fix(memory): keep numpy scalars out of metadata and API JSON

[Jehu]

Memory API handlers crashed with TypeError: Object of type float32 is not JSON serializable whenever they returned document metadata containing similarity scores (e.g. the Memory dashboard search returned a 500 instead of results).

Root cause: memory consolidation stores FAISS relevance scores in doc.metadata['_consolidation_similarity']. The scores are numpy scalars; they were pickled into the FAISS docstore and later failed json.dumps() in ApiHandler.handle_request.

• The cosine relevance-score normalizer and both metadata assignment sites now coerce scores to native float.
• handle_request serializes responses with json.dumps(..., default=float), so a numpy scalar left anywhere in a handler payload can no longer produce a 500.

Existing docstores keep their pickled numpy values until each document is rewritten; the default=float fallback keeps them JSON-serializable in the meantime. (Trade-off: a numpy integer in a payload serializes as a float, e.g. 42.0; never observed in practice — only similarity floats carry numpy types.)

Verification

• Reproduced on a live instance: dashboard search failed with the TypeError above; 41/52 documents in usr/memory/default/index.pkl carried numpy _consolidation_similarity values.
• One-off conversion of that docstore: all 52 documents round-trip through json.dumps.
• Verified default=float serializes np.float32, np.float64, and np.int64.
• python -m py_compile on the three changed files. No automated tests exist for these helpers, so validation was manual.

[3clyp50]

Thanks, the diagnosis and the source casts look right. I pushed follow-up commit 454d45c4 to narrow the compatibility fix.

I removed default=float from the shared API handler because it changed unrelated response types. For example, np.bool_(True) became 1.0, np.int64(42) became 42.0, and large integers could lose precision.

Legacy _consolidation_similarity values are now converted in the memory dashboard formatter instead. The formatter works on a copy, so stored document metadata is not rewritten as a side effect.

I added regression coverage for the cosine normalizer returning a native float, dashboard JSON serialization with a legacy np.float32 score, and preservation of the original stored NumPy value.

Verification:

• 23 focused memory, dashboard, timezone, and API tests passed
• 94 tests passed with the PR layered on the current ready branch
• the PR still merges cleanly