feat(memory): add task-relevance ranking with boost tags and min relevance filter

[Siddhant-K-code]

What

Adds task-aware relevance ranking to memory recall. Callers can now boost specific tags, provide task context for source matching, and filter out low-relevance results.

Why

As memory grows, the signal-to-noise ratio degrades. Returning everything that's semantically similar isn't enough — the agent needs context ranked by what matters for the current task, not just what exists.

Changes

New RecallRequest fields

• boost_tags — tags that receive a +0.1 relevance boost during ranking
• task_context — free-text task description; memories whose source matches get a +0.05 boost
• min_relevance — filter threshold (0-1); memories below this score are excluded from results

Ranking formula

relevance = (1 - recency_weight) × similarity + recency_weight × recency
          + 0.1  (if any tag matches boost_tags)
          + 0.05 (if source appears in task_context)
          + 0.05 (if text appears in task_context)
clamped to [0, 1]

Tests

6 new tests: boost tags tipping ranking, min relevance filtering, zero min (no filter), task context source boost, relevance clamping, empty boost (no effect).

Usage

# Boost auth-related memories for the current task
curl -X POST localhost:8080/v1/memory/recall -d '{
  "query": "how does login work?",
  "query_embedding": [...],
  "boost_tags": ["auth", "security"],
  "task_context": "fixing the login flow in code_review",
  "min_relevance": 0.3,
  "max_results": 5
}'

Closes #78