feat(memory): add MEMTIER five-signal retrieval scoring

[bug-ops]

Summary

Implements four retrieval scoring signals from arXiv:2605.03675 (MEMTIER) in tiered_retrieval.rs, addressing issue #3987.

• Recency decay: exp(-ln2 / half_life_days * age_days) — entries decay exponentially; lambda expressed as half-life in days for intuitive config
• BM25/TF-IDF: inline implementation (~60 lines), no new dependencies; scores each candidate against query tokens across the candidate set
• Cognitive signal: fact_access_log count proxy (normalized) — entries accessed more frequently after tool calls rank higher
• Tier boost: additive weight for consolidated/semantic entries over episodic ones

Design

All signals are applied in a new score_candidates() step inserted between recall_routed() and assemble_within_budget(). Combined via weighted sum:

final = sim_weight * similarity
      + recency_weight * recency
      + tfidf_weight * tfidf
      + cognitive_signal_weight * cognitive
      + tier_boost_weight * tier_boost

All new config fields in TieredRetrievalConfig default to 0.0 except similarity_weight = 1.0, preserving exact backward-compatible behaviour with no config changes.

Performance

• DB queries for recency and tier signals are weight-guarded — zero extra SQL in default config
• TF-IDF: O(n×m) for n≤20 candidates × ~10 query terms ≈ 10 µs
• All three new store functions instrumented with tracing::instrument

Test plan

• 1425 unit tests pass (21 skipped — require Qdrant)
• 13 signal-specific unit tests added: recency math, TF-IDF scoring, empty input, all-zero weights guard, weighted-sum formula + sort order, backward compat with default config, tier boost ranking, cognitive signal normalization

Closes #3987