perf(tui): bound live thinking stream rendering to O(1) tail

[kingsword09]

• Bound active thinking stream rendering to a constant-depth tail buffer (max 24 lines / 4096 chars)
• Avoid re-parsing the entire accumulated reasoning trace as Markdown on every delta
• Preserve the complete reasoning trace for expansion (Ctrl+O) and full-text search
• Protect surrogate pair boundaries during tail trimming
• Add unit tests for live tail bounding, full trace retrieval, and UTF-16 safety

Fixes #117

[kingsword09]

Benchmark evidence for issue #117

Added scripts/bench-tui-thinking.ts (commit 5a2ed28, runnable via bun run bench:tui-thinking). Three modes:

Mode	Purpose	Runtime
bun scripts/bench-tui-thinking.ts	compare HEAD vs pre-fix 7f3d73b on the same corpus	~1s
bun scripts/bench-tui-thinking.ts --verify	regression gate: flatness + absolute ceilings	~2s
bun scripts/bench-tui-thinking.ts --stress	100k-line stream + CPU/RSS/heap + legacy checkpoint probes	~35s

---

1. Compare mode — reproduces the reported O(n²)

Same corpus (mixed CJK prose, bold/code, fenced TS blocks, lists), same machine:

trace size	legacy last-bucket (ms/frame)	HEAD last-bucket (ms/frame)	legacy total	HEAD total
200 lines (~11 KB)	8.4	0.40	739 ms	93 ms
500 lines (~27 KB)	41.5	0.34	8.2 s	166 ms
1000 lines (~54 KB)	168	0.33	62.9 s	322 ms
2000 lines (~109 KB)	625	0.33	7.9 min	665 ms

At ~50 KB (the issue's "saturation" point) the legacy component is already at ~168 ms/frame (~6 fps). At 109 KB it hits 625 ms/frame.

2. Stress mode — 100k-line stream (5.6 MB trace)

HEAD completes the full 100,000-line stream:

• wall 34.5 s / CPU 34.9 s (single core)
• RSS 261 MB (delta +188 MB), heapUsed 24 MB, external 9.3 MB
• last 100 frames average 0.344 ms/frame, max 0.521 ms/frame
• getSearchText() retains all 5,594,601 characters for /search and expansion

Legacy cannot complete 100k lines — extrapolated to ~8.2 hours. Instead it is probed at 7 geometric checkpoints (each in a fresh view so total benchmark time stays linear):

checkpoint	legacy end-frame (ms)
500	47.6
1,000	192.3
2,000	686.0
4,000	2,572.0
8,000	9,967.4
16,000	39,654.6
32,000	158,946.1

Linear regression slope 4.927 ms/line, R² ≈ 1 (every doubling of trace → ~4× per-frame cost, confirming O(n²)). Extrapolated end-of-stream frame:

• 50k lines: 246 s (4 min stuck on one frame)
• 100k lines: 493 s (8.2 min stuck on one frame)

3. --verify regression gate (usable in CI)

All four trials pass (exit 0):

• last-bucket ≤ first-bucket + 0.5 ms (flatness)
• largest trial last bucket ≤ 4 ms/frame
• 2000-line total ≤ 1500 ms

If someone ever regresses the tail buffer back to full-trace rendering, the last bucket jumps to hundreds of ms and the gate fails immediately.

---

Summary

metric	legacy (7f3d73b)	HEAD (3920d74)
per-frame cost vs trace length	O(n) → O(n²) total	O(1) per frame
100k-line end-of-stream frame	~493,000 ms (extrapolated)	0.344 ms
100k-line total render	~8.2 hours (extrapolated)	34.5 s
speedup at 100k lines	—	~856×
tail-buffer memory ceiling	unbounded trace in Markdown	24 lines / 4 KB cap
full trace retained for search/expand	yes	yes (this.text + getSearchText())

The numbers reproduce the issue's reported "25% CPU / 594 MB / decaying throughput on a 50 KB thinking trace" and confirm the fix keeps active-stream rendering bounded regardless of trace length.