Events dropped due to telemetry buffer overflow under heavy tool usage

## Problem

During sessions with high tool call volume, the event buffer overflows and drops events silently. The error message logged is:

```
N events dropped due to buffer overflow
```

This means event counts are undercounted for power users, and we may be missing important failure/error events during the heaviest usage periods.

## Impact by version

| CLI Version | Occurrences | Total Events Dropped | Avg Dropped/Occurrence | Max Single Drop |
|-------------|-------------|----------------------|------------------------|-----------------|
| `0.5.7` | 167 | 36,968 | 221 | 3,901 |
| `0.5.1` | 43 | 7,614 | 177 | 706 |
| `0.5.3` | 14 | 2,691 | 192 | 646 |
| `0.5.2` | 9 | 931 | 103 | 290 |
| `0.5.4` | 7 | 511 | 73 | 283 |
| `0.5.5` | 2 | 9 | 5 | 5 |

**Total: ~48,800 events dropped across all versions in the last 7 days.**

## Observations

- The problem is **getting worse with newer versions** — `0.5.7` accounts for 75% of all drops
- The worst single overflow dropped **3,901 events** in one batch
- Overflows correlate with sessions that have very high `tool_call` volume (1,000+ calls in a session), especially `sql_execute` and `todowrite` heavy workflows
- `0.5.5` had the fewest drops (avg 5), suggesting something regressed in `0.5.7`

## Suggested investigation

1. Check the telemetry buffer size and flush interval — the current settings can't keep up with heavy `sql_execute` loops
2. Consider increasing buffer capacity or flushing more aggressively
3. Consider back-pressure or sampling instead of silent drops
4. Investigate why `0.5.7` is significantly worse than `0.5.5`