Avoid warning contexts for verified native tensor method bindings

docs/measurements/native-tensor-bindings-20261027.md

@@ -0,0 +1,238 @@
+# Native tensor bindings without repeated warning contexts
+
+The final implementation in signed commit `37f22e58954da403105c2c5510b7c7b6b3162ee6`
+reduces ordinary traced training time on the measured CPU and H100 workloads.
+It preserves native method binding and the existing metadata classifier, while
+avoiding the warning-managed getter for verified native tensor methods. JSON
+encoding, tensor hashing, instrumentation selection and the checker are unchanged.
+
+## Fresh ordinary measurements
+
+| Workload / timing scope | Baseline | Candidate | Time reduction |
+| --- | ---: | ---: | ---: |
+| CPU MNIST, complete three-step API-traced body | 286.400 ms | 222.665 ms | 22.25% |
+| CPU small Transformer, complete three-step API-traced body | 682.324 ms | 527.587 ms | 22.68% |
+| CPU, 2,000 complete metadata conversions | 1.607038 s | 1.190048 s | 25.95% |
+| H100 Llama, median of steady-step run medians | 999.855 ms | 842.885 ms | 15.699% |
+
+Percentages divide saved time by baseline time. CPU measurements have four fresh
+processes per variant, ordered ABBA twice. Component conversion is separate from
+training, starts after one warmup conversion per object and excludes tensor
+hashing; its percentage is not a fraction of total training time. CPU training uses full API collection with model tracking disabled.
+
+H100 job **55326071** uses the same 6,163,712-parameter TorchTitan Llama debug
+model, native FSDP, batch one, sequence 256, and one H100. It collects selectively
+using the same 722 frozen invariants and 244-API policy. There is no model tracker
+or concurrent online checker. Each fresh process trains ten seed-42 steps; steady
+statistics use synchronized steps 3–10. Every measured step remains in the evidence.
+
+| H100 pair / order | Baseline median step | Candidate median step | Time reduction |
+| --- | ---: | ---: | ---: |
+| 1 / AB | 999.855 ms | 842.430 ms | 15.745% |
+| 2 / BA | 1,000.320 ms | 842.885 ms | 15.738% |
+| 3 / AB | 998.851 ms | 843.650 ms | 15.538% |
+
+The **median paired reduction is 15.738% (1.1868x)**; the table's ratio of run
+medians is a separate aggregate, 15.699%. Including cold steps 1–2, median complete
+ten-step training-body time is **12.579847 to 10.990844 seconds, 12.631% less**.
+Timings exclude
+imports, setup, state evidence, outer loop housekeeping, queue drain and shutdown.
+The separate seed-43, 20-step holdout has medians of 1,002.488 to 842.039 ms over
+steps 3–20 (16.005% less); this is one validation pair, not a fourth primary pair.
+
+The single H100 plain-control run has a 44.456 ms steady median: baseline/candidate
+are still 22.49x/18.96x that time. CPU single plain controls give 8.60x/6.69x for
+MNIST and 48.05x/37.15x for the tiny Transformer. Plain controls import TrainCheck
+but install neither API wrappers nor tracked parameters. These ratios have only
+one plain repetition, and the workloads are small; they are not general overhead
+estimates. No multi-GPU scaling or concurrent-checker improvement is established.
+
+## What changed and what remains
+
+Ordinary metadata conversion visits hundreds of Tensor methods. Their values
+usually contribute no output fields, yet generic lookup enters and exits a warning
+context for each one. The shortcut verifies exact Tensor/Parameter type, ordinary
+attribute lookup, native descriptor ownership, unchanged inheritance and an empty
+native instance dictionary. It then calls the descriptor type's native binding
+operation. Properties, custom lookup, subclasses, instance shadowing, foreign
+descriptors and TrainCheck-specific fields retain the original guarded getter.
+
+Only original MRO identities and live namespace proxies are cached. Current
+namespaces and native dictionary storage are rechecked for each attribute, so
+mutable values and descriptor results are not cached. A changed MRO disables the
+shortcut, including after bases are restored. The original classifier handles
+live primitive-type configuration exactly once. Native allocation failure retains
+NOT_FOUND/backoff behavior; internal assertions remain visible.
+
+The implementation deliberately retains the real bound object. A prior temporary
+value and its type must be released at their ordinary points, while the bound
+method must keep its receiver/function alive during a later property. Four
+regressions demonstrated that the earlier skip-based candidate changed these
+behaviors. All four fail on verbatim `287b3a2` sources on Python 3.10 and 3.12 and
+pass with the final binding implementation. A fifth regression verifies that a
+function's own `__get__` attribute cannot override Python's native descriptor
+protocol. Historical CPU/H100 reports remain separately labeled; their faster
+skip-based measurements are not reused as results for the final implementation.
+
+This does not combine warning contexts around arbitrary properties, cache tensor
+metadata, change serialization, or introduce asynchronous dumping. The CPU
+component diagnostic reduces guarded getters from 1,192 to 60 per two conversions,
+while retaining all 28 field names per object. It preserves warning-filter
+contents/identity, the warning handler, tensor values and Python/NumPy/Torch RNG.
+Those are operation counts, not percentages of runtime attributed to warnings.
+
+A separate H100 diagnostic, job **55326074**, counted steps 3–10:
+
+| Scope | Binding helper calls | Native bindings without guarded getter | Remaining guarded getters |
+| --- | ---: | ---: | ---: |
+| All threads | 1,940,048 | 1,698,000 (87.52%) | 242,048 |
+| Window owner | 1,243,312 | 1,100,304 (88.50%) | 143,008 |
+| Other trace-producing thread | 696,736 | 597,696 (85.79%) | 99,040 |
+
+Exact Tensor receivers account for all native bindings: 1,698,000 of 1,788,000
+calls (94.97%). Exact Parameter receivers are unobserved, not a measured zero
+eligibility rate. The 152,048 other-receiver calls all use the fallback. Every
+successful shortcut still constructs the real method and runs classification.
+The count percentages do not attribute a fraction of total runtime to warnings.
+
+The collector preserves returned objects and exceptions, verifies bindings before
+installation and at restoration, and records zero exceptions escaping the binding
+helper or guarded getter. Internal NOT_FOUND handling remains unchanged and is
+not a separate exception counter. Its counted window takes 1.39489x its ordinary
+control's duration; diagnostic
+durations do not enter the performance table. Both corpora have 416,364 rows,
+208,182 complete calls and 142,621,017 bytes with zero raw audit issues. Every
+per-step/API/event-kind count and per-step byte total matches, as do the complete
+recorded state and policy. This proves volume/integrity parity, not raw byte equality.
+
+## State, traces and checking
+
+All 16 CPU traced runs match fresh plain controls' recorded initial/final model,
+gradient, optimizer and stored module state, RNG and losses. Every CPU MNIST run
+has 520 API records / 351,073 bytes; every Transformer run has 1,338 records /
+1,014,134 bytes. Every call has one PRE and one POST, with no VAR records. All
+eight paired raw comparisons preserve complete ordered event sequences, fields,
+arrays and duplicates. Remaining differences are identities and timestamps;
+raw equality is false and cross-process identity-graph equivalence is not claimed.
+
+All nine H100 phases pass independent source, environment, command, policy,
+complete recorded state and loss checks. Seed-42 phases match the plain control;
+the seed-43 candidate matches its selective baseline. GPU tensor evidence includes
+names, shapes, dtypes and byte hashes, rather than archived full optimizer-state
+checkpoints. No identity, tensor value or array was removed to establish state
+parity.
+
+Raw comparisons of the first seed-42 selective pair and the seed-43 holdout
+uniquely match both complete ordered per-thread streams in each pair. Each variant
+has 416,364 records / 142,621,017 bytes in the first selective pair and 715,004 /
+244,304,737 in the holdout. These full field comparisons cover those two pairs;
+raw/parser integrity audits separately cover all eight selective runs. No fields, array elements or
+duplicates are removed. Differences remain in identities/timestamps, constructor
+`object_id` and CUDA `cuda_stream`/`native_handle` values; raw equality is false.
+Constructor snapshots store literal `id(receiver)` in both source archives. All
+800/1,600 stream-metadata observations preserve alias equality and a bijective
+observed handle correspondence. No structural/type/shape/array/exception payload
+differences occur, including the same two step-one `_BypassDispatchCache` exceptions.
+No global cross-process identity graph or inter-thread ordering equality is claimed.
+
+All eight raw/parser audits and eight checker phases completed successfully for
+this exact revision. The first seed-42 selective pair and seed-43 holdout each
+execute the same 722 frozen rules. Every cohort has **719 active rules: 718 passes
+and one violation; three rules are dormant**. There are no checker errors, unknown
+outcomes, unobserved coverage or variable-check skips. Decision counts match across
+variants: 114,373 per ten-step corpus and 218,713 per holdout corpus. No variable
+precondition is requested, consistent with the API-only scope; this study does
+not establish model-tracker or variable-checking coverage.
+
+Within each corpus, all 722 complete `checker_result` values match structurally
+between observed and independent direct checking. The comparator preserves scalar
+types, fields, identities and array order; five legacy NaN literals match only at
+the same paths. Cross-variant rule outcomes and decision counts match, while
+complete checker results match for 721/722 rules. Rule 614 retains fresh-run
+identity/timestamp differences in its failure evidence; those fields are not
+removed or remapped to claim equality.
+
+The shared violation is rule 614, `cuda.stream() always calls
+StreamContext.__init__()`: its frozen child argument requires CUDA handle
+385452624. Actual handles are 386050160/386062656 in the ten-step pair and
+386059824/386070112 in the holdout. Both handle aliases agree, all other required
+fields match, and no fields are missing or extra. Dormant APIContain rules 216/276
+require reference worker ID 23438301021888; rule 717 requires unequal threads and
+has zero decisions. Rules were not reinferred or filtered. This is unchanged
+frozen-rule behavior, not evidence of zero violations or universal rule activation.
+
+Offline validation ran on two separate labmach CPU cores, one native thread per
+worker and an 8-GiB process-group cap per lane. Its scheduling and operational
+timings are separate from inference-performance claims. Source and frozen input
+inventories remained unchanged. All 98 downloaded report files were verified.
+
+## Sources, environment and validation
+
+- CPU baseline: `6dccb47d0a8bc192fbe7cbbea700b66106bc2381`.
+- H100 baseline: composed `8cc73ab35590caa1a5d7286b16462d045b6d2da0`; candidate
+  overlays only `dumper.py` and `metadata_methods.py` from signed `37f22e5`.
+  CUDA hashing, selective policy and workload code are identical across variants.
+  The production PR targets main and does not depend on the study's other changes.
+- TorchTitan: `2210414902d84d1219fd63ab8efeb24375a2a1a1`.
+- Frozen invariant SHA-256:
+  `bad55c222f4eb3406bdfaec527ab1a75eaf66cdbe8f06c42d7f0775f4ee9224b`;
+  compiled policy SHA-256:
+  `df4c91296774b2dfd3af09d0a25e21e25f26ff1de93cb97ab959fa77379c3211`.
+- Labmach CPU: Python 3.11.4, PyTorch `2.16.0.dev20261012+cpu`, NumPy 2.1.0;
+  one Torch/native thread, CUDA hidden. MNIST is the existing 1,199,882-parameter
+  Net with six cached normalized examples, NLL and Adadelta. The Transformer is
+  21,052 parameters: two width-32 layers, feed-forward width 64, two heads,
+  sequence 16, batch two, fixed WikiText prefix/vocabulary, NLL and clipped SGD.
+  Seeds, data and benchmark-definition hashes are retained in every run manifest.
+- Expanse: H100 80GB HBM3, Python 3.12.3, PyTorch
+  `2.16.0.dev20261012+cu130`, CUDA 13.0. Eight CPUs / 48 GiB allocated; GPU work
+  runs on compute nodes. Ordinary job completed in 4:36. Batch peak memory covers
+  the entire job, not individual variants; no memory-cost improvement is claimed.
+- Final local full suite: **335 passed, two failures** also reproduced on the
+  measured main baseline (`test_traincheck_internal_attr_hidden` and
+  `test_inject_eval_loop`). All 26 focused metadata tests pass on Python 3.10 and
+  3.12; the exact composed source passes 33 metadata/classification tests on both.
+  The matching labmach runtime also passes all 33.
+- Pinned changed-file pre-commit, explicit test lint/format checks and strict
+  scoped mypy pass. The legacy dumper retains its configured mypy coverage rather
+  than claiming a repository-wide strictness migration. Earlier all-files checking
+  found only the documented, unchanged Black whitespace failures in `config.py`
+  and `proxy_wrapper/hash.py`; those unrelated edits remain excluded.
+
+## Durable evidence and reproduction
+
+Local artifact root: `outputs/metadata-fastpath-20261027/` in the canonical
+checkout. CPU evidence is in `cpu-binding/`; new H100 packages and results are in
+`gpu-binding/` and `gpu-binding-diagnostic/`. Final checking evidence is in
+`validation-binding/job-55326071/validation-summary.json` and
+`focused-attribution.json`; raw comparison reports are under ordinary job 55326071's
+`trace-comparison/`. Exact commands, frozen source
+inventories, raw traces, timings, diagnostic scripts and failure logs remain
+outside Git. Historical studies are preserved separately. No credentials are
+included in source or evidence packages.
+
+| Artifact | SHA-256 |
+| --- | --- |
+| CPU evidence archive | `7418a17bd9bf05c36119ac56eb23abb28bd5280e552afafbe2a6df39931e8edd` |
+| CPU candidate source archive | `61ff85dce4b0e94633e6fc932d6bc051744fe9c17d8625cf83a5d44c7aae5381` |
+| H100 candidate source archive | `517c5afa213ec1693d42aa04f9afd61676dd1680f94b2f73734f0759dc3e68b5` |
+| H100 ordinary results, job 55326071 | `3754fa9de95c586a4338ff49a5daec3ccac67a4d40689cfb07bf09c6e66a403b` |
+| H100 diagnostic results, job 55326074 | `d6ddd8bb9852209d6f12f64e23f7f68a56d3e3d1f5535645b13b79644aa90df4` |
+| Complete fresh validation reports | `4323fff37a0085d75de432645025918b3b262ec56b316f2bb19d002690288abd` |
+
+From the artifact root, rerun saved analyzers:
+
+```sh
+python analyze_cpu_binding.py --root cpu-binding/results --archives cpu-binding \
+  --output cpu-binding/verified-analysis.json
+python analyze_binding_gpu_results.py \
+  --ordinary gpu-binding/job-55326071/results/ordinary \
+  --output gpu-binding/job-55326071/ordinary-analysis.json
+```
+
+For new experiments, use fresh output directories and the frozen packages' README
+commands. CPU runs use `cpu-binding/run_study.py --root STUDY --candidate-commit
+37f22e58954da403105c2c5510b7c7b6b3162ee6`; H100 submission verifies package checksums
+before `sbatch comparison.sbatch`. Source inventories are verified before and after
+execution. CPU processes have 90-second / 4-GiB / 256-MiB-trace bounds; H100 phases
+and jobs have explicit time/trace/allocation limits. No limit was reached.