[Feature]: perf: cross-layer performance optimization across TUI / Desktop / Provider / EventWire

[ADX15xs]

Version line

v2 — Go rewrite (1.x), main-v2 (active development)

What problem does this solve?

Reasonix has two frontend runtimes — a TUI (Bubbletea) and a Desktop app (Wails + React) — that share a common control.Controller backend. Over time, profiling revealed three categories of performance bottlenecks:

1. Rendering frame rate (TUI) — The transcript view re-wraps the entire conversation on every token delta; lipgloss styles and goldmark renderers are recreated every frame; the status bar is built twice per frame (once in Update counting lines, once in View displaying them).
2. IPC frequency (Desktop) — Every streaming token delta is emitted as a separate Wails bridge call; the Go side allocates a new string per event via Emit; no coalescing or backpressure exists; the React side rebuilds component trees on every update without useDeferredValue.
3. Allocation / GC pressure (cross-layer) — SSE stream parsing copies every line from []byte to string; tool call arguments use += for delta accumulation (O(n²)); shell command output accumulates via string concatenation; JSON parsing for the todo panel fires every frame regardless of whether the data changed; project files are re-read from disk on every access.

These issues are most visible during long streaming conversations (>10k tokens), repeated tool call execution, and large shell command outputs — all common in daily use.

Proposed solution

Deliver three batches of zero-risk, independently verifiable optimisations across all four layers:

Batch 1 — Immediate gains (low risk):

• Replace SSE string operations with zero-copy []byte (scanner.Bytes() + bytes.HasPrefix/bytes.Equal/bytes.TrimSpace)
• Cache lipgloss Style objects at the package level, rebuilt only when width or colour changes (wrap, theme, queue indicator styles)
• Cache goldmark mdRenderer instances per terminal width via sync.Map
• Replace += string concatenation with strings.Builder in connectorBlock, MCP view width calculations, and shell output accumulation
• Incremental flushableMarkdownPrefix using a lastParagraphBreak byte offset
• File caching (loadProjectsFile with a 5-second TTL and sync.RWMutex)
• Single-pass streamedRows rune scanning instead of strings.Split + regex
• Replace map[event.Kind]string with a [KindCount]string array in the wire layer
• Extract event.Coalesce() as a shared exported function for continuous deltas

Batch 2 — Core optimisations (medium risk):

• Incremental transcript wrapping (wrapDirtyFrom + lineWrapCounts), only re-wrapping the dirty tail instead of the entire transcript
• Eliminate double status-bar construction via buildStatusLines caching (one call in Update, zero in View)
• Reuse package-level buffers (renderRowsBuf/renderBarBuf/blankRow) in renderTranscript, plus strings.Builder output assembly
• Desktop Go event coalescing: buffer consecutive Text/Reasoning deltas, flush on non-coalescable events, async emitter with ring-buffer backpressure (drop oldest 1/4 at queue capacity 1024)
• Desktop React: memo on Transcript/StatusBar/UserMessage components, useCallback stabilisation for handleToggleColdPage/handleToggleWarmTurn, 77 useCallback wrappers in App.tsx
• Wire type deduplication (remove desktop/wire.go, share eventwire.ToWire)
• scrollVersion keyed on stable identifiers (id + kind + status) only, ignoring volatile text.length
• return s guards in applyEvent reducer to return the same reference when nothing changed

Batch 3 — Dependent / cross-cutting optimisations:

• Tool call argument accumulation via argsBuilders map[int]*strings.Builder in both OpenAI and Anthropic SSE handlers
• Rewrite coalesceEventDeltas with a single strings.Builder per coalesced run, eliminating O(n²) string copy during long flush intervals
• Cache JSON parsing in the todo panel (todoParsedArgs/todoParsedTodos, invalidated via setTodoArgs/refreshTodoCache)
• Cache bottom-panel row count (buildBottomPanelsRowCount() called once in Update instead of rendering twice)
• useDeferredValue(text) in MarkdownRenderer to defer expensive markdown rendering during streaming; extract REMARK_PLUGINS/REHYPE_PLUGINS as module-level constants

All optimisations preserve the control.Controller transport-agnostic contract, maintain existing test coverage, and are independently verifiable (go vet + go test + npx tsc).