[Essoz]

Repeated function-call time-window queries scan all event rows before grouping, including full-column type membership tests. Add a lazy process/thread timestamp index for conservative native numeric cases, then restore original row order and retain the existing pandas grouping, assertions, and fresh record construction. Unsupported dtypes, mixed-precision/scalar comparisons, and malformed queries fall back to the original mask.

This PR is stacked on #144. The full-trace experiment applies the separate #145 pruning-state correction to both variants; that correctness change is not bundled in this PR.

Measured result

Three sequential alternating pairs on the complete saved tiny-Llama trace (573,494 events, six enabled relations, one CPU worker/native thread):

Pair/order	Baseline	Indexed
1: baseline, candidate	102.655 s	76.748 s
2: candidate, baseline	101.796 s	76.788 s
3: baseline, candidate	101.610 s	76.658 s

Median 101.796 → 76.748 s: 1.326×, 24.61% less wall time. All six complete outputs match the corrected reference: 836 invariants and 104 failed-hypothesis records, with arrays and multiplicity preserved. Raw output and full/stage context files are also byte-identical across variants. Source/input integrity checks passed; independent audit verified the retained source/output/context evidence and recomputed the statistics.

Median sampled child-session PSS increased 2.129 → 2.168 GiB (+40.15 MiB). The first baseline peak was higher than later baseline peaks; all per-run peaks are retained in the report. Fresh-process timings include lazy index construction. A separate 200k-row/1k-query synthetic benchmark measures 2.662 → 0.340 s, with first-query build cost shown separately.

Validation and limits

• 18 new differential tests; 61 combined local tests and 67 in the matching Linux runtime, including six shared Isolate pruning state between independent precondition solves #145 tests.
• Independent parent-method oracle: 144 dtype/offset cases and 6,480 queries in each of two pandas/NumPy runtimes, zero differences.
• Pinned Ruff, Mypy, Black, and isort checks pass. Production and benchmark checkpoints are signed.
• Conservative fallbacks cover dtype/null/category behavior, numeric precision, strict/reversed bounds, incomplete/duplicate/exception records, and unusual scalar comparisons.
• Cache contract requires stable time/type/PID/TID values, dtypes, and row order. Frame replacement or length changes invalidate it; same-size edits require a new TracePandas. Storage grows O(N).
• Shared host, three pairs on one saved trace, six enabled relations, no CPU affinity; this does not establish training overhead or distributed scaling. CPU Torch matches the original CUDA wheel's source revision, but Python/build differ. Do not multiply this incremental result by the earlier single-pair cache result.

Measurement report and compact provenance include reproduction instructions, all run timings, memory tradeoffs, source/trace/output digests, and retained archive identity. Raw trace/output records remain private study artifacts.

Engineering hardening

Typed numeric time-index storage and query boundaries, retaining conservative dtype/scalar validation and exact fallback semantics (8dc1fd1), plus signed parent hardening updates. All 97 affected inference tests, all three benchmark smoke checks, pinned hooks, and six stronger Mypy 1.9 module/tool/test scopes pass. A new bounded replay of source ce25c12 plus the common #145 production patch emits all 836 invariants and 104 failed hypotheses byte-for-byte identical to the corrected reference; both context files are also identical. The verification run took 76.794 seconds, without a new paired speedup claim. Exact source/runtime/equality evidence is in docs/measurements/inference-hardening-validation-20261025.json.

Separately composed updated production sources pass 140 affected Linux regression tests, whole-source Mypy (77 files), Ruff, and isort. Whole-source Black still reports three unchanged legacy files, down from six in the original base; changed-file hooks pass. Strict checks are scoped and do not establish repository-wide strict typing. Historical performance numbers above retain their original measured revisions; this hardening pass does not re-estimate them.