Follow-up to #30 (the ask-goldmark span rewrite, now on main). The design already commits to this — docs/design.md, Dialects section — but it was deliberately deferred to its own session rather than ridden along with the core.
What
Under --dialect mkdocs, the goldmark configuration should drop the extension.Linkify extension. Python-Markdown (the mkdocs target renderer) has no GFM autolinking, so a bare URL is plain prose there. Because #30 made the span computation and the render backstop both read from the profile's goldmark config, dropping linkify makes bare URLs automatically stop being treated as autolinks under this dialect — no per-dialect guard logic anywhere.
Dropping linkify cannot split a URL in any mode: every break candidate is a whitespace run and a bare URL contains none, so reflow only moves the whitespace around the token. The only behavior linkify governs in span computation is code-span pairing around a backtick inside a bare URL — near-nonexistent in real prose.
Where the profile choice has to reach
internal/gm.New (document parse) — format.go, internal/render/render.go, and reflow's line reparse (internal/reflow/reflow.go).internal/gm.NewInline (the no-break span profile) — internal/segment/nobreak_gm.go.internal/render.Normalized (the render backstop comparison must use the same profile).segment.NoBreakSpans / segment.CodeSpans — currently take only text string and are called from ~7 sites in reflow.go with no dialect parameter.
Suggested precursor
Do the design's "compute no-break spans once per cluster and share them" refactor first (design.md, "No-break spans: ask goldmark" — the cost/sharing note). That collapses the ~7 recomputing call sites into one, so the dialect/profile choice threads to a single place instead of spreading a new parameter across the segment API.
Verification
Our render backstop only measures self-consistency under goldmark's semantics; it cannot see goldmark-vs-Python-Markdown divergence. The real oracle for this change is a full mkdocs build diff on a real docset (Karl Isenberg's workflow). Land it when that verification is available.
Also note #26: deriveOptions never exercises the mkdocs dialect (hash bits 16/17 free), so the fuzzer does not currently cover mkdocs at all — worth wiring up alongside this so the profile split gets fuzz coverage.
Follow-up to #30 (the ask-goldmark span rewrite, now on
main). The design already commits to this —docs/design.md, Dialects section — but it was deliberately deferred to its own session rather than ridden along with the core.What
Under
--dialect mkdocs, the goldmark configuration should drop theextension.Linkifyextension. Python-Markdown (the mkdocs target renderer) has no GFM autolinking, so a bare URL is plain prose there. Because #30 made the span computation and the render backstop both read from the profile's goldmark config, dropping linkify makes bare URLs automatically stop being treated as autolinks under this dialect — no per-dialect guard logic anywhere.Dropping linkify cannot split a URL in any mode: every break candidate is a whitespace run and a bare URL contains none, so reflow only moves the whitespace around the token. The only behavior linkify governs in span computation is code-span pairing around a backtick inside a bare URL — near-nonexistent in real prose.
Where the profile choice has to reach
internal/gm.New(document parse) —format.go,internal/render/render.go, and reflow's line reparse (internal/reflow/reflow.go).internal/gm.NewInline(the no-break span profile) —internal/segment/nobreak_gm.go.internal/render.Normalized(the render backstop comparison must use the same profile).segment.NoBreakSpans/segment.CodeSpans— currently take onlytext stringand are called from ~7 sites in reflow.go with no dialect parameter.Suggested precursor
Do the design's "compute no-break spans once per cluster and share them" refactor first (design.md, "No-break spans: ask goldmark" — the cost/sharing note). That collapses the ~7 recomputing call sites into one, so the dialect/profile choice threads to a single place instead of spreading a new parameter across the segment API.
Verification
Our render backstop only measures self-consistency under goldmark's semantics; it cannot see goldmark-vs-Python-Markdown divergence. The real oracle for this change is a full
mkdocs builddiff on a real docset (Karl Isenberg's workflow). Land it when that verification is available.Also note #26:
deriveOptionsnever exercises the mkdocs dialect (hash bits 16/17 free), so the fuzzer does not currently cover mkdocs at all — worth wiring up alongside this so the profile split gets fuzz coverage.