retro: Fix PR #600 - only add permissions we actually use

## Retro Findings from PR #600

The implementation added too many speculative permissions instead of following tracer bullets approach.

## Problems
1. **Added untested tools** - Added Read, LS, Glob, Grep, Task, TodoRead, etc. without knowing if they actually prompt
2. **No confidence** - "there are many tools here i have never heard of"
3. **Should add one-by-one** - As we hit actual prompts, not speculatively

## Correct Approach
- Start minimal
- Add permissions ONLY when we get prompted
- Test each addition
- Build confidence through actual use

## Action Items
- [ ] Revert to minimal set of KNOWN prompting tools
- [ ] Document which specific prompts we're solving
- [ ] Add new permissions incrementally as needed
- [ ] Each addition should reference the actual prompt that triggered it

## Example of Better Approach
```json
"allow": [
// Only tools we've actually been prompted for:
"mcp__filesystem__list_files", // prompted when doing X
"LS", // prompted when listing directories
// Add more as we discover them
]
```

Related: #601 (documentation bloat)
Principle: tracer-bullets

Start by inspecting the current permission configuration and PR #600, then compare the added permissions with the actual prompts they address. Revert to the minimal known set, document each supported prompt, and ensure future additions reference a prompt that triggered them.

Written by the indexing model from the issue text.