[Essoz]

Make subclass the default tracker for newly collected model traces. It keeps models as native PyTorch modules and tracks their parameters; proxy remains explicitly selectable through CLI and YAML. This is a default/coverage change, not a performance claim.

Help and examples explain the parameter-focused coverage and migration rule: retain explicit proxy mode for existing proxy-derived invariants, or recollect/reinfer and verify actual activation. Saved instrumentation options remain explicit and unchanged. Selective immediate-variable rules use the selected tracker; consistency-only rules retain the sampler policy.

Merge #154 first. Both PRs target main directly. #157 separately fixes the broader --use-full-instr startup recursion.

Validation:

• 12 new typed regressions cover defaults, explicit CLI/YAML choices, required saved options, selective policy, and a real parameter transition checked for active success and an active failing negative control. Two default-selection tests fail on unchanged main.
• Final composition Fix subclass tracking of Transformer masks and registered assignments #154–Snapshot constructor receivers generally without invoking application getters #157 at 035f7f1976c5d84bd01f5d96bb095fb07c9cc059 passes real two-step Transformer CLI runs: ordinary default (956 API +72 VAR), selective default (4 API +48 VAR), and full default (1,628 API +72 VAR). All 12 parameter paths appear, with complete call pairs and no duplicate raw records.
• All three final runs exactly match plain losses, parameters, gradients, cached mask and Torch CPU RNG. The intended optimizer-data invariant activates and completes two passing decisions in each mode. These counts show coverage, not timing improvement.
• The final composed suite passes 246/247 tests; test_inject_eval_loop fails identically on unchanged main 9ba90ac. Pinned hooks and configured mypy across all 78 production files pass. Strict new-code/tests checking with installed Torch types passes; stronger legacy CLI checking retains its same ten errors, with none introduced.

All commits are signed. Exact commands, source hashes, raw traces and coverage reports are retained at /Users/matt/gitrepos/traincheck/outputs/instrumenter-final-cli-20261027/validation.json; composed checker evidence is in outputs/instrumenter-final-validation-20261027. Ordinary, selective and full CLI times are single whole-process completion checks, not overhead estimates.