markdown: predicted heading slugs diverge from GitHub's actual anchors (emoji VS16 dropped, No-category characters kept), so internal links to those headings break #190

resolve_anchors binds a link to a heading-coincident anchor to the fragment the heading "will have in the rendered Markdown" (src/render/markdown/anchors.rs:1-5). For that binding to work, gfm_slug has to reproduce the anchor GitHub actually generates. It doesn't for two classes of characters, in opposite directions, and every internal link to such a heading lands on a fragment GitHub never assigned.

Divergence 1: variation selectors — anydoc drops them, GitHub keeps them

gfm_slug (src/render/markdown/anchors.rs:180) keeps only is_alphanumeric, the hardcoded mark ranges in is_combining_mark (anchors.rs:199), connector punctuation, and hyphens. U+FE0F (VARIATION SELECTOR-16) is in none of those, so it is filtered out. GitHub's anchor algorithm (github-slugger, which GitHub renders with) removes a fixed punctuation set; VS16 is not in it, so the anchor keeps the invisible selector.

Divergence 2: No-category characters — anydoc keeps them, GitHub drops them

² (U+00B2) and ½ (U+00BD) pass Rust's char::is_alphanumeric (category No is numeric), so gfm_slug keeps them. github-slugger removes both, so the GitHub anchor does not contain them.

Repro

Minimal .docx (styles.xml maps Heading1, bookmark inside each heading, w:hyperlink w:anchor=... after it), converted with 0.2.4 (commit 261fc25, identical to v0.2.4):

# ❤️ Love      → [go to heading](#-love)      GitHub anchor: #️-love   (fe0f kept)
# 2️⃣ Second   → [go to second](#2⃣-second)   GitHub anchor: #2️⃣-second (fe0f kept)
# x² formula   → [go to x](#x²-formula)       GitHub anchor: #x-formula (00b2 dropped)

Exact fragments from the conversion (hex-escaped):

anydoc:        2d 6c 6f 76 65            = "-love"
github-slugger: fe0f 2d 6c 6f 76 65      = "️-love"

anydoc:        32 20e3 2d 73 65 63 6f 6e 64      = "2⃣-second"
github-slugger: 32 fe0f 20e3 2d 73 65 63 6f 6e 64 = "2️⃣-second"

anydoc:        78 b2 2d 66 6f 72 6d 75 6c 61   = "x²-formula"
github-slugger: 78 2d 66 6f 72 6d 75 6c 61      = "x-formula"

vs github-slugger 2.0.0, which reproduces GitHub's anchors. Heading text such as ## ⚠️ Deprecation notice or ## ✨ What's new is common in real documents, so the wrong fragment is emitted for exactly the headings people decorate.

Verified against the npm build @firecrawl/anydoc 0.2.4 (source on main is unchanged: only README commits since v0.2.4).

Direction

gfm_slug's keep-list is close but not equal to github-slugger's remove-list. Matching github-slugger exactly for these two classes — keep U+FE0F (it is a mark GitHub's slugger preserves), and drop the No/Nl characters github-slugger removes (or, more robustly, derive the keep predicate from the same character set github-slugger removes) — would fix both directions. A differential test over a corpus of tricky headings against github-slugger would keep this property from regressing.