feat(core): opt-in OR fallback for zero-hit semantic searches

[kongtou20070406]

Problem

search(text) correctly preserves raw SQLite FTS5 semantics, so an ordinary
multi-token string is an implicit AND query. That is useful for precise lookup,
but a semantic-history request can return no context when even one conversational
token is absent.

On a frozen 50-item development split of the official small Mem2ActBench release
(commit b00726940b5abbe9bd324bdd7a2cb272f5c62a29), the current query retrieved
labelled evidence for 0/50 tasks. A safe, quoted OR query over at most 12 tokens
retrieved labelled evidence for 13/50 at 17.06 ms p95. A blind Luna Max reader
then improved end-to-end exact tool calls from 0/50 to 9/50 and argument F1 from
0.0313 to 0.3047 (9 paired wins, 0 harms; one-sided exact McNemar p=0.001953).
Oracle evidence reached 22/50, so this is useful but not ceiling-level recall.

Minimal proposal

Add an opt-in fallback: 'or' search option. It runs the existing raw/safe
primary query first and only retries with the existing safe tokenization joined
by OR when the primary result is empty.

• Default and explicit FTS5 behavior remain unchanged.
• Existing source/project/session/time/meta/visibility filters and rank/limit
  remain identical on both attempts.
• The skill may opt in for semantic-history recall, but not for exact scopes or
  sentinels where an empty result is meaningful.

Acceptance evidence

• Unit regression: default multi-token AND remains unchanged; non-empty primary
  results never broaden; the opt-in zero-hit path broadens; visibility/meta and
  structural filters still apply.
• Enabled/disabled retrieval and downstream-reader ablation.
• Frozen private-history regression, including an unanswerable abstention guard.
• Query latency remains within the existing gate.