[Essoz]

Selective instrumentation was omitting fields that its chosen invariants inspect. A fresh native TorchTitan trace produced 61 violated rules even after repairing a separate harness JSON-loading bug: 54 missing-return cases, 1 incomplete distinct-argument case, 5 positional-only/variadic selector cases, and 1 process-local CUDA stream-handle portability failure. Missing context-constructor inputs also made rules dormant.

This compiles the requirements into a union of needed evidence:

• Returns for ConsistentOutputRelation.
• Complete arguments for DistinctArgumentRelation and the existing InputOutputParam contract, independent of rule order or narrower overlapping requirements.
• Context lifecycle events and complete constructor inputs for context preconditions, including inverted conditions.
• Full argument metadata when the selector cannot represent a required positional-only/variadic argument, unresolved name or absent signature. Supported mappings remain selective; malformed metadata raises explicit errors.

It preserves the invariant set, predicates, literal identity conditions and checker behavior. No rule or field is dropped to force a pass. FSDP/model-tracker compatibility and live-reader issues are outside this PR.

Measured correctness result

Final H100 job 55325270, composed source c2233f1, unchanged 722 rules, 6.2M Llama, one GPU, sequence 256/batch 1:

• 61 → 1 violated rules; 716 → 719 active rules. Both the 10-step same-seed and 20-step different-seed cohorts have 718 active passes, 1 active violation, 3 dormant rules, 0 errors/unknown outcomes/skips.
• The remaining rule 614 correctly reports that the learned literal stream handle differs in the new process. Two other rules retain reference-worker-TID dormancy; a third was dormant in the reference.
• Exact complete observed/direct CheckerResult parity 722/722 per cohort; all report input hashes match audited traces and the frozen rules. All four selective traces pass raw/parsed-row and complete/unique call-pair audits.
• All three paired training trials and the holdout retain matching state-byte digests, DTensor metadata, losses and RNG evidence.

This is a correctness repair, not a speedup. Correct selective collection takes 22.64x plain steady-step time in this tiny every-step workload. Earlier lower numbers omitted required evidence and are not valid checking results. Full study and evidence provenance.

Engineering validation

Baseline main 6dccb47; this independent main-based PR contains four signed commits. 69 policy/constructor/trace-integrity tests pass on the independent branch. The test-discovery fixture repair also passes both unittest invocation styles locally (28 each); before repair discovery had 3 failures and 3 errors while qualified invocation passed. The composed local pytest selection passes 83 tests. Both invocation styles also pass 28/28 on labmach Python 3.11.4 / Torch 2.16.0.dev20261012+cpu, using the SHA256-verified updated fixture and single-thread settings.

New regressions use real wrappers, JSON policy/trace loading and actual checkers, including negative controls, inversion, literal-ID dormancy, order-independent unions, required-field failures and malformed types. Healthy output/distinct-input/unsupported-argument regressions fail before their respective fixes.

Configured Ruff, mypy, Black and isort hooks pass. New tests/helpers pass explicit stricter mypy. The stronger legacy collect_trace.py check improves 10 baseline findings to 5, with none added; this is not a repository-wide strictness migration.

python -m unittest tests.test_selective_output_policy \
  tests.test_selective_context_requirements tests.test_selective_argument_kinds \
  tests.test_constructor_snapshot tests.test_trace_integrity