v0.71.1

v0.71.1: search completeness and verify text fixes

- search: exhaustive=true now lifts the symbol cap (was silently 25 of N);
  a capped symbol list says it is a sample; scoped fetch grows until the
  in-scope set is complete; non-positive limit clamps; exhaustive bounded
  at 2000 with a stated cap
- search: context=N results print each line once with runs merged, path
  once per file (was duplicating lines and repeating the path per line)
- verify: missed sites were dropped from the text form the agent reads

Gates on the release binary: go test ./... green; ci_invariants all
invariants held (re-baselined under scorer v3); ab_gate v0.71.0 vs this
PASS (9 pairs, recall +0.030, cost 1.14x).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EmbKfexHPJPniutoFBdq6M