Status: draft (2026-11-03), issue #127. Full design in the repo at docs/superpowers/specs/2026-11-03-server-comments-from-the-reader-parse-design.md. It follows the one-editor-dialect work released in 2.5.7, which moved the editor onto its syntax tree.
A %%-delimited Obsidian comment is hidden from readers, and where it opens and closes depends on code and math (a %% inside either is literal). The editor now takes code and math from its CommonMark tree. The server still uses a hand-written text mask in wikantik-api. Wherever that mask misreads block structure, the author sees text dimmed as hidden that readers are shown. The remaining shapes involve a wikilink holding a backtick, backticks in HTML or autolinks, escaped backticks, pipes inside a code span in a table cell, a table on a list item's first line, and display math in a list item.
The tag index has the same weakness: it finds inline #tag text with a scan that uses the same mask, while the page draws tags from flexmark's own nodes.
The server already has a parser that reads block structure exactly as readers see it, because it is the reader. Ask it instead of approximating it.
ReaderComments in wikantik-main is the only server implementation. It returns at once for a body with no %% (1,459 of 1,467 prod pages pay nothing). Otherwise it parses the raw body as the render does and takes code and math nodes as opaque spans. It then pairs markers with the editor's rule, using a binary search per marker. Display math takes its extent from the parsed node, not from the pre-pass record, so it ends with its list item as the render draws it. Every caller migrates: render, reader text, block ids, access-rule classification, references, embeds and the page API. If the parse is unreadable, it fails toward hiding and logs a WARN, and the page's access rules still lock.ReaderTags keeps the InlineTag nodes the renderer would draw, through one shared predicate, so the page and the index are one rule by construction. Query-time tag filters read the structural index's stored inline tags instead of re-scanning bodies. Then the hand masks are deleted: ObsidianComments, MathMask, QuotedFences, the two GFM table helpers, and the tag scanner with its masks.CodeMask. Their failure mode is a wrong Page Graph edge, not a leak, and they are measured at 51 characters off on the prod corpus. They get a follow-up issue.Expected values come from the rendered page, never from the code under test, and each test is shown failing against the current mask first.
comments.json, later inline-tags.json) run against the new server classes, with one new case per #127 shape.ReaderTags.%% costs zero extra parses and one with it costs one. A ratio-based linear-time guard (not milliseconds) covers comment- and math-dense bodies.Each phase is releasable alone. Phase B needs a reindex after deploy.