fix: tool_search streak counter fires injection at 3 and tool_describe at 5 (Closes #1373)

[Lexus2016]

Automated evolution PR for issue #1373.

Root cause: callers pass session_id=agent.session_id or empty string, and the old streak counter treated any falsy value (incl. empty string) as do-not-track, so the streak never incremented and the fallback directive never fired.

Fix:

• Map empty-string session_id to stable key default_session so production fires; None still opts out (preserves pure-function test contract).
• Add per-session previous-queries tracking (rolling, capped at 8).
• At streak >= 3 (search_streak_threshold): inject full tool list + previous queries + fallback directive.
• At streak >= 5 (new search_streak_describe_threshold): additionally inject auto_describe with the top hit full schema.

Tests: 8 new tests in TestSearchStreakEmptySessionNowFires covering all four spec requirements.

[Lexus2016]

Reviewed and merging. This correctly implements the root cause my diagnostic PR (#1445) surfaced, and goes further by adding previous_queries so the model can see it is circling.

One risk worth recording rather than blocking on: all empty-session callers now share a single __default_session__ counter. Sub-agents built by delegate_task run in the same process, so if a parent and a child both reach the counter with an empty session id, their streaks add together and the directive fires earlier than either warrants.

Worst case is a nudge one or two searches early, which is mild — and the alternative (no counter at all) is what this PR is fixing. Worth watching in the next digest: if tool_search streak counts drop but the directive appears in sessions that only searched twice, this is why.

Note the diagnostic warning from #1445 is removed here, which is right — there is no longer a silent-off path to warn about. If a future change reintroduces one, the lesson is that a counter returning 0 forever looks identical to a counter that is working.