fix(memorize): silent memory item drops in XML parser and LLM-mode retrieval

[Meur3ault]

📝 Pull Request Summary

Two parser bugs in src/memu/app/memorize.py that silently drop memory items during memorize(), plus a follow-up so the recovered items stay reachable from LLM-mode retrieval.

---

✅ What does this PR do?

Three atomic commits in src/memu/app/memorize.py:

1. Preserve memory items with empty <categories> (cbfac4b)

_parse_memory_element required both <content> and <categories> to be truthy:

if memory_dict.get("content") and memory_dict.get("categories"):
    return memory_dict
return None

But the prompts at prompts/memory_type/{profile,event,knowledge,behavior}.py all state "categories": [...can be empty]. Memories that don't fit a pre-configured category were dropped with no log line. Now only <content> is required; empty categories defaults to [].

2. Expand the XML root-tag whitelist (e116f4d)

_find_xml_boundaries recognised only ["item", "profile", "behaviors", "events", "knowledge", "skills"]. MemoryType in database/models.py includes tool and singular behavior/event/skill. The whitelist is extended to cover both forms of every MemoryType. Latent today because every built-in prompt wraps in <item>, but any custom prompt using semantic root tags returned [] with only a Could not find valid root tag warning.

3. Link uncategorized items to a fallback category (8284ff1)

After fix 1, items with no matched category land in the DB but with no CategoryItem relation. LLM-mode retrieval joins items via that table (retrieve.py:_format_items_for_llm), so those items remained unreachable from the LLM path. (RAG-mode was unaffected because its recall_items does a global vector search.)

An uncategorized category is now auto-created at category init (memorize_config.enable_uncategorized_fallback, default True). Items whose extracted category names match nothing get linked to it. The category is seeded with a static summary at creation (equal to its description) because retrieve.py:_rank_categories_by_summary filters on cat.summary; the seeding goes through update_category() so it persists on SQLite and Postgres backends where get_or_create_category returns a detached or copied instance. Dynamic summary updates skip the fallback since aggregating heterogeneous items into one summary is noisy and wastes tokens.

---

🤔 Why is this change needed?

The first two bugs cause silent data loss during the core memorize() flow — extracted facts simply never reach storage and there is no log to attribute the loss to. The third change closes the gap so that data preserved by fix 1 is also retrievable through the LLM-mode path, not just RAG-mode.

Related discussion: the long-term memory drift reported in #381 is made strictly worse by silent drops, so fixing this is upstream of any drift work.

---

🔍 Type of Change

• Bug fix
• New feature
• Documentation update
• Refactor / cleanup
• Other

---

✅ PR Quality Checklist

• PR title follows the conventional format (fix: ×3 commits)
• Changes are limited in scope and easy to review (three atomic commits, ~50 source lines, ~330 test lines)
• Documentation updated where applicable (config field has a description=; no public API docs needed)
• No breaking changes — the new enable_uncategorized_fallback flag defaults to True and adds one uncategorized entry to list_memory_categories(); setting it to False restores the old behaviour exactly
• Related issues or discussions linked (see "Why" section)

---

📌 Optional

• Edge cases considered:
  • User pre-configures a CategoryConfig(name="uncategorized") → not duplicated, user's config wins
  • User's existing uncategorized has a summary → not overwritten (not cat.summary guard)
  • Reinforcement skip path: items being reinforced keep their original category links
  • All three DB backends (in-memory, SQLite, Postgres) verified via update_category() flow
• Follow-up tasks (left for separate issues):
  • MemoryCategory.embedding is set at init but never consumed by retrieval (route_category re-embeds cat.summary instead)
  • Cold-start retrieve returns [] from _rank_categories_by_summary when no category has been summarised yet — the fallback's seeded summary partially mitigates this

---

23 new unit tests across tests/test_xml_parser.py and tests/test_uncategorized_fallback.py, no API key needed. The fallback tests use the in-memory backend with a deterministic stub LLM client, so they exercise the real category creation, item persistence, relation linking, and summary-update code paths end-to-end. make test reports 100 passed, 1 skipped; make check passes pre-commit, mypy, and deptry (the pre-existing uv.lock 1.5.0 vs pyproject.toml 1.5.1 mismatch on main is not touched by this PR).