Parity gaps against the Python agent (rules refresh, key stamp, stats, docs) #36

I audited guard-agent-php 3.2.4 against the Python agent (guard-agent 3.2.4) and found these divergences, each reproduced with a failing test first.

• start() schedules the dynamic rules refresh without a project id. The reference runs the rules loop only if self.config.project_id (client.py:128-129).
• Persistence keys are stamped with hrtime(true), a monotonic clock. The reference stamps wall-clock time.time_ns() (_buffer_redis.py:54). The AGENTS.md claim that reload is oldest-first is also untrue: the reference does not sort on reload and neither does this agent.
• EventBuffer::getStats() lacks auto_flush_running, which the reference buffer stats carry (buffer.py:57-74).
• AGENTS.md is stale: it says version 0.1.0 with no shipped tags, a guard-agent/<version> User-Agent (the code sends guard-agent-php/), omits the encryption onError stage, and claims the Python, TS and Go agents sign post-gzip bytes (they sign the uncompressed body).

Stats keys: the agent, buffer and transport key sets otherwise equal the reference. The arrays are host-facing only (the PHP adapters merge them into agentStats) and are never serialized into a status push, health endpoint or SaaS payload, so I keep camelCase and document the mapping.

The persist-at-flush, tick or shutdown deferral stays as is (maintainer-approved in #33). I only make the docs state it accurately.