Consolidate ring buffer implementation and clean dead code paths #34

Part of #32

Problem

Ring transport was added on top of the existing native backend. Over time, ring-related branches were added into HookedRootModule methods that ring never actually calls in production.

Dead code in hook_points.py

Ring's actual call chain (both HF and vLLM paths) only touches:

• HookPoint.forward() → torch.ops.ring.producer()
• HookPoint.name setter (converts hook name → ring type/id)
• HookPoint.enabled flag

It never enters these HookedRootModule methods:

• prepare_monitoring() — ring enters and immediately returns
• run_with_cache() — ring branches at lines 717-728 and 780-782 are unreachable; HF ring path uses generate_with_monitoring, vLLM uses DMXGPUWorker.execute_model
• get_caching_hooks() — ring skip logic at lines 1004-1006 is unreachable for the same reason

All _using_ring_transport / _using_native_backend if/else branches in these methods are dead code for ring mode.

Hook type constants triplicated

HOOK_TYPE_RESID_PRE = 0, HOOK_TYPE_LN1 = 1, etc. are defined in three places:

• Python: ring_transport.py (lines 51-71)
• C++ enum: csrc/ring/tensor_meta.h (lines 28-50)
• C++ display names: csrc/ring/tensor_meta.h (lines 56-67)

Plus a legacy name mapping dict in ring_transport.py (lines 221-242). Adding a new hook type requires synchronized edits to all four locations. Should define once and share.

No-op forward hooks

_make_ring_hook() in ring_transport.py creates empty hooks that do nothing. install_ring_hooks() registers them but the actual work is in HookPoint.forward(). Dead code.

Tasks

• Remove dead ring branches from HookedRootModule (run_with_cache, prepare_monitoring, get_caching_hooks)
• Single source of truth for hook type constants
• Remove no-op forward hook registration (_make_ring_hook, install_ring_hooks)