (which import this file directly) live in #24, not here - if either of you is planning a similar pass over those, they depend on this PR's version of md_to_json.py/build_topic_index/build_image_index, which is why I kept both functions' return signatures unchanged even while fixing their internals (see the CONTAINER_TAGS/dead-code reply for specifics).

[hal-eisen-adfa]

Round 2 — all 15 previous findings verified fixed; 5 new ones introduced by the fixes

I re-ran the converter over a fresh 304-topic kotlin-web-site corpus and ran your test suite (30/30 pass, 304/304 converted, exit 0). Every one of the 15 findings from my last pass is genuinely fixed — I verified against real output, not just the diff:

Finding	Evidence
TAG_RE ate <table>	30 in source → 30 in output (was 0); orphan "type":"tab" blocks: 0
merge_attr_lines deleted {...} paragraphs	guarded on non-empty parse_attrs
<tabs> dropped non-tab children	whatsnew24.md intro prose survives
{ style = "note" }	kotlin-tour-welcome.json now emits attrs: {"style": "note"}
bare attrs blanked	411 populated width="N", 0 blanked; group-key populates
image attr braces leaked	15 → 0 genuine
color injection / duplicate style=	validated at load_config, merged not appended
main() exited 0 on failure	exit 1 confirmed, --allow-failures works
src= on every element	the 8 remaining warnings are all genuine missing .png — no false positives
Windows separators	as_posix() throughout
image re-copy / stale JSON	ctime unchanged on 2nd run; stale JSON pruned

Two things I want to be explicit about, because you pushed back and you were right:

• I checked your "the 2 residual {width= hits are inside <!-- comments" claim. It holds — both are in commented-out markdown.
• I checked converter.warnings / build_image_index's tuple against #24. find_missing_assets.py does read both. That's not dead code and I was wrong to call it that — keeping the signatures stable was the correct call.

Also: finding the heading {id=...} suffix bug yourself, during verification, on a corpus path neither reviewer's checks exercised, is exactly the right instinct. ~39 pages is not a small catch.

---

The 5 new findings

All five are introduced by this round's fixes — none existed in 7f67b2fb. Two are serious. The pattern worth noting: three of the five are cases where the fix is correct in exactly the case the new test exercises and wrong just outside it.

1. group_containers now crashes on a top-level unmatched closer (line 588) — dropped len(stack) > 1 guard, KeyError.
2. rmtree can delete the source docs (line 812) — no output_dir != docs_root guard. Reproduced.
3. The style= merge emits invalid CSS (line 352) — no ; separator; both rules get dropped.
4. fold_image_attrs eats the separating space (line 382) — minor.
5. COLOR_RE.match raises TypeError on a non-string (line 741) — minor.

Details inline. 1 and 2 are each a couple of lines; I'd want both before this lands, especially since #24 wires this into CI. Everything else here is in good shape — this was a thorough pass and the corpus numbers back it up.

Resolving all 15 previous threads.

[alexmmiller]

Pushed 5ec1f60 addressing round 2. Replied inline to each of the 5 new line comments individually; summary here.

Fixed (all 5 round-2 findings)

1. group_containers crash on a top-level unmatched closer (line 588) - restored the len(stack) > 1 guard dropped during round 1, so stack[-1]["type"] is never evaluated against the bare root frame (which has no "type" key). Verified '</note>\n\nAfter.\n' no longer raises and still emits the kind: "tag" warning.
2. rmtree could delete the source docs (line 812) - added an explicit topics_out_dir.resolve() == topics_dir.resolve() guard before the rmtree, exiting 1 with a clear error instead of deleting docs_root's topics/.
3. Style-merge produced invalid CSS (line 352) - used your suggested sep logic verbatim: only inserts "; " when the existing style value is non-empty and doesn't already end in ;.
4. fold_image_attrs ate the word-boundary space (line 382) - the remainder is no longer re-.strip()'d after slicing; only the truthiness check strips, the content itself keeps its leading space.
5. load_config raised a bare TypeError on a non-string color value (line 741) - added an isinstance(config[key], str) check ahead of the regex match so it hits the intended clean error message instead.

Tests

One regression test per fix, plus tightened test_fold_image_attrs_preserves_trailing_prose to assert the exact string (" Slack:", not the .strip()'d "Slack:") - per your own note that the old assertion wouldn't have caught #4 either way. 34/34 tests pass (29 existing + 5 new/tightened).

[hal-eisen-adfa]

Code review

Found 3 issues:

1. Self-closing container tags open a container that never closes, silently nesting the rest of the page inside it. In TAG_RE, the greedy ([^>]*) absorbs the trailing / before the optional /? can match, so group 1 (closing) stays empty and <tab title="A"/> is recorded as an opening marker. Reproduced against HEAD: <tab title="A"/> followed by a paragraph yields [{"type": "tab", "blocks": [{"type": "paragraph", ...}]}] with no warning emitted. No test exercises a self-closing tag. This was flagged on PR ADFA-4739: Pipeline for producing template-based Kotlin documentation #21 against the byte-identical file (discussion_r3732882620, "A self-closing tag opens a container that never closes, swallowing the rest of the page") but was not carried over into this PR.

OfflineDocumentationTools/ProcessDocs/ProcessKotlinDocs/ProcessKotlinWebsiteJSON/md_to_json.py

Lines 116 to 120 in 5ec1f60

	# "table", consuming "<table>" as tag "tab" with attrs "le".
	CONTAINER_TAGS = {"tabs", "tab", "note", "tip", "warning"}
	TAG_RE = re.compile(r"^<(/?)(" + "|".join(CONTAINER_TAGS) + r")(?![\w-])([^>]*)/?>$", re.I)

2. Indented (4-space) code blocks are emitted as raw unescaped HTML instead of code. convert_node handles markdown-it's fence token but has no case for code_block, which CommonMark produces for indented code. Those fall through to the generic fallback below and become {"type": "html", "html": <raw content>} — so code containing < or & (e.g. List<String>) is passed through as markup rather than displayed as code, and the block loses its code typing. The string code_block does not appear anywhere in md_to_json.py or in tests/test_md_to_json.py.

OfflineDocumentationTools/ProcessDocs/ProcessKotlinDocs/ProcessKotlinWebsiteJSON/md_to_json.py

Lines 552 to 557 in 5ec1f60

	# Fallback: anything not explicitly handled (images are inline-only,
	# so plain "image" blocks don't occur at block level; captured via
	# paragraph HTML instead).
	return {"type": "html", "html": self.rewrite_urls(str(t.content or ""))}

3. The documented image block type is never emitted, so the published schema is wrong. No code path in convert_node/convert_nodes constructs {"type": "image", ...} — markdown-it only ever produces image as an inline token nested in a paragraph, which render_inline turns into an <img> inside that block's html string. The fallback's own comment states this ("images are inline-only, so plain "image" blocks don't occur at block level; captured via paragraph HTML instead"), contradicting the docstring schema below and the matching image entry in README.md. Since producing this JSON schema is the stated deliverable of this PR, a template author on the ADFA-4739 follow-up would write dead handling for a block that never appears.

OfflineDocumentationTools/ProcessDocs/ProcessKotlinDocs/ProcessKotlinWebsiteJSON/md_to_json.py

Lines 39 to 43 in 5ec1f60

	{"type": "list", "ordered": false, "items": [{"blocks": [...]}]}
	{"type": "table", "headers": ["a", "b"], "rows": [["1", "2"]]}
	{"type": "image", "src": "...", "alt": "..."}
	{"type": "hr"}
	{"type": "tabs", "attrs": {"group": "build-system"},

Frequency note, so these can be triaged rather than taken at face value: against a kotlin-web-site clone at 5aff3dc (298 files under docs/topics), issue 1 has 0 occurrences and issue 2 has 1 — in async-programming.md, whose content contains no < or & and so is not currently mangled. Both are latent rather than actively breaking output today. Issue 3 affects the contract the follow-up PR builds against.

Also confirmed: all round-1 and round-2 review findings are genuinely fixed at 5ec1f60a, re-derived from the current file rather than from the "fixed in X" replies.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[alexmmiller]

Pushed 5336ee5 addressing all 3.

1. Self-closing tags (TAG_RE) - fixed by splitting the self-close marker into its own capture group (([^>]*?)\s*(/?)>$ instead of a single greedy ([^>]*)/?>$), so it can no longer be swallowed by the attrs group. convert_node's html_block dispatch now emits an immediate open/close pair for a self-closing tag instead of a bare opener. Verified <tab title="A"/> followed by a paragraph no longer nests that paragraph inside the tab, and confirmed <table>/</table> still correctly don't match. Covered by test_tag_re_captures_self_closing_marker_separately, test_html_block_self_closing_tag_emits_open_and_close_markers, and test_self_closing_tag_does_not_swallow_trailing_content.
2. Indented code blocks - added a code_block case to convert_node alongside the existing fence one, emitting the same {"type": "code", ...} shape. Verified List<String> x = ... (4-space indented) now comes out typed as code with the raw text intact, not as an unescaped html block. Covered by test_indented_code_block_renders_as_code_not_html.
3. Documented image block type - agreed this was a docs-vs-reality mismatch rather than a code bug, so fixed the docstring and README instead of adding new "promote an image-only paragraph to a block" behavior that could change what ADFA-4739's templates receive. Both now state explicitly that images are always inline content inside their containing block's html, never a standalone block.

38/38 tests pass (34 previous + 4 new).

[hal-eisen-adfa]

Automated review at high effort, run against the live corpus rather than the diff alone: I sparse-cloned JetBrains/kotlin-web-site (303 topics, 7,863 links), ran the converter over it, and audited the output. Every finding below is reproduced.

Baseline health is good — 303/303 topics convert with only upstream-broken-link warnings, no schema leaks (_raw, tag_marker, and tabs-without-tabs are all zero), and the 38-test suite passes.

1 HIGH, 3 MEDIUM, 7 LOW inline. The HIGH is the one that matters: slugify() strips punctuation instead of hyphenating it, which breaks 209 real cross-references. The rest are mostly latent or documentation-contract issues.

Things I specifically tried to break and couldn't: all four bug classes fixed across the earlier review rounds hold up on the live corpus — TAG_RE's (?![\w-]) correctly rejects <table>, self-closing <tab/> yields a balanced marker pair, merge_attr_lines preserves { it.length }-style paragraphs (1 real occurrence, in functions.md), and the style-attribute merge produces valid CSS. A 32-vs-31 heading discrepancy I chased in compatibility-guide-17.md turned out to be a heading inside an HTML comment — correct behavior, not a loss.

[alexmmiller]

Pushed 4fddb34 addressing all 11 findings from the latest automated review. Replied inline to each thread individually; summary here.

HIGH (1)

1. slugify() deleted punctuation instead of hyphenating it (line 211) - measured at 209 dead #anchor links on the live corpus. Fixed using your suggested regex; verified all four of your repro examples now match their linked anchors exactly.

MEDIUM (3)

2. fold_image_attrs swallowed a {...} group that parsed to nothing (line 393) - folded_any now only sets on a successful parse_attrs, mirroring the guard merge_attr_lines/extract_trailing_attrs already use.
3. {...} after a link rendered as literal text (line 428) - extended fold_image_attrs to fold a trailing group onto the matching link_open's attrs (via a new _find_link_open helper), not just after images/headings.
4. A container left open at EOF warned nothing (line 608) - group_containers now checks len(stack) > 1 after the loop and warns, symmetric with the existing unmatched-closer branch.

LOW (7)

5. .md links with a path prefix neither resolved nor warned (line 281) - MD_LINK_RE now absorbs a path prefix and resolves by bare filename, same as build_topic_index.
6. Image srcs with a path component silently dropped (line 295) - resolve_image_src now strips to bare filename before lookup; also dropped the now-genuinely-dead startswith("/") duplicate check.
7. Unbalanced raw-HTML fragments (line 536) - documented as an explicit contract in the module docstring rather than changed, per your own framing.
8. Blank lines inside a raw HTML run dropped (line 572) - elif line.strip(): → else: so they're preserved in the joined output.
9. main() wrote into the source before validating it wasn't the source (line 828) - the aliasing guard now runs before theme.json/images/ are touched, not just before the topics rmtree.
10. Stale images never pruned (line 833) - added _prune_stale_files, run after the copytree, keeping the _copy_if_changed mtime-skip optimization intact (a pruning pass instead of switching to rmtree-then-recopy).
11. markdown-it-py had no version floor (requirements.txt:8) - pinned >=2.0. Flagged the stale ADFA-4739 comment you mentioned for whoever's on that branch, since it's a different file I can't fix from here.

Tests

13 new regression tests, several reusing your exact repro snippets/corpus examples. 55/55 pass (42 previous + 13 new).

[alexmmiller]

Full code review of ProcessKotlinWebsiteJSON/, and six fixes

Reviewed the whole directory at 4fddb348. The code is solid — 55/55 tests passed, ruff check --select=E9,F,B,SIM,C4 clean, and against a fresh kotlin-web-site clone it converted 303/303 pages in 1.1s, byte-for-byte deterministic across runs, with stale-output pruning working for both topics/ and images/ (including now-empty directories).

Six defects found and fixed in 1d7f6a7e. All six are latent — I checked each against the live corpus and none of them fire there.

The fixes

#	Defect	Fix
1	An explicit {id="..."} only beat a later auto-slug. In the other order the auto-slug claimed the id first and both headings shipped the same id, so any link to the explicit one landed on the auto one.	New reserve_explicit_heading_ids pass registers every explicit id before any heading converts, so explicit wins in both directions. Scoped to inlines directly inside a heading — a paragraph merely ending in {id=...} is not an anchor. Two genuinely identical explicit ids are a source bug this can't fix, so both are kept as authored and it warns.
2	A punctuation-only heading (## ...) slugified to "" — id="" is invalid HTML and unlinkable.	Falls back to section, numbered by the existing de-dup loop.
3	extract_title matched a [//]: # (title: ...) comment inside a fenced code block, both setting a bogus title and deleting that line out of the code sample being displayed.	New fenced_spans skips matches inside ``` / ~~~ fences — handles tilde fences, longer fences nesting shorter ones, and an unterminated fence running to EOF.
4	A raw html_block run of nothing but blank lines emitted {"type": "html", "html": ""} (a truthy list of empty strings).	flush_raw requires actual content. Blank lines within a run that has content are still preserved, which is what keeps <pre>/<script> bodies intact.
5	COLOR_RE's flat {3,8} accepted #12345 and #1234567, which no browser does.	Narrowed to the lengths CSS defines. This value is interpolated into a style="" attribute, so the validator should mean what it says.
6	build_tree popped an already-empty stack on a stray closing token — IndexError: pop from empty list with nothing naming the cause.	Ignores the stray closer and warns. Defensive; markdown-it always balances.

How I verified these are safe

Since every finding is latent, the meaningful check wasn't the unit tests — it was a differential corpus run. I converted all 303 pages with the converter at 4fddb348 and again with the fixed one, and diffed:

• 0 differing output files
• identical 57-warning set

So these change nothing about today's output; they only close cases the corpus doesn't currently contain.

On those 57 link warnings, for the record: they're 27 distinct references, and I checked every one against the checkout. All are genuinely absent from topics/ — cross-repo links into kotlinx.coroutines and dokka. No false positives from the resolver.

Docstring

The "Block shapes" list omitted note/tip/warning, which group_containers genuinely emits (the README already documented them, and points readers at the module docstring "for full shapes"). Now added.

Also recorded, after testing it: a <tab> written outside any <tabs> comes through as a bare {"type": "tab"} block that a template written against the documented shapes won't render — and it does not warn. I documented the actual behaviour rather than changing it, since that's beyond the scope of this review. Worth filing separately if you want it to warn.

Tests: 55 → 91, coverage 66% → 81%

Beyond one regression test per fix, this closes the coverage gaps the review turned up — features that had no test at all:

• %variables% substitution (substitute_vars / load_variables) — a documented feature that reaches every rendered string.
• Block-level <note> — covered by nothing. Note the corpus can't cover it either: all 5 <note> uses in kotlin-web-site are the single-line form inside releases.md's hand-written HTML tables, which is the documented passthrough case and behaves correctly.
• Both <tabs> fallback paths — synthesising tabs from adjacent code fences, and dropping a wrapper with nothing tab-like inside. Each had previously fixed a content-dropping bug, and neither was tested.
• Direct tests for the blockquote/list/table/hr converters, and a guard that the internal _raw key never reaches the JSON.

(Coverage is under-reported either way — main() is exercised by 10 subprocess tests that coverage can't see.)

Note for #24

fix/ADFA-4739 had a stale copy of md_to_json.py — it merged this branch before 4fddb348 landed. I've synced it (80e234a9); md_to_json.py and its tests are now identical on both branches.