[Chen Yufeiyang (feiiiiii5)]

Fixes #3738.

Description

Seed deduplication used the coarse behavioral key for standalone seeds and for singleton groups, so seeds that differed only by seed type, data type, or response JSON schema were silently skipped: for example a prompt and an objective with the same value, or two singleton groups whose only difference is the response schema. Both execute differently, so only one of them was stored.

The standalone identity now covers the seed type, data type, response schema, and behavioral inputs, and singleton groups (and groups whose complete membership is already stored) are compared by the full member identity already used by multi-seed groups. Exact duplicates are still skipped and the multi-seed group path is unchanged.

Tests and Documentation

New file tests/unit/memory/memory_interface/test_interface_seed_dedup_identity.py with 5 cases:

• same value, different seed types: both stored and retrievable by objective filter;
• same value, different data types (text vs url): both stored;
• two singleton groups that differ only by response_json_schema: both stored;
• exact singleton duplicate: still skipped;
• exact standalone duplicate: still skipped.

Evidence:

# parent revision: 3 failed, 2 passed (exit 1)
# this branch:      5 passed (exit 0)
pytest tests/unit/memory (full)          1927 passed, 1 skipped
related seed/memory unit tests           1311 passed
ruff check / ruff format --check / ty    clean

No documentation or JupyText change: this is an internal dedup semantic fix, and the docstring of add_seeds_to_memory_async is updated to state the identity it now uses.

Not run: the full test suite and CI (no Azure SQL instance; the identity key is computed in Python, but that backend was not exercised locally).