u2ui 1.5.3 → 1.5.7

package/js/rte/AGENTS.md

@@ -0,0 +1,142 @@
+# RTE Agent Instructions
+
+These instructions apply to this directory and all descendants.
+
+## Project status — important
+
+RTE is new and not used in production or by downstream consumers yet. There
+is no backward-compatibility obligation. Prefer the clean final API and remove
+superseded names or designs instead of adding aliases, deprecations, shims, or
+legacy layers. Existing behavior still needs tests, but compatibility with an
+earlier RTE draft is not a reason to keep a weaker design.
+
+## Start here
+
+- Read [`README.md`](./README.md) for the architecture and requirements.
+- Read [`PLAN.md`](./PLAN.md) for implemented phases and the next responsibility.
+- Read the README beside every responsibility you change.
+- Treat `../rte0` as a read-only behavioral reference. Never modify it or any
+  other file outside `rte` while working on RTE.
+
+## Priorities
+
+- Keep changes focused, compact, readable, and internally consistent.
+- Prefer a small general solution to a collection of special cases. RTE is a
+  generic engine and must support use cases that were not anticipated.
+- Treat reported examples as regression cases for a general rule. Do not ship
+  tag-, class-, fixture-, or browser-specific behavior when the underlying
+  concept can be expressed by configuration, the content model, or a reusable
+  primitive.
+- KISS and YAGNI apply, but every contract must remain composable, replaceable,
+  and extensible.
+- Shipped defaults are useful starting points, never fixed policy. Applications
+  may replace them.
+- Fix obvious, local, undisputed defects encountered in the discussed area.
+  Do not expand work into subjective or speculative cleanup.
+- Preserve existing comments and debugging functions unless their removal was
+  explicitly requested.
+
+## Architecture
+
+- Use native `contenteditable`, `Selection`, `Range`, Input Events, and modern
+  Web APIs. Do not build a parallel document model.
+- Never use `document.execCommand()`.
+- Keep one shared core per selection context as the normal architecture. A core
+  may coordinate multiple editable surfaces and multiple roaming or static UIs.
+- Mutable state belongs to its core, surface, transaction, or module instance.
+  Module globals are only for immutable constants and stateless helpers.
+- Nested editable hosts are isolation boundaries for events, selections,
+  traversal, normalization, and commands.
+- DOM mutations owned by the editor go through transactions and explicit point
+  mapping. Preserve forward and backward selection intent.
+- Keep responsibilities narrow. Content validity, repair planning, mechanical
+  mutation, selection mapping, input routing, commands, history, sanitizing,
+  browser policies, and UI remain separate modules.
+- Later layers depend on public contracts, never private implementation details.
+- Browser workarounds must be feature-detected, isolated near the affected
+  primitive, and covered by a named regression test. Do not branch on user-agent
+  strings.
+- Security sanitizing and structural normalization are separate policies.
+  Never pass untrusted HTML to unsafe DOM sinks.
+- Scope normalization to the smallest affected subtree and expand only when a
+  repair crosses that boundary. It must converge, be idempotent, and preserve
+  meaningful content.
+
+## Configuration
+
+- Give every module safe, useful, host-aware defaults without requiring setup.
+- Prefer inheritable CSS custom properties for serializable behavior. Use
+  JavaScript configuration for functions, policy objects, and other values CSS
+  cannot express naturally.
+- Different editing hosts have different semantics. Lists, inline-only hosts,
+  tables, generic blocks, atomic content, and nested editors must not be forced
+  through one generic block rule.
+- Use current browser standards; compatibility with obsolete browsers is not a
+  goal. Modern CSS, including nesting, is allowed.
+
+## Code style
+
+- Code and code comments are English.
+- Use standard ESM: `import`/`export`, `const`/`let`, classes, and functions. No
+  IIFEs, `var`, prototype assignment, or accidental globals. Assign to
+  `globalThis` only for deliberate compatibility or test instrumentation.
+- Use clear, established, natural, preferably short names. One concept gets one
+  term throughout the project.
+- Avoid redundant API names: context already carries meaning (`module.add()`,
+  not `module.addModule()`). Prefer symmetric names for symmetric operations.
+- Keep control flow compact with small helpers and short early returns. Do not
+  compress code until its intent becomes harder to see.
+- Place related code together. Add a file only for a real responsibility,
+  shared use, or separate testability—not merely to reduce line count.
+- Remove middle-man helpers that only forward arguments without adding policy,
+  validation, normalization, or composition.
+- Do not hardcode application-specific tags, classes, commands, or policies in
+  generic primitives.
+- Use braces for multi-line and non-trivial control flow. Keep the local style
+  consistent when changing existing code.
+- Do not run broad automatic formatters. Never run `deno fmt` in this workspace.
+
+## HTML and CSS
+
+- Keep production presentation independent from the engine. UI modules consume
+  editor state but do not become part of the core.
+- Configure behavior through documented `--u2-rte-*` properties with good
+  fallbacks.
+- Keep the playground compact and diagnostic. It must call production modules,
+  never contain a second implementation of editor behavior.
+- Escape or sanitize user-controlled HTML with the selected sanitizing policy.
+  Parsing a trusted test fixture is not a substitute for a safe external-input
+  path.
+
+## Documentation and tests
+
+- Every production file has a dedicated comprehensive test file and focused
+  documentation in its responsibility README.
+- Every responsibility README documents purpose, public contract, invariants,
+  browser considerations, and concrete TODOs.
+- Update documentation and `PLAN.md` in the same change whenever a contract or
+  implementation status changes. Do not describe planned behavior as existing.
+- Tests cover normal behavior, boundaries, invalid input, cleanup, teardown,
+  nesting, direction, cancellation, event order, and state isolation where
+  relevant.
+- A browser quirk becomes a minimal failing regression test before it becomes a
+  policy.
+- Tests leave no fixtures, selections, listeners, active cores, or other state
+  behind.
+- Do not weaken assertions, add timing guesses, or hide failures behind
+  browser-specific skips. Test the stable contract and isolate genuinely
+  engine-specific behavior.
+- Run `/u2/js/rte/tests/` through HTTP. A revision is cross-browser verified
+  only when that exact revision passes in current Chromium, Firefox, and WebKit.
+- Use `/u2/js/rte/playground/` for visual inspection of normalization, mapping,
+  selection, and input behavior.
+
+## Before handing off
+
+- Review the diff for unrelated changes, redundant code, naming consistency,
+  cleanup, and stale documentation.
+- Ensure every changed production responsibility has matching tests and docs.
+- Report the exact browser revisions or engines actually tested. Never infer a
+  cross-browser pass from one engine.
+- Mention changes outside the discussed area prominently, with reason and
+  effect.

package/js/rte/PLAN.md

@@ -0,0 +1,487 @@
+# RTE implementation plan
+
+Each phase ends with documentation, dedicated tests for every production file,
+real-browser verification, and no known failing invariant. Later phases may
+refine earlier contracts, but may not bypass their responsibility boundaries.
+
+The working `../rte0` is a behavioral reference throughout the project. Existing
+behavior, especially host-specific editing, selection handling, and browser
+fixes, is first captured as black-box tests. RTE then preserves or deliberately
+improves it through smaller responsibilities; structural weaknesses in the old
+implementation are not mistaken for behavioral failures.
+
+## Resume work
+
+Phases 0–3 provide the core, selection/range primitives, content model, scoped
+normalization, playground, and input pipeline. The command layer, the prevented
+input path, and Enter close the phase-3 gap between configuration and behavior.
+Phase 4 has started with immutable mark types, serializable values, equivalence,
+directional conflicts, deterministic sets, removal, and replaceable DOM
+adapters. Generic non-collapsed range commands now apply and remove one mark,
+including reusable inline elements and neutral-span cleanup, without
+`execCommand()`. Selection and caret state, toggle, and pending ordinary text
+input are implemented. Adapters can explicitly remove semantic wrappers while
+preserving unrelated attributes, and the standard HTML adapters for bold,
+italic, underline, strike, code, and links exercise the whole path. Pending
+overrides now carry through native IME composition without
+intercepting its mutations. Complete mark sets now resolve exclusions, remove
+configured values absent from an exact target, and apply their canonical result
+inside one mapped operation.
+
+The browser runner covers every phase below: mark algebra, pending overrides
+carried through native IME composition, complete and nested canonical mark sets,
+state-based history with its commands, lists and their nesting levels,
+model-placed element insertion, the HTML source responsibility with its dialog,
+value marks and the contextual link editor, CSS-declared content classes shared
+by the style control, the sanitizer and presentation cleanup, one shared
+placement policy, the full remove-format ladder, registry-owned shortcuts, the
+attribute and element policies on imports with their aliases, the table
+structure commands and the image frame with their contextual handles, and the
+session rule that takes all of it away at once — everything the convention
+client draws in one shadow root per editor. The playground configures its
+prototype surface live from a panel built out of the registered modules, so
+every host property can be tried without editing the page.
+
+How many tests that is, the runner says itself; the number is not repeated here,
+because a number in prose is stale the moment the next test is written. A result
+is cross-browser evidence only after that exact revision has run in current
+Chromium, Firefox, and WebKit. The last such confirmation was Chrome 152 and
+Firefox 154; WebKit verification of that revision is pending.
+
+Remove-format now runs a ladder that does not stop at the application's own
+presentation: six presentation rungs in `Unstyle` (styles, presentational
+attributes, foreign classes, formatting wrappers, declared content classes, and
+the remaining semantic inline elements) plus a seventh `blocks` rung in the
+command that reduces lists, tables, quotes, and headings to the host's default
+block. The action is unavailable only when the selection is already plain text,
+and a collapsed caret reaches the whole content rather than nothing.
+Keyboard shortcuts moved from toolbar controls to the command registry, so a key
+works whether or not a control is on screen, and `Tab`/`Shift+Tab` reach the
+list-nesting commands only inside a list.
+
+What may arrive is now its own question. `--u2-rte-import-elements` defaults to
+the `@content` preset — headings, text, lists, tables, media and the text-level
+semantics — while `--u2-rte-elements` stays what a host tolerates in content it
+already owns. Neither can widen the sanitize policy.
+
+An element the import list does not carry, but whose meaning a listed one does,
+is replaced rather than dropped: `<b>` becomes `<strong>`, `<i>` becomes `<em>`.
+Nothing did that afterwards — the bold mark recognizes `<b>`, but only a mark
+command ever makes an element canonical — so a strict list would otherwise have
+turned every pasted emphasis into plain text.
+
+A native paste now meets the attribute policy. Nothing is parsed there — the
+browser inserts its own payload — so it was the one import path without an
+allowlist, and layout ids, tracking attributes and inline styles came straight
+through. The pipeline applies `SanitizePolicy.clean()` to the nodes that
+arrived, before presentation cleanup and structural repair, so an id removed in
+the first stage is what lets the third dissolve the wrapper carrying it.
+`--u2-rte-import-sanitize: none` opts a host out.
+
+Pasted content is now cleaned where it landed. The cleanup scope was derived
+from the caret, which sits at the *end* of a paste, so a pasted document was
+only ever repaired in its last block — the reason legacy markup survived intact
+even after the planner learned to remove it. The nodes that arrived are already
+tracked, and now decide the scope.
+
+Redundant markup is now removed at any depth. A neutral generic block was only
+ever reshaped as a direct child of the root, so legacy documents kept their
+nested bare `div`s — valid HTML, which is exactly why the content model alone
+never removed them. `canonical` also stopped being a synonym for `structural`:
+it now dissolves a generic inline wrapper that carries nothing.
+
+The same work made normalization measurably cheaper. Measured in Chrome 152 as
+the best of five runs against the previous revision: `ContentModel.allows()`
+14.8 → 3.2 µs, one paragraph normalized 213 → 85 µs, a clean 264-element article
+18.7 → 7.1 ms. The model no longer allocates a string per rule lookup and skips
+the ancestor-exclusion walk for children nothing can exclude; the normalizer no
+longer asks the executor to confirm the plans that change nothing, and no longer
+re-validates every element the walk just collected.
+
+A review pass consolidated four duplicated concepts into single definitions
+(`isEditingBoundary`, `elementOf`, `indexOf`, `narrow`), removed a duplicated
+availability check and a second placement policy, and made `edit.model` resolve
+lazily. Measured in Chrome 152: an availability check that does not consult the
+content model dropped from 10.0 to 1.8 µs, the list scan from 31.4 to 21.0 µs,
+and one toolbar refresh from 490 to 437 µs. A lazy `config()` was measured and
+rejected: building eight accessor descriptors costs more than parsing the values
+eagerly.
+
+Measured again in Chrome after the host-shape work, then sampled with the CPU
+profiler over 400 keystrokes. What a keystroke costs was never normalization: a
+third of it was reading CSS the caller did not need. `Commands.model` and
+`Edit.model` narrow the content model by one property and were parsing all
+eleven, and the toolbar resolved a computed style four times per refresh. Both
+now read what they use, without a cache anywhere — the settings stay live.
+
+| | before | after |
+| --- | --- | --- |
+| `Commands.model` | 8.7 µs | 0.7 µs |
+| `commands.enabled('bold')` | 11.3 µs | 6.1 µs |
+| `Toolbar.refresh` (8 controls) | 160 µs | 120 µs |
+| one keystroke, roaming toolbar | 519 µs | 415 µs |
+
+What is left is flat: `getBoundingClientRect` for the toolbar's placement is the
+largest single item at 14% of a keystroke, then event dispatch, range cloning,
+and the model's own rules. A click inside a surface costs 8.0 µs against 6.0 µs
+for the same event outside the editor, a keydown 4.6 against 3.4, and drawing
+the saved selection 0.5 µs per selection change.
+Treat the number shown by `/u2/js/rte/tests/` as authoritative.
+
+## 0. Foundation
+
+Status: implemented. The foundation and policy suite passed in current
+Chromium, Firefox, and WebKit. Later phase tests extend the same browser runner.
+
+- CSS configuration with semantic host defaults.
+- One core per selection context and multiple editable surfaces.
+- Direction-preserving selection snapshots based on live ranges.
+- Synchronous transactions, dirty-node collection, lifecycle events, and
+  nested transaction reuse.
+- Dependency-free browser test runner.
+
+Exit criteria:
+
+- Static checks pass.
+- Foundation suites pass in current Chromium, Firefox, and WebKit.
+- Focus, selection, nested editors, cancellation, and teardown leak no state.
+
+## 1. Positions and ranges
+
+Status: ownership, live points, range splitting, text/block traversal,
+containment, and explicit operation mapping are implemented. The browser
+baseline passed in all target engines. Generated cases cover range traversal and
+the content-preserving point-map operations; `remove`, `replace`, and `move`
+still need an oracle of their own.
+
+- Define logical DOM points with affinity at text, element, and atomic bounds.
+- Map points through split, wrap, unwrap, replace, move, merge, and remove.
+- Traverse selected content by blocks and editable boundaries.
+- Preserve forward/backward selections and collapsed carets through every
+  primitive operation.
+
+Tests emphasize every boundary offset, empty nodes, `<br>`, replaced endpoints,
+atomic nodes, nested hosts, and generated forward/backward ranges.
+
+## 2. Content model and normalization
+
+Status: the immutable policy engine, standards-oriented HTML rules, pure repair
+planning, mapped repair execution, scoped fixed-point normalization, and a
+visual playground are implemented. Generated cases assert convergence,
+idempotence, preserved text, and valid output across levels and host tags.
+
+- Model permitted parents/children from HTML content categories plus editor
+  policy overrides.
+- Implement minimal, structural, and canonical normalization as convergent
+  rules: repair, lift, split, unwrap, convert, merge, and remove.
+- Preserve meaningful attributes, classes, and styles while removing neutral
+  wrappers and empty metadata.
+- Reduce transaction touches to the smallest stable dirty scopes; expand only
+  when a repair crosses the boundary.
+- Make normalization idempotent and canonical output deterministic.
+
+The useful ideas in the old `NodeCleaner`—allow-lists, modernization, removal of
+empty metadata, block-only wrapper detection, and Unicode normalization—become
+separate policies. Computed-style comparison, prototype mutation, and unrelated
+whole-tree cleanup are not carried forward.
+
+## 3. Input pipeline
+
+Status: post-native event classification, CSS cleanup triggers, live target
+ranges, local normalization scope, composition deferral, nested-host isolation,
+selection mapping, disposal, command routing with native text data for
+prevented input, and replacement by an already prepared DOM fragment are
+implemented. `ExternalInput` consumes rich paste/drop `dataTransfer` and the
+native target range, applies the selected sanitizer and optional Unstyle
+policy, and invokes mapped fragment insertion while failures stay fail-closed.
+Without an explicitly selected sanitizer, native paste/drop remains
+browser-owned; the pipeline removes classes and styles only from observed added
+roots before structural normalization while preserving elements that predate
+the input.
+Collapsed backward and forward deletion at a
+mergeable block boundary are routed explicitly; ordinary character and
+selected-range key deletion remain native. Contextual plain-text and quotation
+import are next.
+
+- Route `beforeinput`, `input`, composition, paste, drop, delete, and Enter by
+  `inputType` and host policy.
+- Let native editing proceed only where its result is interoperable and covered
+  by postconditions; otherwise prevent it and execute an explicit command.
+- Sanitize external fragments before insertion, then normalize only the fragment
+  and insertion neighborhood.
+- Give lists, inline hosts, tables, generic blocks, and atomic content semantic
+  default behavior that CSS can override.
+
+Tests use trusted browser input where automation allows it and direct event/
+command contracts for exhaustive edge cases.
+
+## 4. Marks and inline commands
+
+Status: started. Mark values define equivalence, directional exclusion,
+canonical ordering, and removal. Replaceable adapters parse existing elements
+and render canonical wrappers for semantic tags, classes, attributes, styles,
+and custom policies. Generic commands apply and remove one configured mark over
+arbitrary non-collapsed ranges, preserve selection direction, reuse suitable
+inline elements, join adjacent canonical wrappers, clean up neutral spans, and
+derive selection or structural caret state for toggle. A closed adapter
+universe can now report and set exact canonical mark sets while resolving type
+conflicts in the same transaction. Redundant nesting is removed, canonical
+single-child wrappers follow mark rank, and newly exposed siblings merge to a
+fixed point without crossing meaningful or atomic boundaries. Formatting other
+than the standard HTML policies remains application-defined.
+`PendingMarks` routes ordinary `insertText` through mapped input and applies the
+same caret overrides to a live native composition range only after IME ends.
+The ready-made adapters cover bold, italic, underline, strike, code, and links.
+They render one canonical tag, remove their semantic wrappers explicitly, and
+preserve unrelated attributes. Links use a small structured value without
+coupling URL policy to mark representation.
+
+- Apply, remove, and toggle over arbitrary ranges without
+  `document.execCommand()`.
+- Support semantic elements, class tokens, attributes, style declarations, and
+  custom mark adapters through one algebra.
+- Split boundaries minimally; merge equivalent siblings; eliminate redundant,
+  conflicting, or empty wrappers.
+- Keep staged remove-format separate as Unstyle because it spans multiple mark
+  representations. A future link-control module owns URL entry and validation.
+
+Tests combine partial text, multiple blocks, nested marks, overlapping removal,
+backward selections, atomic content, repeated toggles, and exact undo.
+
+## 5. Structural commands
+
+Status: started. The command registry, the `Edit` execution context, the
+`PointMap.split()` primitive, and the Enter/line-break commands are implemented.
+Enter also exits an empty item from a nested list without crossing its surface,
+splits content-model text blocks such as headings, and creates the configured
+default block after a non-default text block's exact end. Backspace and Delete
+merge compatible content-model blocks at either boundary and treat every caret
+representation inside an empty block identically.
+`BlockStyles` converts one or several known text-block wrappers, preserves
+selection and unrelated attributes, reports mixed state, and supports custom
+selector/tag/write/clear definitions without treating layout blocks as text
+styles.
+`Lists` owns one closed group of container elements and takes item elements from
+the model's `defaultChild`, so no command names `li`. `toggle(tag)` converts
+blocks into items, converts an existing list of another kind, and lifts items
+back out; a partly selected list is split so only the selected run changes, and
+a result that meets a list of its own kind joins it. `indent` and `outdent` move
+a run between nesting levels and claim `formatIndent` and `formatOutdent`.
+`insertNode(create, inputTypes)` inserts one prepared element where the content
+model accepts it, splitting the caret's block only that far, and covers the
+horizontal rule.
+
+- Extend paragraph and heading conversion to loose pre-normalization text.
+- Blockquotes, links, tables, and configurable atomic nodes.
+- Preserve ordered-list numbering across a split or lift.
+- Delete a non-collapsed selection before splitting or inserting, instead of
+  leaving that case native.
+
+## 6. History
+
+Status: implemented as a state model. `History` records one entry per change of
+one surface, storing the content as a cloned fragment with a path-addressed
+selection that survives content replacement. A `MutationObserver` covers every
+origin, so native typing, commands, paste, drop, and unrelated application
+scripts all enter history without the engine owning their mutations.
+Transactions supply the grouping boundaries: the `input` trigger coalesces into
+one entry per interval, every other trigger records a discrete step before and
+after itself. `undo`/`redo` are ordinary commands claiming `historyUndo` and
+`historyRedo`; the input pipeline routes Ctrl/Command+Z and +Y from `keydown`
+because a browser stops reporting those input types once its own stack no longer
+matches replaced content. The convention client installs one history per
+rich-text surface and contributes both toolbar controls.
+
+- Store diffs instead of whole content states once transactions record
+  reversible operations, keeping the state model for unowned mutations.
+- Group composition deliberately rather than relying on its mutations alone.
+- Expose the coalescing interval and an off switch as `--u2-rte-history`.
+- Decide whether the entry limit should be measured in bytes rather than
+  entries, and verify the cost on large documents.
+
+## 7. Sanitizing and serialization
+
+Status: started. `SanitizePolicy` now defines one immutable element, global and
+per-element attribute, URL-protocol, comment, and data-attribute contract. The
+first `NativeSanitizer` adapter parses into a detached fragment exclusively
+through `Element.setHTML()`, intersects surface elements without broadening the
+security policy, and fails explicitly when that safe sink is unavailable. It
+has no unsafe fallback. Native paste/drop needs no parser fallback because its
+payload stays browser-owned. `Source` now serializes a surface and parses source text back through the same
+policy. A canonical serializer and contextual parsing remain open.
+The independent immutable `Unstyle` policy supplies ordered presentation
+levels to both the selection command and safely parsed external fragments, so
+paste/drop will not grow a second formatting-cleanup list.
+`insertFragment` now replaces a selection with such a prepared fragment through
+mapped removal and insertion. Its content-model lifting handles inline, block,
+list, and nested-link contexts without browser editing commands.
+`ExternalInput` composes the three responsibilities for applications that
+explicitly replace rich paste/drop. The default cross-engine path instead
+cleans the browser's inserted DOM without reading its HTML payload.
+
+- Resolve the surface's structural element policy once and use it as the upper
+  bound for cleanup, command availability, paste/drop sanitizing, and toolbar
+  choices. A sanitizer may narrow that policy but never broaden it.
+`--u2-rte-classes` declares the class names a host treats as content. It is one
+declaration with three consumers: the style control offers exactly those names,
+`SanitizePolicy.clean({classes})` keeps only those from external HTML, and
+`Unstyle` narrows the class attribute to them and leaves a wrapper carrying one,
+because a declared class is content rather than presentation.
+
+- Keep additional sanitizer adapters application-supplied behind the same
+  policy contract unless a general engine use case requires one.
+- Define allowed global/per-element attributes, URL protocols, class/data
+  policy, inline-style policy, and comments separately from allowed elements.
+- Keep security sanitizing separate from structural normalization.
+- Serialize canonical HTML without editor-only markers or browser debris.
+- Test hostile fragments, contextual parsing, mutation-XSS regressions, URL
+  attributes, custom elements, pasted office markup, and policy narrowing.
+
+## 8. UI adapters
+
+Status: started. The optional roaming `Toolbar` binds application-owned markup
+to the active surface's resolved command registry, reflects availability and
+boolean/mixed state, keeps saved selections across UI focus, supports simple
+Ctrl/Command shortcuts, shows and hides with the active surface, and
+delegates placement and presentation. Application-supplied manual popovers and
+the convention toolbar now enter and leave the browser top layer with the same
+visibility contract. `--u2-rte-toolbar-when: selection` optionally requires a
+non-collapsed saved selection.
+
+The first batteries-included convention-client prototype is implemented: one
+side-effect import plus `--u2-rte` and `--u2-rte-toolbar` lazily wires standard
+Enter/input behavior, Bold, and one shared default toolbar. It allocates no UI
+before the first active rich-text surface and leaves `plaintext-only` native.
+The explicit `rte.js` API remains the composable engine layer. Optional
+extensions install atomically into current and future surfaces through
+`Editor.add()`, contribute controls, and are removed everywhere through
+`Editor.delete()`. The optional `blocks.js` module exercises this with a
+command-valued Paragraph/H1/H2/H3 select and replaceable custom style policy.
+The optional `breaks.js` probe adds one editor-wide `setup()` lifetime, one
+per-surface `attach()` lifetime, a non-transactional view command, and a
+presentation-only top-layer overlay without inserting markers into editable
+HTML. Optional `unstyle.js` contributes a staged action whose first applicable
+level changes only selected content; levels and labels are replaceable.
+
+- Publish command availability and active/mixed state as observable editor
+  state.
+- Extend roaming controls beyond pressed buttons and command-valued selects to
+  menus and application-owned custom controls. A select's `options` may already
+  be a function of the surface, so a control's choices can come from the host's
+  configuration; the client refills it before the toolbar is shown.
+- Presence follows the configuration and availability the selection: a control
+  this editor does not offer is absent, one that cannot act right now is
+  disabled. A toolbar that rearranged itself as the caret moved would move its
+  targets out from under the pointer. `--u2-rte-toolbar-unavailable: hide`
+  trades that away where a host prefers a toolbar of only what it can do.
+- Implement static UIs bound to one surface; allow several UIs and command
+  subsets to coexist.
+- Verify reusable placement policy across carets, ranges, writing modes, and
+  viewport edges.
+- Exercise the lifecycle with another contextual UI before adding dependency,
+  ordering, or asynchronous setup concepts.
+
+### Optional extension probes
+
+These are concrete architecture tests, not commitments to one large plugin
+framework. Each extension must install explicitly, add no resources while it is
+absent, work for current and future surfaces, and release everything through
+`dispose()` when it or its editor is removed.
+
+1. **Visible line breaks — prototype implemented.** It is non-destructive but needs
+   extension-owned presentation and per-surface state. Provide a CSS default,
+   an optional toolbar toggle, and an always-on mode that needs no button. Do
+   not insert marker nodes into editable HTML. Chromium does not render
+   generated content on `<br>`, so the prototype uses one scheduled top-layer
+   overlay without a mutation observer. Firefox and WebKit still need visual
+   confirmation.
+2. **HTML source editor — implemented.** `Source` serializes a surface by
+   walking its DOM and reports where the selection lands in that text, so the
+   view opens at the caret without inserting marker nodes. Formatting breaks a
+   level into lines only where whitespace cannot be significant, and writing
+   removes exactly those breaks again, so a read/write round trip is lossless.
+   Writing always parses through the sanitizer and never assigns `innerHTML`.
+   `source.js` adds the top-layer dialog, whose text area is wrapped in
+   `<u2-code>` so the source is highlighted where that element is defined and
+   plain where it is not; the engine itself gains no dependency. Mapping the
+   dialog's own selection back into the DOM and a canonical serializer remain
+   open.
+3. **Link editor — implemented.** `valueMark(adapter)` generalizes the value
+   half: one command creates, changes, and removes a mark whose value is
+   content, and at a caret it acts on the whole mark it sits on, so an existing
+   link's address can be changed without selecting its text. `link.js` adds the
+   contextual form, which follows the caret the way the table and image handles
+   do: it appears at a link, anchored on it, and goes when the caret leaves, so
+   the toolbar control is left with creating one. Its address field is plain text
+   because native url validation rejects relative paths, fragments, and
+   application schemes; protocol and attribute policy stays with the sanitizer.
+   The form applies as it is typed rather than on an Apply button, marking each
+   run as ongoing input so history keeps an address as one step, and it edits the
+   one link it appeared on. `normalize`, `suggest`, and `complete` leave what an
+   address means to the application, the last of them filling a list of entries
+   that may carry markup. Editing the link text itself remains open.
+4. **Table tools — implemented as commands.** `Tables` groups insertion,
+   row and column insertion and deletion, and table removal. It names no
+   section, row, or cell tag: each comes from the model's `defaultChild`, one
+   level apart. Every change is an ordinary mapped mutation, so it is one undo
+   step and the caret lands where the edit happened. Spanning cells make the
+   index-counted actions unavailable rather than shifting the wrong cells.
+   `tables.js` puts them on the table itself: row handles down its left edge and
+   column handles along its top, each lined up with the cell the caret is in,
+   repositioned from selection, change, input, scroll and resize rather than
+   from a timer. Cell movement and span-aware column arithmetic remain open.
+5. **Image sizing — implemented.** `selectedElement()` and
+   `elementAttributes()` make the one element a selection covers addressable and
+   give it a value command over a fixed set of attributes; `images.js` adds the
+   frame and the pointer interaction. The drag moves the frame only and the size
+   is written once on release, so a resize is one undo step and the overlay owns
+   no size policy: `width` and `height` are what the sanitize policy allows on an
+   image. What is resizable is a selector. Its three handles sit on the trailing
+   edges only, because the flow holds an image's top and start edges: the corner
+   keeps the proportion, the two edges change one measurement each. An alt-text
+   editor and keyboard resizing remain open, as does a hook for regenerating the
+   file server-side rather than letting the browser stretch it.
+
+All roaming and contextual extension UI must be created in the core's
+`Document` or `ShadowRoot`, use the browser top layer through Popover or Dialog
+where appropriate, and share one placement policy.
+
+Everything the convention client draws lives in one `Chrome`: a single shadow
+root per editor holding the toolbar, the contextual handles, the link form and
+the source dialog. An editor is chrome inside someone else's document and has to
+survive their `button {}` rule; one encapsulated root also means the application
+sees a single element instead of one per piece of UI. There is no `::part`
+surface on purpose — it would make the chrome's internal structure a public
+contract, and an application that wants different chrome builds its own on the
+commands, which `Toolbar` has always supported by binding markup it is given.
+The chrome now follows the active surface through the closest modal dialog,
+open popover, or fullscreen boundary and returns to its original root when that
+boundary closes. Native inertness makes descendant mounting mandatory; core
+retention keeps the temporary UI child out of focus deactivation, Source, and
+History even when the top-layer element is itself the editing host.
+
+The handles inside it are one shared `Handles` component that owns no editor
+concept, so anything with a rectangle can use it, inside this engine or outside
+it: given a document instead of a shadow root it brings its own. Shadow-root style ownership,
+focus transitions, disposal, and coexistence of multiple UIs are release-gate
+tests, not application-specific details.
+
+## 9. Browser policies and release gate
+
+Status: started with `rangeRect()`, a non-mutating geometry fallback for
+collapsed native ranges. Browser policies live in dedicated, independently
+tested files rather than one global fixes bundle; semantic editing behavior
+stays in its owning command, input, normalization, or UI responsibility.
+
+- Isolate feature-detected Chromium, Firefox, and WebKit differences by the
+  primitive or input behavior they correct.
+- Promote every discovered browser quirk into a minimal regression test before
+  adding a policy.
+- Run generated range/DOM cases and interaction suites repeatedly in all target
+  engines.
+- Verify teardown, memory-sensitive listener ownership, accessibility, IME,
+  mobile selection, touch, clipboard, drag/drop, and Shadow DOM.
+
+The first release requires all suites green in the latest stable target
+browsers, complete responsibility READMEs/TODOs, and no `document.execCommand()`
+usage anywhere in RTE.