[Essoz]

Daniel's parallel-inference branch and the optimization PR stack diverged from main, leaving a tracer import conflict and no reviewable base PR to bring that work into main. This PR supplies that shared base; it includes Daniel's existing parallel inference/checking implementation and the already-reviewed deterministic sampling change from #139.

Merge current main ed54c7d into Daniel's tip 122e769 on a separate branch. Preserve both sides' module-path/device initialization behavior and Daniel's interpreter-shutdown guard. The conflict resolution changes only tracer import ordering and formatting required by the configured hooks; no new inference algorithm or behavior is introduced. Daniel's original branch is unchanged.

Dependencies and review order:

• This PR → Bound example reservoirs across construction and inference merges #140 → Index trace contexts and call records during invariant inference #141 → Cache process and thread identities during precondition inference #144 → Index function-call time windows while preserving pandas query semantics #147 and Add fresh-run validation and reproduce object-ID false alarms #148.
• This PR → Fix recorder re-entry and atomic trace buffer delivery #142 → Preserve TorchScript builtin identity when wrapping PyTorch APIs #143 → Handle dynamic non-string type-name metadata #146.
• This PR → Isolate pruning state between independent precondition solves #145.

After a parent PR lands in main, retarget its direct children to main before merging them; repeat through the stack. The existing PRs retain their feature boundaries and signed history. #140 is moved off the already-merged #139 branch. Merge commits are preferable for this stack because they preserve ancestry; squash/rebase merging a parent requires refreshing its descendants again.

Validation on 2026-10-26:

• Base: 11 sampling regressions and pinned Mypy 1.9 over 76 production files pass.
• Full composed stack: 140 production + 28 harness regressions pass under Python 3.10.16 / Torch 2.2.2.
• Full composed Mypy: 77 production files pass; stricter harness checks pass over 9 files.
• Shutdown smoke checks pass on this base and Fix recorder re-entry and atomic trace buffer delivery #142/Preserve TorchScript builtin identity when wrapping PyTorch APIs #143/Handle dynamic non-string type-name metadata #146: identical arguments/results/exceptions, one application call, and zero bookkeeping calls. Base tracer AST matches Daniel's after accounting for import order.
• All nine descendant branches merge cleanly into both their intended parent and main ed54c7d; all stack composition merges are clean.
• Conflict-file commit hooks pass. Full-stack Ruff/Mypy/isort pass; Black retains the same three pre-existing failures in config.py, proxy_wrapper/hash.py, and proxy_wrapper/subclass.py. No checks were disabled.
• Remote CI is intentionally disregarded because its self-hosted runner is unavailable.

This is integration/correctness work, not a new performance measurement. Earlier PR performance numbers retain their original revisions and limitations. Reproduction logs and exact composed revisions are retained in outputs/conflict-resolution-20261026/ in the experiment checkout; no traces or machine-specific artifacts are committed.