perf: cache plain text stream checks

[sam-saffron-jarvis]

What

• Adds an incremental plain-text eligibility check for assistant stream tails.
• Reuses the previous plain-text scan result while streamed content grows by ordinary prose characters.
• Falls back to the full markdown/math scanner when punctuation or newlines could introduce markdown syntax.

Why

During long plain-text responses, the web UI was rescanning the entire streamed assistant body every render just to decide whether it could keep using the cheap text-node append path. That is unnecessary for the common case where deltas are just words/spaces.

This keeps the conservative markdown behavior, but avoids repeated full scans for the boring case. Boring, in UI perf, is where the money is.

Tests

• node internal/serveui/static/markdown_streaming_test.js
• go test ./internal/serveui ./cmd
• go build ./... && go test ./...

[SamSaffron]

Consolidated this idea into a single local changeset on main with the other selected PR ideas.