[Refactor][TCPCG] Remove obsolete runtime registries and split-op plumbing (8/9)

[Oasis-Git]

This is step 8/9, split from the original TCPCG removal PR #41634.

Depends on #42290 (6/9) and #42291 (7/9); merge those dependencies first. This branch targets upstream main, so its diff includes unmerged prerequisites. Review this step alone using the isolated step-8 diff.

Series

1. Relocate shared graph tensor and DSA head-gate helpers — [Refactor][TCPCG] Relocate shared graph tensor and DSA head-gate helpers (1/9) #42285.
2. Add explicit batch support to BCG eager regions — [Refactor][TCPCG] Bind live forward batches in BCG eager calls (2/9) #42286.
3. Retire TCPCG backend selection, configuration, documentation, and tests — [Refactor][TCPCG] Retire backend selection and configuration (3/9) #42287.
4. Consolidate Radix attention eager regions and isolate Inkling-specific behavior — [Refactor][TCPCG] Consolidate attention eager regions and isolate Inkling policy (4/9) #42288.
5. Migrate DSA and DeepSeek eager regions — [Refactor][TCPCG] Migrate DSA and DeepSeek graph breaks to eager methods (5/9) #42289.
6. Migrate MoE and Mamba eager regions — [Refactor][TCPCG] Migrate MoE and Mamba graph execution to eager methods (6/9) #42290.
7. Migrate diffusion eager wrappers — [Refactor][TCPCG] Standardize diffusion eager graph wrappers (7/9) #42291. Depends only on step 2.
8. Remove obsolete runtime contexts, registries, and split-op plumbing — this PR.
9. Simplify quantization paths separately because of their torch.compile implications.

The disputed relocation of inactive compiler/reference files is excluded pending a separate decision.

Changes

After caller migration, remove the legacy runtime wiring that BCG no longer needs.

• Stop publishing the TCPCG forward context and remove runtime module registries and split-op registration use.
• Replace attention/MoE layer collections used for graph eligibility with an attention-layer count and companion-layer flag.
• Remove obsolete graph branches and transitional eager-decorator compatibility; update runners, platform hooks, collectives, and test utilities together.
• Keep the torch.compile framework and torch_compile_decoration.

The isolated comparison uses refactor/tcpcg-migrations-base, an integration base containing both the SRT migration through step 6 and the independent diffusion migration in step 7. It therefore shows only this cleanup, without repeating those migrations.

Inactive compiler/backend/context files remain at their original paths pending the separate reference-code placement decision. Quantization-specific context indirection is handled in step 9.

Validation

Pre-commit passed. The combined CPU suite at this stage passed 68 tests plus 16 subtests using the macOS import shim. Accelerator graph capture/replay, end-to-end model accuracy, and full main CI at this intermediate stage remain to be validated.

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.