[Essoz]

Every Tensor metadata conversion repeats stable base-class lookup checks, and
scalar arguments undergo a throwaway JSON serialization before their final
serialization. This change caches only immutable lookup-hook facts and skips the
compatibility probe for exact None/bool/float/str and signed-64-bit integers.

All descriptors and mutable class/instance state remain live. The cache adds no
references to application descriptors or their names; a regression checks the
observable reference counts. Existing field selection, getter order, warnings,
method-binding lifetimes, compile/FakeTensor handling, larger-integer failures,
snapshot timing, JSON encoding and trace records are preserved.

Ordinary labmach CPU comparisons use four fresh runs per revision in ABBA×2
order, one pinned performance core and one native thread. Full API tracing is
enabled; model tracking is disabled for timing. These are measured operation/
training-body reductions, not GPU or total process-runtime claims.

Workload	Baseline	Candidate	Less time
800 add/mm calls, Tensor + Parameter	1.516760 s	1.479685 s	2.44%
MNIST, 10 steps	0.667927 s	0.646459 s	3.21%
Transformer, 10 steps	1.771281 s	1.725027 s	2.61%

One paired operation trial and one paired MNIST trial were slower; the report
preserves every sample.
The scalar component alone improves 35–37% on two local series, but that is not
the training speedup. Fixed cache containers are 344 B/process in the local
runtime, excluding code/class objects and allocator overhead. Python-validated
pickle capture, guarded JSON fallback bypass, and byte batching did not justify
a production change; their measurements and limits are in the report.

Validation: 70 focused Python 3.10 regressions and 55 Python 3.12 tests pass;
configured changed-file hooks, explicit new-test lint/format and strict scoped
mypy pass. Full suite: 357 pass with two failures independently reproduced on the
unchanged baseline (test_traincheck_internal_attr_hidden, test_inject_eval_loop).
All 24 instrumented operation/training runs match the recorded plain state,
losses and RNG. Twelve complete ordered raw comparisons retain every ID/time
field and find only those process-specific differences. Separate three-step
combined API+subclass-model-tracker runs also match recorded state/loss/RNG and
API/VAR payloads. Existing Transformer pre/post observer coverage remains partial
and is documented; no invariant checker ran. H100 validation is deferred until
access returns; no GPU or online-checker improvement is claimed.

See the measurement report
for raw evidence locations, commands, source/harness hashes, memory/drain scope,
repeat timings, output comparison limits and rejected experiments.

Dependency: targets codex/metadata-method-fastpath (#162).
Baseline: bccc56fab83d525198286f1855b9dd7f6866135f;
corrected implementation: 161313a93512b07410fa2b1072613333a5a2e492.
Earlier f9b221f timings are historical and are not used for the final results.
Commits are signed. No merge has been performed.