[Essoz]

Wrapping a registered PyTorch builtin changes its function ID, so TorchScript loses the existing operator mapping. A real-file tanh-based bias-GELU function then fails to compile with TypeError: ... expected, got builtin_function_or_method, although eager execution still works.

Before publishing a wrapper, copy the original function's existing TorchScript builtin mapping. Feature-detect the private lookup/registration APIs and leave unregistered functions unchanged. Retain registered wrappers so Python cannot reuse an ID still present in TorchScript's registry.

Validation:

• Fresh-process CPU reproduction changes both broken wrapper controls from compilation failure to success with aten::tanh in the graph. Outputs, input gradients, and bias gradients exactly match the unwrapped baseline within each platform.
• All 13 semantic/trace checks and four helper tests pass on macOS/PyTorch 2.2.2 and Linux/PyTorch 2.2.2+cu121. Eager tracing still emits one paired tanh call with arguments and return snapshots.
• The portable checkout suite also passes against the trace-integrity source plus this integration: four helper tests and one regression covering six fresh subprocesses.

Scripted graphs execute the operator directly and emit no inner Python tanh records. This fixes the focused builtin-identity failure; full Megatron startup and checkpoint-shape compatibility remain unverified. Registered wrappers stay alive for the process lifetime, so repeated creation retains additional objects.

Stacked on #142 (codex/trace-integrity). See docs/experiments/torchscript-builtin-validation.md for controls, reproduction, and compact evidence.

Engineering hardening

Added a discriminated compatibility-result type, typed registry and wrapper factory, validation of present callbacks/operator metadata, per-wrapper status metadata, and a warning emitted once per process for absent private support (d4ebb12). Eight added tests cover malformed metadata, callback failures, unsupported capability, and warning behavior. All 13 TorchScript tests (including six fresh subprocess controls) and 15 tracing tests pass. Stronger Mypy 1.9 checks pass for the compatibility module and contract tests. Stronger legacy tracer/dumper diagnostics decrease from 7 to 4; the four remaining diagnostics predate this change and are documented.

Separately composed updated production sources pass 140 affected Linux regression tests, whole-source Mypy (77 files), Ruff, and isort. Whole-source Black still reports three unchanged legacy files, down from six in the original base; changed-file hooks pass. Strict checks are scoped and do not establish repository-wide strict typing. Historical performance numbers above retain their original measured revisions; this hardening pass does not re-estimate them.

Integration with current main

Based on #142 (codex/trace-integrity), with signed merge checkpoint f37be1efcb50. The shared base is #150, which reconciles Daniel's parallel-inference tip 122e769 with main ed54c7d. History is preserved; no force push or GitHub PR merge was used. All nine PR heads merge cleanly with their declared bases and current main, and the full stack composes without conflicts.

The resolved composition passes 140 production + 28 harness tests locally (Python 3.10.16 / Torch 2.2.2), pinned Mypy over 77 production files, and stricter Mypy over 9 harness/test files. Conflict-file commit hooks pass. Full-stack Ruff/Mypy/isort pass; Black retains the same three pre-existing formatting failures (config.py, proxy_wrapper/hash.py, proxy_wrapper/subclass.py). Remote CI is ignored because the self-hosted runner is unavailable. These checks validate integration; existing performance measurements retain their original revisions and are not remeasured here.

Land the parent in main first, then retarget this PR to main before merging it. Repeat for its children. Use merge commits to preserve stack ancestry, or refresh descendants after a squash/rebase merge. Exact heads and local logs are retained in outputs/conflict-resolution-20261026/.