[saminsohag]

Summary

Fixes the flicker seen while streaming and after a reply finishes, and makes the wave and blurIn character animations actually look different from fade.

Flicker

• Blocks replayed their entrance when a reply went quiet. The incremental view swapped its root between a Semantics wrapper (text arriving) and the bare column (quiet, after 250 ms). The type change re-inflated the whole document, so every table, fence and formula played its block animation again (growIn collapsed and regrew), table scroll offsets reset, and SVGs blanked for a frame — also around every pause longer than 250 ms mid-stream. The wrapper now stays; only its arguments flip, and with container: false and no label it adds no semantics node, so the settled tree is unchanged.
• Raw TeX flashed at the end of $$ blocks. When the closing $$ arrived split across chunks, the open formula was completed with a stray $ inside, failed to parse, and the error fallback drew the source for one chunk (and jumped the height 50 → 76 → 50 px in a probe).
• A display formula dropped back to raw TeX as it closed. The inline hold scanned inside closed \(…\) / \[…\], read the _ in (z)_i as emphasis, and cut the tail mid-formula until the next paragraph began. Closed formulas are now opaque to the hold, like code spans.

Animations

• wave — the crest uses the hue of ColorScheme.primary, made vivid and kept clearly lighter/darker than the text. It used to be a grey derived from the text colour, and Material 3's pastel dark primary alone (#BCC2FF on #E4E1E9) is still invisible.
• blurIn — opacity now runs ahead of the blur, so words are visibly soft while readable; the blur scales with font size (0.2 em, clamped 1.5–8 px). Previously it was under 1 px by half opacity.
• Fade window — entrances were cut off at 64 characters behind the head; the defaults (300 chars/s × 0.25 s = 75) already exceeded it. The window now grows with reveal speed, 64 → max 256.

Performance

• Neighbouring characters at the same progress step (1/32) share one span.
• The view rebuilds only segments with characters actually still fading (RevealEngine.fadingFrom, a binary search over the stamps) instead of a fixed window.

A/B, same session, debug VM, min of 3 runs: at defaults (300 chars/s, 0.25 s) fade/blurIn unchanged and wave ~7% faster; at 1200 chars/s fade +5%, blurIn +10% (the full entrance now plays instead of being clipped), wave unchanged. Not device measurements.

API

Additive only — no breaking changes:

• revealStyleFor(..., {Color? accent, double? fontSize})
• applyReveal(..., {Color? accent, double? fontSize})
• RevealEngine.window, RevealEngine.fadingFrom, RevealEngine.maxFadeWindow

Without accent, revealStyleFor's wave is exactly the old colour (tested). fadeWindow stays at 64, now the starting size.

Also: a local demo/ folder is git-ignored and excluded from the root analyzer.

Tests

New regression tests, each confirmed to fail without its fix:

• reveal_regressions_test.dart — a finished block does not enter again (with isStreaming left on and off)
• math_streaming_test.dart — half a $$ closer never paints raw source; a closed formula is not held for the markup inside it
• reveal_effect_test.dart (new) — wave lands in the accent; crest stands out with a pastel dark accent; blurIn soft while visible and scaled by font size
• reveal_engine_test.dart — window grows / keeps stamps / is capped; fadingFrom is exact

Three existing span-shape / engine tests were updated where they pinned the fixed 64 window or used uniform progress (which now correctly merges into one span).

flutter test: 2041 passed, 39 skipped. flutter analyze: clean. Changelog entries are under ## Unreleased.