feat: cache discovery results for domain retry

[mvoutov]

Problem

When a domain fails during doc init, retrying with --mode chunked --domains "failed-domain" currently re-runs the full discovery (2 parallel Claude agents, ~2 min) even though we already have the results from the first run.

Solution

Cache discovery findings to .claude/skills/.discovery-cache.json after the first successful discovery run. Structure:

{
  "timestamp": "2026-03-23T...",
  "domains": [...],
  "findings": "...",
  "repoHash": "abc123"
}

When --domains is used with --mode chunked:

1. Check if cache exists and is recent (< 1 hour or same repo hash)
2. If yes, load cached domains + findings, skip discovery
3. If no, run discovery as normal

This makes domain retries instant — scan only, no LLM calls until actual generation.

Current workaround

The --domains flag now skips discovery (merged in latest), but it also loses the discovery findings context that helps Claude write better skills.