discovery: poll for the meta instead of timing out + rich logging (0.6.3)

[flymysql]

Symptom

A node started before the meta crashed the whole host process:

Failed to create dynamic storage backend 'peercache': timed out
...
discovery.py register() → rpc.call("register") → TimeoutError: timed out

DiscoveryClient.start() did a single register() RPC and let a TimeoutError propagate, killing the SGLang scheduler.

Changes

• Poll, don't time out. start() now calls _register_blocking() which retries register() indefinitely (until stop()), logging periodically — discovery: node=… waiting for meta … (attempt N, Ms elapsed): …; retrying in 2.0s. Once the meta is up it registers and continues. No more crash.
• Operability logging (the requested visibility):
  • startup: node identity + endpoints, and which node is the meta (meta-host vs client);
  • meta logs every REGISTERED / re-register / DEREGISTERED / dead-node prune, each with the current member count + id list, and flags heartbeats from unknown nodes;
  • client logs successful registration (after N attempts, with member list), every heartbeat (known=, member count), and membership changes (joined=… left=…).
• register_retry_interval knob (default 2.0s).

Verification

• New test_register_polls_until_meta_is_up: a client started before the meta keeps polling (no exception) and registers once the DiscoveryServer comes up.
• Full suite 53 passed.

Bumps to 0.6.3. Pure-Python (no extension rebuild needed).

Validate on the node (immediate)

pip install --force-reinstall --no-deps "git+https://github.com/flymysql/PeerCache@cursor/discovery-poll-logging-49da"

Start a worker before the meta — it now logs waiting for meta … and joins once the meta is up, instead of dying.