fix: SQLite writes lack retry logic under concurrent load

[KJ-AIML]

Problem

memctrl/store.py sets busy_timeout=30s but has no exponential backoff retry. If two agents (CLI + MCP server) write simultaneously, the second gets database is locked and fails permanently.

Fix

Add a _retry_write() helper with 3 attempts (50ms / 200ms / 500ms). Wrap all write operations:

• insert_memory
• delete_memory
• update_memory
• clear_memories
• insert_tree_node
• save_provenance

Acceptance

Add test_concurrent_writes_dont_crash using threading to verify no unhandled OperationalError.

[KJ-AIML]

Implemented in branch fix/credibility-hardening-2026-05-27.

Changes:

• Added _retry_write() helper with exponential backoff (50ms / 200ms / 500ms)
• Wrapped all 12 primary write methods + 7 additional batch/atomic paths
• Removed unused _with_retry decorator dead code
• Added test_concurrent_writes_dont_crash (20 threads, zero errors)

Audit note: all write paths now protected except _init_db (schema init, runs once at startup, negligible contention risk).

[KJ-AIML]

Fixed in PR #6.