Fix admin deletion cascade and valuation classification

[cursor]

Bug and impact

• Admin user deletion relied on database-level ON DELETE CASCADE, but the default SQLite configuration does not enable FK cascades. Deleting a user could leave executions, reports, chats, token usage, API keys, schedules, and other user-owned rows behind.
• The ETF/index valuation classifier treated ordinary equity names containing broad terms like Trust as ETFs/indexes. REIT-style equities such as Digital Realty Trust could skip DCF/comps and receive aggregate ETF-style valuation output.

Root cause

• delete_user only deleted the User ORM object and assumed cascades would remove all related rows.
• _is_index_or_etf used broad substring matching across metadata and names (FUND, TRUST, provider names), causing false positives.

Fix

• Explicitly delete user-owned dependent rows in dependency order before deleting the user, independent of DB cascade enforcement.
• Tightened ETF/index detection to exact instrument metadata values and specific fund/ETF name phrases, avoiding broad Trust/Fund false positives.
• Added regression coverage for SQLite user deletion and Trust-named equity valuation.
• Aligned an existing valuation bridge assertion with the cent-rounded bridge components returned by the implementation.

Validation

• python3 -m unittest backend.tests.test_admin_service backend.tests.test_valuation_summary