feat(memory): trust-aware retrieval

**Context:** ROADMAP.md → Memory security → Trust-aware retrieval

---

## Component

Agent (Python runtime)

## Describe the feature

Weight retrieved memories by **freshness**, **source type**, and **pattern consistency** when assembling hydration context. Higher-trust sources rank above unverified or anomalous entries.

## Use case

All memories currently compete equally in the token budget. A single poisoned or outdated rule can dominate context and mislead the agent.

## Proposed solution

1. Scoring function in memory retrieval path (`agent/` + orchestrator hydration).
2. Metadata on memory entries: `source_type`, `verified_at`, `corroboration_count`.
3. Configurable weights per Blueprint.
4. Telemetry: log top-k retrieval scores for debugging.
5. Tests with fixture memories and expected ranking.

## Other information

- Design context: `docs/design/MEMORY.md`.

- [ ] This might be a breaking change

Start with docs/design/MEMORY.md, then trace the memory retrieval path under agent/ and orchestrator hydration. Define how freshness, source type, pattern consistency, metadata, Blueprint weights, and top-k telemetry fit together. Done means fixture-based tests verify the expected ranking and the retrieval scores are logged for debugging.

Written by the indexing model from the issue text.