- looksLikeWikiNumberedList now checks for ANY multi-hash heading (##, ###,
  etc.) and requires consecutive # lines (no blank lines between them) to
  distinguish wiki numbered lists from markdown headings
- Add edge case tests: multiple h1 headings, h3 headings, tilde with numbers,
  punctuation-adjacent formatting, tab/newline whitespace

Replace goldmark extension.Strikethrough with hugo-goldmark-extensions/extras
which natively parses ~sub~, ^sup^, ~~del~~, and ++ins++ syntax. The ADF
converter now emits spec-compliant marks:

- ~text~ -> {"type": "subsup", "attrs": {"type": "sub"}}
- ^text^ -> {"type": "subsup", "attrs": {"type": "sup"}}
- ~~text~~ -> {"type": "strike"}
- ++text++ -> {"type": "underline"}

Previously, subscript/superscript/underline were converted to HTML tags
(<sub>, <sup>, <u>) by the wiki converter, which goldmark parsed as RawHTML
nodes that the ADF converter silently dropped.

Changes:
- shared/adf/convert.go: swap extension.Strikethrough for extras extension,
  add extrasKindToMark() mapping extras AST kinds to ADF marks
- wiki.go: remove sub/sup HTML conversion (goldmark handles natively),
  convert wiki +text+ to ++text++ for goldmark extras Insert extension
- Remove unused wikiSub/wikiSup regex patterns

- Document WikiToMarkdown as intended for MarkdownToADF/goldmark-extras
  consumption, not standalone markdown renderers (finding 2)
- Document punctuation-adjacent formatting limitation in
  replaceWikiFormatting comment (finding 3)
- Add test: adjacent h1 without blank line is intentionally treated as
  wiki numbered list (findings 1, 4)
- Add tests: WikiToMarkdown sub/sup passthrough and underline conversion
  for goldmark extras (finding 2)

- Expand comment on ## heuristic explaining Jira nested numbered list tradeoff
- Document intentional before/after boundary asymmetry (opening vs closing punctuation)
- Add test for superscript in compound word (x^2^y) to match existing tilde test
- Add test locking period-before-delimiter boundary behavior

The standard parser (for plain markdown) now uses extras.Delete only
(double-tilde ~~text~~ strikethrough). Subscript, superscript, and
insert are only enabled in the wiki parser, used when input is
detected as Jira wiki markup.

This prevents even-tilde compound words like "signal~webapp~frontend"
from being mangled by goldmark subscript processing in non-wiki input.

Also addresses review findings:
- Fix variable shadowing: rename loop var t -> s in end-to-end test
- Add ASCII delimiter assumption comment on replaceWikiFormatting
- Clarify ^ anchor scope in boundary pattern comments
- Add even-tilde and even-caret compound word tests

Coverage gaps filled:
- Shared-package parser split tests: ToDocument vs ToDocumentWiki vs ToJSON
  proving standard parser omits subsup/underline, wiki parser produces them
- Inline-only wiki formatting (H~2~O, x^2^, +important+) through
  MarkdownToADF: proves auto-detection correctly falls through to safe parser
- Nested wiki ## under # false negative test lock
- Square bracket boundary test ([-deleted-])

Fixes:
- Add ] to wikiBoundaryAfter so [-deleted-] converts correctly
- Strengthen WikiToMarkdown doc comment: explicitly notes pipeline coupling
  to adf.ToDocumentWiki and warns callers MUST use wiki parser

Adds contract note: auto prioritizes not corrupting plain markdown
over detecting wiki edge cases. Callers that know the input format
should bypass heuristics via WikiToMarkdown + adf.ToDocumentWiki.

Rename makes the pipeline coupling explicit: the output is not
general-purpose markdown but a dialect tuned for adf.ToDocumentWiki.

Strengthen CompoundWordsEndToEnd test to verify compound words appear
within single ADF text nodes rather than just checking concatenated text.

Also document ToJSON parser choice to match ToDocument/ToDocumentWiki.