Known Inconsistencies

MarkedInput now renders as one editing host. The container is the only contenteditable; everything below it either inherits editability or is explicitly atomic:

<div contenteditable="true">
  <!-- THE editing host. Flips to "false" under readOnly -->
  <span>text</span>
  <!-- bare text surface: no attribute, inherits editability -->
  <mark contenteditable="false">value</mark>
  <!-- value-only mark: atomic by contract, no tabindex -->
  <span>text</span>
  <span>
    <!-- slot mark: root stays BARE -->
    <i contenteditable="false">control</i>
    <!-- only the controls around the slot are frozen -->
    <span>slot text</span>
    <!-- slot content lives in the ONE host -->
  </span>
  <span>text</span>
</div>

Who writes what: bind.ts writes the per-token topology (bare text surfaces, ce=false value marks, frozen controls around a slot host); SelectionDriver owns the container’s contenteditable, gated only by readOnly; TokenModel.control() marks registered controls (grips, menus, overlays) ce=false at registration, because a control is not document content.

Every input still goes through the guard: beforeinput is cancelled and lowered to a model edit (keyboard/beforeInput.ts + keyboard/input.ts), which fails closed — an input type the guard cannot express as an edit is dropped rather than left to mutate the host.

Measurement scope. The rows below were measured in Chromium (live storybook, trusted events) on 2026-08-11 for the pre-migration state and 2026-08-12 for the post-migration state, plus the core/adapter suites. Chromium-only is a maintainer decision (2026-08-10); Firefox and Safari users of the published packages exist and are not covered by any of these measurements.

---

Resolved by the single-host migration

Defect (pre-migration, measured)	Post-migration state
Click steal (never filed — found while re-measuring): clicking into a non-focused span never moved the caret, because the model re-applied stored anchors and stole focus back before Chromium placed the click. It also ate triple-click on any non-focused span.	Gone — resolved before it was ever documented. Its two sources (per-span focus + the focusin selection sync) no longer exist. Clicks into the first three spans land where clicked, every time.
Cross-mark delete only touched the originating span. After the rewrite it was worse: a finished cross-mark drag left every host ce=false and focus on BODY, so Backspace was silently dropped.	Works. A sweep from span 1 across a mark into span 2 followed by Backspace deletes the text and the mark, in one guarded edit.replace. The whole drag-time editability flip is deleted — one host has nothing to escape.
Tab trapped in the field, cycling <span contenteditable> and tabindex=0 marks (post-rewrite it was fully dead: focus reached the mark and was snapped back).	Tab leaves the field, natively — mark roots lost tabindex. Pinned by keyboard.react.spec / keyboard.vue.spec (“leaves the field on Tab”). Caveat, measured: a consumer whose Mark renders a real <button> still gets a tab stop — ce=false does not remove a native button from tab order. That is the consumer’s element, not markput’s.
Home/End trapped at span offset 0 / span end.	Full visual line. One host means one line box for the browser’s own line motion.
Word navigation (Alt/Ctrl+Arrow) stopped dead at the span boundary.	Crosses marks. Alt+ArrowLeft word-jumps across marks; repeated presses keep moving.
Shift+Arrow selection could not cross a mark (it grew to the span end and stuck).	Expected resolved by topology — not separately measured in the post-migration sweep. Shift+Arrow is plain native selection extension inside one host, and the plain-arrow and drag-sweep equivalents were both verified; treat this row as inference, not measurement.
Drag selection did not cross marks (mid-drag it did, via the ce=false flip, but mouseup left the editor dead).	Works, with no flip machinery. The native sweep crosses ce=false atomics inside one host; focus never leaves the container, and the resulting selection is an ordinary editable range.
Focus/blur churn: every click on a different span fired focusout+focusin on the container (3 clicks + 1 exit → 3× focus / 4× blur).	Single entry/exit. The container is the only focusable element markput writes, and the focusin listener is deleted; only focusout remains. Derived from the topology and the deleted listener — event counts were not re-recorded.

---

Still open

Triple-click selects the mark-bounded run, not the line

Chromium bounds paragraph selection at any inline contenteditable=false atomic. This was control-measured without markput: a plain <div contenteditable> containing a bare inline <span contenteditable="false"> shows the same bound. It is an engine limit, not a markput DOM-topology problem, and nothing in the migration could have fixed it. No workaround ships: faking it would mean owning paragraph selection ourselves.

Empty-gap mouse targeting

The parser guarantees an empty TextToken('') between two adjacent marks, rendered as a bare <span></span>.

• Arrow-reachable: one caret stop, container-anchored between the two marks. anchorFor resolves it.
• Not mouse-reachable: on the live DOM that span computes to 0.0px wide, so there is no pixel to click — clicks at the gap midpoint land on a neighbouring mark instead. (The design doc’s “5/5 clicks land on the gap span” came from a static fixture and does not transfer; see the addendum in docs/records/one-host-migration.md.)

A ZWSP or min-width filler remains rejected: a filler is real text in the DOM, so it contaminates range.toString() on every copy, and it costs a second arrow stop at every gap. Reach the gap with an arrow key.

A click on a mark’s own presentation lands at the mark’s nearest EDGE

A consumer’s mark styling can own pixels no token’s text covers — the markdown list preset is display: block; padding-left: 1em, so the 1em band left of the first glyph belongs to the mark’s element and to no text. Chromium hit-tests a click there to that element, and the model owns no position inside a mark’s presentation, so the boundary answers the mark’s near edge instead. Typing then inserts outside the mark: on - `Code snippets` and `code blocks` a click in the padding plus X yields XCode snippets and code blocks — the row un-lists.

Until 2026-08-22 that same click dropped the keystroke entirely: the boundary’s island guard read the inherited isContentEditable, which a slot mark’s bare root and every element below it get from the container, so it declined, domAnchors() declined with it, and the fail-closed beforeinput guard cancelled the key with no model edit. Reading the contentEditable property instead — a consumer’s explicit island and nothing else — is the fix, and the near edge is the documented fallback it uncovers.

Whether “start of the slot content” would be a better answer than “before the mark” is undecided; it is a change to nearestMarkEdge, not to the guard.

Native undo/redo is dead

Cmd+Z / Ctrl+Z does nothing — for mark deletion and for plain typing. Two independent reasons, both measured:

1. Every edit is preventDefault() + a model write, so the browser’s undo stack is empty — there is nothing to undo.
2. historyUndo / historyRedo arrive as beforeinput types the guard cannot express as an edit, so the fail-closed guard cancels them.

Measured identical before and after the migration; the topology never had anything to do with it. The fix is an editor-owned history, which is a separate feature. (The old page’s “basic undo within a span works ✅” was false and has been removed.)

IME / composition is unhandled

insertCompositionText is not cancelable, so the guard lets it through by design — composition writes to the host directly and the model reconciles afterwards. Unhandled by design; needs its own design pass.

Markup injection via paste

Pasting the literal text @[injected](primary:99) creates a mark instead of inserting plain text. Mechanism is unchanged by this migration: the paste is intercepted, lowered to a string splice, and the spliced window is re-parsed (keyboard/input.ts → tokens/tree/transactions.ts → valueBoundary.ts), so pasted markup is markup.

No focus ring

outline: none is set on the editor’s surfaces in packages/core/styles.module.css; nothing draws a focus indicator in its place. Unchanged by this migration.

No ARIA role or accessible name

The container carries no role, no aria-multiline, no label. A screen reader does not announce it as an editable text field. Unchanged by this migration — but the container is now the single focusable editing host, i.e. the natural carrier for role="textbox" + aria-multiline="true". That is a scoped follow-up task, not a topology problem.

Chromium-only scope

All behavior above was measured in Chromium only (maintainer decision, 2026-08-10). Firefox previously measured materially worse on gap carets and edge positions. Firefox and Safari users of @markput/react / @markput/vue exist; their behavior is unverified.

---

Works correctly

Re-verified subset only — each row is either in the post-migration live sweep or pinned by a suite. Rows that were never re-verified are not listed.

Behavior	State
Typing and onChange	Characters insert at the caret; the value updates per keystroke.
Enter inserts \n	Enter is a newline in the value, not a DOM line break — the guard lowers insertParagraph/insertLineBreak to '\n', so the browser never builds a <div>/<br> inside the host.
Caret navigation across marks	Left/Right walk through a mark one position per keypress, natively — no dead stop, no double press, no custom arrow handler in the core at all.
Select-all, then type / paste / delete	Ctrl+A selects the editor’s whole contents (one host, one boundary) and the next input replaces the whole value. In block layout, Ctrl+A then Enter replaces the document with one fresh row — BREAKING, introduced by 603adfac.
Cross-mark selection delete / replace	A selection spanning marks deletes or is replaced as one guarded edit, including the marks it covers.
Copy	text/plain is the native range.toString() over the one host, so it spans marks; markput also writes text/html and its own markup MIME so a copy/paste round-trip keeps the marks.