Adds a soft, GPU-cheap fade-in to each newly-mounted top-level markdown
block while a message is streaming. Maps user's mental model of
"line-by-line" onto markdown blocks: paragraph, heading, list item,
fence, blockquote, table, hr — fires once per new block at mount.

Mechanics
- TypewriterMarkdown sets data-streaming="true" on its wrapper while the
  message is streaming. Absent (not "false") on completed/replayed
  messages so the CSS rule cannot match historical transcripts.
- MarkdownCore appends "streamdown-root" to the className passed through
  to Streamdown's root div — a stable CSS hook with no extra wrapper.
- globals.css adds a keyframes + block-level selector list scoped via
  prefers-reduced-motion. Inline elements (<strong>/<em>/<a>) are
  intentionally excluded because parseIncompleteMarkdown can transiently
  insert/remove them as remend repairs unterminated tokens.

Why this works
- Streamdown keys each block by useId+index and wraps it in React.memo,
  so existing blocks are reconciled in place when their content grows.
  Only newly-arriving top-level blocks mount fresh — and CSS animation
  on a freshly-mounted child fires exactly once at mount.
- No JS bookkeeping, no render-phase work, no prop changes.

Per-character pacing inside SmoothTextEngine still operates inside each
block; the new fade-in layer is purely additive at block boundaries.

Codex P2 on PR #3221: replay rows are emitted with isStreaming=true
and streamPresentation.source="replay" while the backend rebuilds
history on reconnect. Gating data-streaming on isStreaming alone
would set the attribute and animate every block of every replayed
message — exactly what we wanted to avoid.

Compute isLiveStreaming = isStreaming && streamSource !== "replay"
and gate the data-streaming attribute on that. Completed messages
already have isStreaming=false, so the attribute is naturally absent
in their case too.

Test extended to assert all three cases: live → "true", replay → absent,
completed → absent.

Replace the per-block fade-in (which fired only once per paragraph at
mount, then text grew silently inside) with a per-word blur-fade-in.

Mechanism:
- rehypeSplitWordsForFade walks the HAST tree and wraps each
  whitespace-bounded word in <span class="sd-word">. Skips inside
  <pre>/<code>/KaTeX/Mermaid subtrees so code, math, and diagrams
  stay structurally intact.
- MarkdownCore swaps in this rehype plugin only while
  parseIncompleteMarkdown=true (i.e., during streaming) — completed/
  static markdown still renders without word spans (no extra DOM weight).
- globals.css replaces the per-block keyframes with sd-word-blur-in
  (filter: blur(6px) → blur(0), opacity 0 → 1, 500ms ease-out).
  Still gated on [data-streaming="true"] and prefers-reduced-motion.

Why per-word
- The user reported the prior per-block effect was invisible in
  thinking blocks (which are typically a single paragraph — block
  fade-in fires once at mount, then text grows silently inside the
  same <p>). Per-word also animates inside an already-mounted block:
  each new whitespace boundary mounts a fresh span and triggers a
  fade for that word and only that word.
- Stable identity at every word position: as text grows from "Hel" →
  "Hello", react-markdown reconciles the same span in place; only
  fresh whitespace boundaries mount new spans. Completed words stay
  static while only new words fade in.

Tests
- New rehypeSplitWordsForFade.test.ts covers the core invariants:
  word/whitespace tokenization, partial trailing word, empty input,
  pre/code skip, KaTeX skip, recursion into inline (<strong>/<em>).

Upstream regression: @smithy/util-retry@4.3.7 was published at
2026-05-02T04:09:26Z with workspace:^ refs that escaped the smithy
monorepo. Any lockfile-free bun install (used by both
scripts/check-bench-agent.sh and scripts/smoke-test.sh — which mimic
'bun x mux@latest' for end users) fails to resolve those refs.

Reproduction outside this repo:
  mkdir t && cd t
  echo '{"dependencies":{"@aws-sdk/credential-providers":"^3.940.0"}}' > package.json
  bun install --ignore-scripts
  # → 'Workspace dependency "@smithy/types" not found'

This breaks every PR opened/pushed after 04:09Z (PR #3219 was lucky to
merge ~30 min before). Add an npm-style overrides pin to <=4.3.6 (the
last known-good release, 2026-04-28) until smithy republishes.

Verified locally: 'bash scripts/check-bench-agent.sh' now passes, and
@smithy/util-retry resolves to 4.3.6 in a fresh lockfile-free install.

The prior static gradient mask obfuscated the line currently being typed
at low opacity, so users saw partial words at varying transparency. Pivot
to: hide the in-progress line entirely, then fade in each completed line
as a unit at the moment its visual row wraps to the next.

Mental model (rendered transcript A → B → C):
- While B is partially typed, the user sees only A. B is hidden.
- When B's visual row completes (text wraps to row C), B fades in over
  ~250ms with no horizontal stagger, no tear, no jitter.

Implementation:

CSS (globals.css)
- Sharp linear-gradient mask hides the bottom 1.6em (~1 visual line at
  line-height 1.6). A 0.4em fade band right above the boundary softens
  the edge.
- New `@property --reveal-y` (typed <length>, initial 0px) is an extra
  offset that pushes the transparent strip upward.

JS (TypewriterMarkdown.tsx)
- ResizeObserver on the markdown root tracks rendered height growth (the
  signal that maps directly to visual rows in the user's viewport — not
  AST/HAST).
- On any positive delta, Element.animate() pulses --reveal-y from
  `delta` → 0 over 250ms. The mask gradient sweeps past the new
  content, fading it in temporally.
- ResizeObserver presence is guarded so happy-dom test environments
  degrade silently; real Chromium always has it.

Why this is genuinely viewport-aware:
- Trigger is rendered geometry. A long paragraph wrapping to N visual
  rows reveals one row at a time; one AST text node doesn't fool the
  reveal. Width-driven re-wraps from sidebar collapse, window resize,
  and zoom all participate naturally because they change offsetHeight.

Co-authored-by: Mux <noreply@coder.com>

The mask was unconditionally hiding the bottom 1.6em (~1 visual line),
which blanked out short replies and stream starts before the first
wrap. Codex P1: 'users can see a blank assistant message until a
second line wraps or the stream completes, which makes live streaming
look stalled'.

Gate the mask on a new `data-stream-hide-tail` attribute, set by JS
only when rendered height is at least ~1.5 line-heights (between 1 and
2 lines). Below that, the message renders fully visible. The threshold
is computed from the element's actual computed line-height so it
scales with font-size and line-height overrides.

Also skip the WAAPI pulse on the first-time threshold crossing
(1 line → 2 lines): the about-to-be-completed first line was already
visible without a mask, so a pulse would briefly hide and re-fade it
in (a flicker). Subsequent wraps (2+ lines stable, growing) still
pulse normally.

Co-authored-by: Mux <noreply@coder.com>

`el.toggleAttribute(name, force)` sets the attribute with an EMPTY
value, not "true", so the CSS selector
`[data-stream-hide-tail="true"]` never matched and the mask was
never applied. Use a presence selector `[data-stream-hide-tail]`
instead. Codex P1.

Co-authored-by: Mux <noreply@coder.com>

Upstream published @smithy/util-retry@4.3.8 ~32min after the broken
4.3.7, fixing the workspace:^ resolution failure. Verified clean
install of latest in a fresh root works.

Revert:

1. package.json: drop the `overrides` block pinning util-retry to
   <=4.3.6.
2. scripts/smoke-test.sh: drop the ROOT_DIR capture and node-script
   that injected the workspace's overrides into the smoke-test root's
   package.json. Restore the original `cat >package.json <<EOF`
   minimal init.

Codex P1: the override injection defeated the purpose of
SKIP_SHRINKWRAP=1 (lockfile-free end-user-equivalent installs), which
exists to catch upstream regressions that `bunx mux@latest` /
`npx` users would hit. By mirroring our root overrides, the smoke
test was masking exactly the failure mode it's designed to catch.
With the upstream fix in place, the workaround is no longer needed
and the smoke test can return to its canonical form.

Co-authored-by: Mux <noreply@coder.com>