Block mutating tool calls in pi extension to surface context before edits

[gregology]

Closes #68

The pi extension was delivering when: before context after the edit had already happened — it stored context during tool_call and then appended it to the tool_result. That completely defeats the purpose of before timing for mutating tools.

Now, when pi calls edit or write, the extension checks for matching context and if there is any, blocks the tool call with { block: true, reason: ... }. The agent sees the context first, then re-applies its change with that knowledge. Non-mutating tools (like read) still get context appended to the tool result as before.

The alternative was to inject context into the tool input rather than blocking, but pi's block API is the right fit here — it forces the agent to actually process the context before proceeding rather than potentially ignoring injected input. Blocking also gives the agent a chance to revise its edit based on the context, which is the whole point of before timing.

Updated docs in getting-started.md and cli-reference.md to describe the new behavior.

Dev Ghost metrics

Total duration: 7m 0s
Turns: 75
Tool calls: 57
Tokens: 389,932 input / 4,954 output

Stage	Model	Duration	Turns	Tool calls	Tokens (in/out)	Cache read	Cache creation
triage	claude-opus-4-6	1m 57s	31	27	79,859 / 1,654	67,141	12,710
implementation	claude-opus-4-6	3m 45s	21	14	20,405 / 52	9,240	11,162
evaluation	claude-opus-4-6	0m 22s	5	3	53,731 / 696	44,558	9,166
docs_review	claude-opus-4-6	0m 42s	17	12	178,825 / 2,193	157,078	21,733
craft_pr	claude-opus-4-6	0m 12s	1	1	57,112 / 359	28,367	28,742

Resolves #68