feat: bulk-load documents from the mirror bundle with prefetch

[m1ngsama]

A cold search reads every document one request at a time: 248 requests and 43 s with 1 s of latency per request.

• prefetch() fetches ${mirror}/bundle.json once and stores every entry whose sha matches the current tree and whose content hashes to that sha, so a stale or tampered bundle can't poison the cache. It returns how many were loaded, and 0 without a mirror or bundle.
• search() prefetches when more than 6 candidates are uncached. With 1 s of latency a cold search drops to 2.2 s and 2 requests.
• The bundle has its own 30 s timeout and never trips the mirror breaker, so losing it only means per-file mirror reads. clear() now also resets the breaker.
• The in-memory file cache holds 1000 entries instead of 200, so a prefetched library isn't evicted. Additive; ships as 0.3.4.