[codex] fix raw object key idempotency

[sarthakagrawal927]

What changed

• Made raw upload object keys content-hash based instead of filename-based.
• Added a regression test that uploads the same bytes under two filenames and asserts the same object key is reused.

Why

• The file table is already deduped by (project, domain, content_hash), but the object-store path still encoded the filename.
• That meant the same bytes uploaded under a different filename could create a second blob and orphan the first one, which breaks the claimed content-hash idempotency.

Validation

• uv run pytest -q tests/test_objects.py
• uv run ruff check src/kb/storage/objects.py tests/test_objects.py

[sarthakagrawal927]

Superseded by #5 which includes all changes from this PR (content-hash keying) plus the path-traversal safety guard and pgvector parent_id fix.