fix(engine): drop null metadata values at retain and recall (#3209)

[NovaLux12]

Closes #3209

Problem

MemoryFact.metadata is typed dict[str, str] | None, but retain accepts
arbitrary user-supplied JSON metadata. A null value (e.g.
{"ocr_engine": null}) was stored verbatim, and every read path that returns
MemoryFact — recall, consolidation, mental-model refresh — then failed
Pydantic validation (Input should be a valid string), permanently wedging
the affected bank.

Change

Normalize on both ends, per the plan in the issue:

• Write path — RetainContent.__post_init__ drops null-valued metadata
  keys at construction, so facts extracted from it (metadata=content.metadata
  in fact_extraction) are stored canonically. Single chokepoint: covers the
  main retain path, delta retain, recovery, and transfer import. Document
  metadata in retain_params is untouched (built from the raw request dicts).
• Read path — MemoryFact.parse_metadata drops null-valued keys instead
  of coercing them to the string "None", so already-stored legacy rows
  become readable again with no data migration. Existing string coercion for
  other non-string values is preserved.

Verification

• New tests:
  • tests/test_response_models.py (read path): null metadata value omitted,
    null inside a JSONB-string is omitted, all-null becomes {}.
  • tests/test_retain_metadata_normalization.py (write path): RetainContent
    drops nulls, keeps non-string values as-is, and _build_contents (the API
    ingestion funnel) normalizes a null-metadata request.
• uv run pytest tests/test_response_models.py tests/test_retain_metadata_normalization.py tests/test_retain_orchestrator_mapping.py -q → 27 passed.
• Consumer slice (test_large_document_replacement.py,
  test_retain_entity_prune_race.py, test_validation_result_enrichment.py,
  test_response_exclude_none.py): 16 passed; the 9 errors are all the
  same pre-existing environmental ImportError: pg0-embedded is required for embedded PostgreSQL (optional embedded-db extra not installed in this
  venv), unrelated to this change.
• ruff check, ruff format --check, and ty check pass on all changed files.

[chaoyuan-guo]

One write-side gap remains: delta updates bypass RetainContent normalization.
Both _delta_metadata_only and the partial-delta survivor sync pass raw retain_params.metadata to update_memory_units_metadata_and_tags. As a result, new or changed facts omit null-valued keys, while existing SQL-backed memory_units can receive them during re-retain. MemoryFact filters them on read, so recall hides the inconsistency.
Could these two paths use normalized RetainContent.metadata, while documents.retain_params.metadata continues to preserve the raw input?
The current _build_contents test only checks construction. Regression coverage should exercise:

• unchanged content, reaching _delta_metadata_only;
• a partial delta with unchanged survivors.

In both cases, documents.retain_params.metadata should preserve the null value, while memory_units.metadata should omit the null-valued key.