Change the default host from 127.0.0.1 to ::1 across all runtime code:
- ServerArgs default host in both srt and multimodal_gen
- All fallback loopback addresses in model_runner, data_parallel_controller,
  PortArgs, shm_broadcast, encode_server, encode_grpc_server, and gpu_worker
- Use format_tcp_address for proper bracket wrapping in tcp:// URIs
- Normalize localhost to ::1 in multimodal_gen scheduler_endpoint

Tests, eval scripts, docs, and loopback exclusion checks are left unchanged.

Include the prefill bootstrap address and room number in the error
message when a kvcache transfer fails, making it easier to identify
which prefill instance is unreachable.

Log all parameters passed to store.setup() and store.setup_dummy()
before the call, making it easier to diagnose configuration issues.

The bootstrap_addr was built as f"{host}:{port}" which breaks IPv6
addresses in health check URLs and mooncake session lookups. Wrap the
host with maybe_wrap_ipv6_address so IPv6 addresses get brackets,
producing correct URLs like http://[2803:...]:30000/health.

Wrap all remaining unprotected host:port string constructions with
maybe_wrap_ipv6_address() or format_tcp_address() to prevent malformed
URLs and TCP addresses when using IPv6 addresses.

Files fixed:
- grpc_server.py: gRPC warmup URL
- http_server_engine.py: log message and HTTP URL
- model_runner.py: transfer engine session_id, init_method tcp:// URIs, log messages
- remote_instance.py: init_method tcp:// URI
- encode_server.py: ZMQ tcp:// endpoint
- encode_grpc_server.py: gRPC listen address
- encode_receiver.py: receive_url host:port
- mindspore_runner.py: dist_init_method tcp:// URIs
- ascend/transfer_engine.py: session_id
- dumper.py: ZMQ local_addr
- loader.py: instance:// and http:// URLs for remote weight loading
- remote_instance_weight_loader_utils.py: seed instance service URLs
- mooncake_store.py: master_server_ip parsing and URL (fix broken split on IPv6)
- mini_3fs_metadata_server.py: log message
- elastic_ep/expert_backup_manager.py: ZMQ bind addresses
- elastic_ep/expert_backup_client.py: ZMQ connect addresses
- bench_serving.py: all http:// URL constructions
- compile_deep_gemm.py: base_url construction
- multimodal_gen/benchmarks/bench_serving.py: base_url construction

…check

…line check

ZMQ requires zmq.IPV6=1 on each socket before bind/connect to
IPv6 endpoints, otherwise it silently fails on IPv4-only mode.

- Add info-level init/session_id logs to MooncakeTransferEngine
- Upgrade transfer failure logs from debug to error with full context
  (session_id, addresses, lengths, return codes)
- Upgrade memory registration failures from debug to warning
- Include return code, endpoint, room, and chunk info in conn.py
  session failure logs
- Wrap IPv6 endpoints in log messages with maybe_wrap_ipv6_address

- Log registered buffer ptrs and lengths at init time
- Log failing transfer block address ranges (src..src+len, dst..dst+len)
  to correlate with C++ "address not found" errors
- Add debug logging of base pointers and layer count in send_kvcache
- Fix kv_chunk_idx NameError in session failure log

Now logs: local_session_id, prefill/dst index counts, MLA/MHA backend,
return code, and includes ret/session/indices in the record_failure
message that propagates to KVTransferError.

Log endpoint, room, session_id, and payload metadata at each
ZMQ send_multipart site: aux data sends, status syncs, decode
registration, and decode init.

Wrap all 4 ZMQ send_multipart calls with try/except:
- Log debug before send (what we're about to send)
- Log debug after successful send
- Log error on exception with full context, then re-raise

SGLang default log_level is 'warning', so info/debug logs are
suppressed. Upgrade all new diagnostic logs to warning level so
they appear with default configuration.

- Log why a session is being skipped (num_failures, endpoint, room)
  when hitting the "not alive" early exit
- Prefix session failure log with "Marking session as failed" for
  easy correlation with later "not alive" messages
- Log when a session is cleared from failed state on re-registration
- Upgrade KVArgs registration log from debug to warning

- isort: fix import ordering in gpu_worker.py
- ruff: remove 3 unused imports (socket, is_valid_ipv6_address, maybe_wrap_ipv6_address)
- black: reformat long lines and ternary expressions across 8 files

Add fallback in get_zmq_socket_on_host: if binding to a specific address
fails (e.g., address from get_local_ip_by_remote() is not assigned to any
local interface due to tunneling/NAT66), fall back to binding all interfaces
with tcp://*.