fix(agent): match the Python agent in rules refresh, key stamp and stats

[rennf93]

Closes #36

I audited guard-agent-php against the Python agent (guard-agent 3.2.4) and fixed every divergence I could reproduce. Each behavior change has a test in bin/test_agent.php (section "reference parity") that failed on main first.

• start() schedules the rules refresh only when projectId is set (null or empty means none), like client.py:128-129.
• Persistence keys use a wall-clock epoch nanosecond stamp at microsecond resolution, like time.time_ns() in _buffer_redis.py:54. Reload keeps the reference semantics (keys walked in the order Redis returns them, no sort), and AGENTS.md no longer claims oldest-first.
• EventBuffer::getStats() gains autoFlushRunning (always false, PHP has no background auto-flush), so the stats key set equals the reference set. The arrays are host-facing only and not serialized anywhere, so I kept camelCase and documented the mapping.
• Docs: AGENTS.md and CLAUDE.md (identical copies) now state the real version (3.2.4, untagged), the guard-agent-php/ User-Agent, the encryption onError stage, that the Python, TS and Rust agents also sign the uncompressed body, and that persistence at flush, tick or shutdown is a documented deviation from the reference. README and docs/usage.md are corrected the same way.

I did not change the persistence deferral (approved in #33). I did not touch CHANGELOG.md because release prep owns it.

I ran the php -l sweep, composer validate --strict, bin/test_agent.php against Redis (663 passed), bin/test_shutdown.php (9 passed), the coverage runner (100% reachable lines, no waiver line shifted because the edits keep the line numbers above them) and composer audit --locked, on PHP 8.5 locally.