[Essoz]

Parameter sampling reused the last Python attribute-assignment timestamp. In-place optimizer updates could therefore record changed parameter values at unchanged or backdated times, corrupting observation chronology and derived variable lifetimes.

Both subclass and legacy proxy samples now capture observation-start time for sample and selective-sample, preserving mutation bookkeeping, non-sampling event times, registry eligibility, suppression, and exceptions. Repeating inference/checking also exposed a separate coverage-observer bug: valid NumPy integer lifetime endpoints were rejected. A separate signed commit adds an exact, finite numeric conversion at that dataframe boundary; generic JSON validators and checker predicates remain unchanged.

Measured correctness improvement: on a fresh matched 2×A16, four-update Transformer trial, invalid after-update times fall from 216/216 to 0/216; equal before/after times despite changed hashes fall from 54/216 to 0/216. API/VAR record counts remain 7,504/594 in both versions, with no duplicate API event keys. Across six fixed subclass trials, 1,080/1,080 parameter/update pairs have complete samples and correct chronology.

All nine GPU phases complete successfully. All eight full checkpoint comparisons pass, covering 274 tensor pairs per comparison and matching whole-file hashes, including FP32/BF16, both resume directions, and the deliberately faulty normalization run. This is a correctness change, not a performance claim.

Validation:

• Parameter regressions: old source 8 failed / 13 passed, candidate 21 passed; real SGD disk-trace regression also fails before and passes after.
• Observer regressions: two real-pandas cases fail before; exact nanoseconds above 2**53, observer-on/off parity, missing/malformed values, booleans, nonfinite values, and datetime/duration rejection are covered.
• Composed local suite: 142 passed. Ring18 suites: 87 passed for tracing/model/checkpoint behavior and 55 passed for coverage.
• Pinned hooks pass. Configured mypy passes on 79 core files; strict scoped checks cover both new test modules, the changed tracing test, and the reusable coverage tool. Explicit test/tool lint/format checks pass. Legacy untyped bodies remain outside full checker coverage.

Fresh API+VAR checking finishes with 596/605 active rules, zero dormant/error/unknown outcomes. Healthy/fault violations are 9/8 on rank 0 and 15/17 on rank 1. Two rank-1 fault-only metadata-distinctness alarms do not establish normalization detection. The raw-data diagnosis shows the applied scale was not captured, numeric loss data was excluded, and existing rules do not encode global-token normalization. Healthy object-ID-based false alarms remain visible. Frozen inputs match; every previously completed per-rule report is exactly unchanged after the observer fix.

Full experiment report · Normalization coverage diagnosis

Limits: observation times do not imply exact mutation times or atomic snapshots; an explicit after-update sample remains outside the optimizer API interval. No FSDP, stochastic/dropout, changed topology, partial accumulation windows, or agent-benefit claim. Large evidence stays outside Git: the local archive has 185 hash-verified files; both source snapshots match their committed Python files exactly. Final audit finds no remaining study processes and idle GPUs.

Dependency: stacked on #166 (codex/ring18-distributed-probe), itself stacked on #164. All four commits are signed. No merge performed.