docs: add Phase 2 (Retrieval) implementation plan (#6)

* docs: add Phase 2 (Retrieval) implementation plan

12 tasks: search_vector migration, normalization, full-text search,
vector search, RRF fusion, search orchestration, debug endpoint,
evaluation metrics/runner/CLI, sample-vault fixture retrofit, and the
exit-condition proof. Verified pgvector's cosine_distance() and
websearch_to_tsquery/ts_rank against real Postgres and the installed
pgvector package before writing task code around them, matching the
same discipline the spec's migration mechanics got.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* docs: fix two real bugs found in Phase 2 plan review

- Task 8: the p50 and p95 percentile tests assumed two different
  percentile definitions (linear interpolation vs nearest-rank) —
  no single implementation could satisfy both. Verified by running
  the plan's exact implementation against both test inputs. Keeping
  linear interpolation (matches numpy.percentile's default); fixed
  the p95 expected value to the correct interpolated 95.05.
- Task 1: the migration test's finally block only truncated tables,
  never re-ran the upgrade. If the raw-SQL insert failed, the schema
  was left stranded at revision 0001 for the rest of the test
  session, breaking every later db_session-based test in a way that
  wouldn't obviously point back to this test. finally now re-runs
  upgrade to 0002 (no-op if already there) before truncating.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

Author: leyder-ben