Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@docs/design/docs.md`:
- Around line 37-39: Update the design doc's IME scope to reflect the
implemented behavior: change the "IME composition handling — deferred (basic
`beforeinput` only)" non-goal to state that Korean IME composition (including
the Mobile Safari workaround added in this PR) is supported; mention the
specific implemented coverage (Korean IME on desktop and mobile Safari
workaround) and, if relevant, note any remaining limitations instead of marking
IME handling as deferred. Reference the "IME composition handling" phrase in the
docs so reviewers can verify the update matches the PR changes.
- Around line 10-30: The doc mentions the old package identity `packages/docs`
and `@wafflebase/docs`; update every occurrence to the new package identity
`packages/document` and `@wafflebase/document` (including the Goals section and
the Structure section where package names and import paths are listed) so
references, examples, and any path strings match the implemented package; search
for the symbols `packages/docs`, `@wafflebase/docs`, and any path examples in
this file and replace them with `packages/document` and `@wafflebase/document`
ensuring consistency across the entire document.

In `@docs/tasks/active/20260320-docs-package-todo.md`:
- Line 5: The doc entry currently shows "Status: COMPLETE" while the manual
validation item "manual smoke test" remains unchecked; update the task so status
reflects reality by either (a) completing the unchecked "manual smoke test" item
at the bottom of the file and adding any required verification notes, or (b)
reverting "Status: COMPLETE" back to an in-progress state and adding a TODO
comment indicating the pending "manual smoke test" and owner; locate the
"Status: COMPLETE" header and the unchecked "manual smoke test" checklist item
to make the change.

In `@packages/document/demo.ts`:
- Around line 23-28: The logEvent function uses innerHTML to build a line which
injects the user-influenced detail and is XSS-prone; replace the innerHTML
construction in logEvent with safe DOM creation: create a container div, create
three spans (one for the padded eventSeq number with style, one for the type
with class cls and textContent set to type.padEnd(20), and one for the detail
with textContent set to detail), append them to the container, then append the
container to logEl—ensuring you reference logEvent, eventSeq and logEl when
making the change.
- Around line 71-77: The clipboard copy handler on the 'btn-copy-log' element
uses a .then() chain without error handling; change the click listener callback
to an async function that awaits navigator.clipboard.writeText(logEl.innerText)
inside a try/catch, update the button text to 'Copied!' on success and to an
error state (e.g., 'Copy failed') on catch, and restore the original 'Copy' text
after a timeout in both paths; ensure you reference the same DOM element
(document.getElementById('btn-copy-log') / logEl) and handle the case where the
element or navigator.clipboard may be unavailable.

In `@packages/document/scripts/verify-ime-browser.mjs`:
- Around line 371-430: The script currently treats missing browser executables
as warnings and still prints "[verify:ime] All IME browser tests passed." even
if no scenarios ran; add a sentinel (e.g., ranAnyScenario = false) and set it
true inside each successful runAllScenarios call (for Chromium and WebKit), then
after both engine blocks check the flag and if still false call
printInstallHelp() and exit non‑zero (or throw) so the script fails when both
browsers are missing; reference chromiumBrowser, webkitBrowser, runAllScenarios
and the final success log to locate where to set the flag and where to fail.

In `@packages/document/src/model/document.ts`:
- Around line 61-84: deleteText can hang if pos.offset is at the end of the
block or length exceeds remaining text because resolveOffset may return
available === 0 and the loop doesn't advance; update deleteText to first compute
the block's remaining characters from pos (e.g., sum of inline lengths after pos
offset via getBlock and resolveOffset or by summing block.inlines), clamp the
requested length to that remaining amount and return early if clamped length is
0, then proceed with the existing loop and calls to
resolveOffset/normalizeInlines to perform deletions.

In `@packages/document/src/store/memory.ts`:
- Around line 23-35: getDocument and getBlock currently return live references
allowing external mutation that bypasses pushUndo; change them to return
deep-cloned copies instead (use cloneDocument for getDocument and
cloneBlock/deep clone for getBlock) so callers cannot modify store internals
directly, and ensure initialize/Doc construction uses the cloned document rather
than the original; leave setDocument as-is (it already clones and calls
pushUndo) and keep pushUndo usage unchanged.

In `@packages/document/src/view/editor.ts`:
- Around line 77-90: render() always paints from y=0 and resizes the canvas to
the container height so layout.totalHeight and container.scrollTop are ignored;
fix by making the canvas height match layout.totalHeight (or at least
max(container height, layout.totalHeight)) and pass the current vertical offset
(container.scrollTop) into docCanvas.render instead of 0, and adjust calls that
compute pixel positions (cursor.getPixelPosition and
selection.getSelectionRects) to accept/compensate for container.scrollTop so
hit-testing and rendering align; update recomputeLayout usage if needed to
ensure layout.totalHeight is computed before resizing/rendering.
- Around line 47-56: Edits are mutating a plain Doc instance created from
storeDoc (via new Doc(storeDoc)) so changes bypass the DocStore undo stack;
refactor so all mutations go through DocStore mutation APIs and record undo
snapshots: instantiate or obtain the live Doc via DocStore (use
docStore.getDocument()/docStore.setDocument patterns or add a
docStore.createLiveDoc helper) and ensure TextEditor, applyStyle, and
applyBlockStyle call the store’s mutation methods that internally call
pushUndo() (or explicitly call docStore.pushUndo() before applying changes) so
every edit is recorded; update the code paths referenced (the Doc instantiation
at the top plus the code areas around TextEditor, applyStyle, applyBlockStyle
and the regions noted at 93-101 and 120-148) to use the store-backed mutation
methods instead of mutating the raw Doc object.

In `@packages/document/src/view/layout.ts`:
- Around line 76-85: computeLayout() currently computes a single lineHeight
using getBlockFontSize() which only inspects block.inlines[0], causing later
larger inlines to overlap; change the calculation to base lineHeight on the
maximum font size in scope instead: either update getBlockFontSize() to scan all
inlines in the block (not just block.inlines[0]) or add a helper like
getLineMaxFontSize(line) and compute lineHeight per line as (lineMaxFontSize *
(inline.style.lineHeight ?? block.style.lineHeight ?? 1.5)), then use that
per-line height when setting line.height and updating blockHeight in the loops
inside computeLayout() (and apply the same fix at the other occurrence around
lines 255-260).
- Around line 129-152: layoutBlock currently only wraps at word-level segments
from measureSegments, so oversized segments (seg.width > maxWidth) never break
and overflow; modify the loop that iterates over segments in layoutBlock to
detect when seg.width > maxWidth and perform a character-level fallback: iterate
the characters of the offending inline (use seg.charStart/charEnd and
block.inlines[seg.inlineIndex] to access text), measure sub-segments per
character (or small slices) and emit runs that fill remaining space then wrap,
updating currentRuns, lineWidth, and lines just like the normal path; apply the
same character-level splitting logic to the second wrapping loop referenced (the
block around lines 171-229) so both wrapping places handle oversized tokens.
- Around line 369-394: The hit-test offset is wrong because run.charStart is
stored per-inline (via measureSegments) so treating it as a line-global index
miscomputes offsets for later runs; fix the calculation in the loop over
targetLine.runs (the block referencing run.charStart and targetLine.runs[0]) by
computing the offset within the line as the sum of the lengths of all previous
runs plus bestOffset (e.g. sum targetLine.runs[0..runIndex-1].text.length) and
then return blockId with offset = charsBeforeLine + that sum + bestOffset
instead of using run.charStart.

In `@packages/document/src/view/selection.ts`:
- Around line 96-123: The multi-line selection branch currently builds rects
using the full line block bounds (lb.x and lb.width), causing
centered/right-aligned lines to highlight into margins; update the three rect
constructions to use the actual run/line bounds instead: for the first-line rect
use startPixel.x and startPixel.width (or the computed run width) instead of
lb.width, for the middle-line loop use each line's actual run x and width rather
than lb.x/lb.width, and for the last-line rect use endPixel.x - runX (or
endPixel.width) anchored at the run's x; locate and change the code that pushes
into rects in the multi-line branch (references: rects array, startPixel,
endPixel, lb, startPixel.height) so each rect is constrained to the line run
bounds rather than the full block.

In `@packages/document/src/view/text-editor.ts`:
- Around line 59-67: The key handler in TextEditor is preventing default for
Cmd/Ctrl+Z but not performing undo/redo; update the branch in the TextEditor
keydown handler (the code path inside the TextEditor class that checks for
Meta/Ctrl+Z and Shift/Meta+Z or Ctrl+Y) to call the document undo/redo APIs and
then re-render: invoke this.doc.undo() for plain Meta/Ctrl+Z and invoke
this.doc.redo() for Meta/Ctrl+Shift+Z and Ctrl+Y (or, if your Doc API uses
different names, call the appropriate undo()/redo() methods), keep
event.preventDefault(), and call this.requestRender() after the change so the
editor updates.
- Around line 351-382: The handler computes newPos from this.cursor.position
even when a selection exists, causing arrows without shift to move from the
wrong spot; change handleArrow so when shiftKey is false and
this.selection.range exists you first collapse the position to the appropriate
selection boundary (for 'left' and 'up' use the earlier/minimum of range.anchor
and range.focus; for 'right' and 'down' use the later/maximum), then call
moveLeft/moveRight/moveVertical using that collapsed base position, then clear
the selection via this.selection.setRange(null) and move the cursor to newPos
and requestRender; reference handleArrow, this.selection.range, range.anchor,
range.focus, moveLeft, moveRight, moveVertical, selection.setRange, and
cursor.moveTo.
- Around line 539-576: The software-assembler path in applyHangulResult doesn't
clear the active selection before inserting the first composing jamo, so
previews get appended instead of replacing selections; update applyHangulResult
so that in the composing branch where (result.composing &&
this.hangulComposingLength === 0 && !result.commit) you call deleteSelection()
before setting this.hangulStartPos (and then proceed to insert result.composing
and set hangulComposingLength), ensuring behavior matches the browser-IME path
that clears selection in handleCompositionStart().

---

Nitpick comments:
In `@docs/design/README.md`:
- Around line 26-29: The docs index contains a duplicated row for docs-site.md;
remove the redundant entry so the table only references docs-site.md once
(update the table rows that list [docs-site.md](docs-site.md) to eliminate the
duplicate at the second occurrence), ensuring any descriptive text remains
correct and spacing/alignment of the Markdown table is preserved.

In `@packages/document/index.html`:
- Around line 23-27: The CSS for `#editor-container` uses a fixed width (800px)
which breaks on small viewports; update the selector to use responsive
sizing—replace the fixed width with a fluid width (e.g., width as a percentage
or calc(100% - some padding) plus max-width: 800px) and add a simple media query
fallback for very small screens to adjust height or padding if needed; target
the `#editor-container` rule and ensure border, margin and position behavior
remain unchanged while swapping the fixed width for responsive properties.

In `@packages/document/test/view/hangul.test.ts`:
- Around line 71-73: The test title/comment is inconsistent: it says "한아" but
the assertion expects "하나"; update the test description to reflect the expected
output by renaming the it(...) title from '"한아" → ㅎㅏㄴㅏ' (or similar) to '"하나" →
ㅎㅏㄴㅏ' so the human-readable description matches the assemble(['ㅎ','ㅏ','ㄴ','ㅏ'])
expectation and intent around the assemble function's behavior.

In `@packages/document/vite.config.ts`:
- Around line 1-3: Add a brief clarifying comment at the top of this file
explaining why an empty Vite config is exported (e.g., to preserve
dev-server/default behavior while library-specific build config lives in
vite.build.ts), so future maintainers understand the presence of export default
defineConfig({}) and the relationship to vite.build.ts.