Security: fix medium-severity vulnerabilities (v0.11.7)

[mack42]

Fixes the nine Medium findings from the security audit. Version bumped to 0.11.7. cargo check is clean (no warnings); 85 unit tests + all integration tests pass except the two pubsub integration tests in the untracked WIP tests/ dir that fail identically on main (pre-existing, unrelated).

Findings addressed

15. Insecure defaults / no protected mode. New --protected-mode flag (default on). When enabled and no requirepass/ACL password is configured, non-loopback clients are refused (RESP -DENIED on the plain listener, drop on TLS). Operators binding publicly without auth must set a password or --protected-mode no.

16. Key-pattern ACL only checked args[1]. Added a per-command key spec (command_keys) so all key arguments are checked: multi-key commands (MGET/MSET/DEL/EXISTS/RENAME/SMOVE/…) are fully covered, and non-key commands (PUBLISH/SELECT/SCAN/…) are no longer mis-treated as keyed. Applied on both the RESP and HTTP command paths.

17. AUTH backoff was per-connection (reset by reconnecting). Failures are now tracked per peer IP in shared state, so the exponential backoff survives reconnects and can't be parallelized away; the counter decays after a 60 s idle window and clears on success.

18. SUBSCRIBE/PSUBSCRIBE zero-channel usize underflow panic. These are intercepted before the registry arity check; added an explicit empty-argument guard.

19. Predictable temp filenames / symlink TOCTOU. RDB/AOF (and BGREWRITEAOF) now write through a shared helper that creates temp files with O_EXCL (create_new) and a randomized suffix, so a pre-created symlink can't be followed, then atomically renames.

20. Dump/AOF world-readable. Persistence files are created with 0600 on Unix.

21. RDB CRC64 mismatch ignored. A checksum mismatch now fails the load (returns an error) instead of logging a warning and trusting the data; a stored CRC of 0 still means "checksum disabled".

22. Non-deterministic commands in scripts. Blocked RANDOMKEY/TIME/SRANDMEMBER/ZRANDMEMBER/HRANDFIELD and the SCAN family inside scripts to keep script effects deterministic (matches Redis's historical requirement).

23. XADD sequence overflow + GETRANGE underflow. XADD uses saturating_add for the auto-sequence (exhausted IDs are then rejected by validation rather than panicking); GETRANGE has an explicit empty-value guard.

Notes / residuals

• B8: Built-in rate limiting primitives #22: in this codebase EVAL is not logged to the AOF as a write command and script effects aren't replicated verbatim, so the divergence risk is currently latent; the block is defensive/forward-looking. The proper long-term fix is effects-based replication of script writes.
• B3: Time series data type #17: ACL LOG remains a stub (returns []); making it functional is a feature, not included here.
• B2: Probabilistic data structures #16: the key spec covers common commands; unmapped commands default to treating args[1] as the key (conservative). Channel-level ACLs for pub/sub are still not enforced (separate Low finding).
• Protected mode triggers on any non-loopback peer when no auth is set, regardless of an explicit --bind; operators intentionally exposing an unauthenticated instance must opt out.

Adds a command_keys unit test. Low findings remain out of scope.