GH-34: serve queries from a per-process in-memory index

[blainemotsinger]

Fixes #34.

Replaces the on-disk bleve/bbolt search index with a per-process in-memory index (bleve.NewMemOnly), eliminating the cross-instance index flock that permanently hard-failed query_knowledge with ErrIndexLocked when a second knowledge-mcp instance was alive.

Changes

• search.Index is now a thin in-memory wrapper (search.NewIndex()); queries block on a ready channel until the background IndexAll completes (same cold-start semantics)
• Deleted all disk-index machinery: search.Location, lazy open, lock waits, staleness metadata, corruption recovery; removed search/location.go, search/location_test.go, search/race_test.go
• --index and --no-index-on-startup are accepted warning-logged no-ops so existing configs keep working; --store/--global/--root/--project unchanged
• knowledge-mcp debug (doctor) drops the disk-index diagnostics (location/locked/stale) and keeps resolution/kick/watchdog self-query/doc-count/install-record
• README updated; this repo's own knowledge-store entries refreshed for the new behavior (and their YAML parse issues fixed)

Verification

• gofmt -l ., go vet ./..., go build ./... clean
• go test ./... -count=1 9/9 packages PASS
• make eval-mock 18/18 PASS
• prove evals/t/ 175/175 PASS
• Doctor smoke on the repo itself: exit 0, 10 docs indexed, zero load warnings

Search semantics (query escaping, category/project filters, result merge, scoring) are unchanged; store files remain the single source of truth; cross-process store-file write locking remains out of scope.