Let a reader comment on a line that carries a reference

[saip-block]

Selecting a line to comment on it broke when the line carried a footnote citation. The hover card opened mid-drag, right over the text being swept — and if you got a selection out anyway, the server refused it with "Anchor text doesn't match the plan content — the comment would have nowhere to appear."

Three causes, none of them the same bug.

The preview opened mid-gesture

A held button means the reader is sweeping a selection across the line, not resting on a reference. Each entry point now reads that for itself, so there's no flag to leave stranded when a gesture ends without a mouseup — a link drag, a release outside the window. mouseenter carries the button state, and the focus Chromium fires on mousedown answers with :focus-visible, which a press never matches but keyboard and assistive-technology focus always do. That question is asked as the card is about to open rather than while the focus event is still dispatching, because Chrome settles :focus-visible afterwards and reading it inline intermittently turns a keyboard reader away.

The extractor read the document with the wrong grammar

MarkdownTextExtractor parsed with bare Commonmarker.parse while the renderer parses with footnotes: true. So the canonical text kept [^label] verbatim while the reader saw a superscript number, and no selection carrying that number could ever match. It now parses with the renderer's own EXTENSION_OPTIONS and emits the number the reader sees.

The digits are sentinel-mapped, like every other synthetic character here. That falls out nicely: a selection that merely ends on the marker resolves back to the prose before it rather than to half a citation.

Whitespace — and this one isn't about references at all

A browser hands back a selection with whitespace runs collapsed to a single space; a hard-wrapped paragraph is one flowing line on screen and the reader swept it as one. The extractor keeps the source's own newlines, so any selection crossing a wrapped line was refused, prose or not. Both sides now fold the same way, carrying the position map so ranges still land on real source positions.

Resolution and occurrence counting fold identically. They have to agree, or they disagree about which copy of a repeated phrase was selected and the highlight lands on the wrong one — nothing covered that before, so there's now an example with a repeated phrase spanning a wrap.

Verification

16 examples, each verified to fail when its guard is removed. The system example drives a real Selenium drag across the citation, through the comment form, to a placed anchor; reverting the extractor reproduces the reported error verbatim:

Anchor text doesn't match the plan content — the comment would have nowhere to appear
anchor: "This claim has evidence.1 See §2"

Full suite 1980 examples / 0 failures. RuboCop clean across 489 files. Also confirmed by hand in dev on a plan carrying both a citation and a section link.

---

Demos

Before

before-coplan.mov

After

after-coplan.mov