[Essoz]

Metadata conversion repeatedly tests ordinary functions and bound methods against PyTorch tensor/parameter/module types, then discards them. In the final baseline profiles, built-in isinstance consumed 12.39% of profiled MNIST time and 15.27% of profiled Transformer time.

Recognize exact built-in callable types before those framework checks. Preserve _TRAINCHECK values and retain the existing path for custom objects, custom metaclasses and dynamic __class__ behavior. Attribute getters, warning scopes, ordering, identities, tensor contents, hashes, errors and tracing coverage stay unchanged. Changed helpers have explicit type contracts.

Final matched study: five alternating fresh-process pairs on labmach CPU, three cold steps, both API and subclass model tracing active:

Workload	Baseline median	Candidate median	Reduction in traced training-body time
MNIST	667.776 ms	601.247 ms	9.96%
Small Transformer	959.440 ms	782.596 ms	18.43%

Every one of the ten pairs improves. Baseline 7a40649d and candidate 035f7f19 include the same final #154/#155/#157 changes; only dumper.py differs in their 78 executed production files. These numbers supersede the initial study for this final composition. Normal API exclusions remain enabled (use_full_instr=False); this does not measure the broader flag, GPU/MPS/distributed training or selective checking.

All 30 final cases (six correctness, 20 ordinary, four profile) match explicit plain-training references for initial/final model state, every loss and Python/NumPy/PyTorch RNG progression. Raw trace counts/bytes and complete API pairs agree between variants; all tracked parameters show changing data hashes. Hash algorithm/backend and call counts are unchanged. Transformer manual SGD remains intact, but existing observations do not establish gradient snapshots or pre/post observations for every manual update.

The denominator is the traced training body. Whole-process medians do not improve in these short jobs: imports, setup, writer shutdown and shared-host variation dominate. Median sampled peak RSS is 556.8→561.5 MiB for MNIST and 515.7→515.7 MiB for Transformer; full ranges and individual runs are in the report.

Validation:

• Seven metadata-contract regressions pass on both baseline and candidate under PyTorch 2.2.2 and 2.16 development builds, including exact framework lookup side effects, identities, mutation, warning restoration and propagated errors.
• Final composed suite: 246/247 pass; the sole test_inject_eval_loop failure also occurs on unchanged main. Pinned hooks and configured mypy across 78 production files pass; strict new-code/test checks with installed Torch types pass. Stronger legacy dumper checking has 22 baseline versus 19 candidate errors, with none introduced.
• The measurement report and adjacent JSON contain exact commits/trees/data hashes, all pair values, profile denominators, reproducible commands, resource accounting and limits. Raw evidence and checked harnesses remain in the documented experiment directories; the initial study remains in signed history and outputs.

File writing is already buffered/background on this path; metadata conversion and JSON encoding remain synchronous. Final encoding is only 0.31–0.50% of these profiles. Warning setup remains a measured next target; asynchronous live-object conversion needs a separate snapshot/error contract. This PR adds no asynchronous behavior or hashing/algorithm changes.

All commits are signed. Targets main directly. Production change is independent of the other PRs; the final measurements validate the composed result.