applied: the block-transposition adjacency property is unpinned in release

[tsondru]

Deleting adjacent_runs's b1 == a2 check (smc_nf.rs:2145) reddens 5 shipped tests in debug and 0 in release; all five report the same debug_assert! at smc_nf.rs:2041, so the block-transposition adjacency property has no release-profile guard.

A release lane does not close it: the same perturbation under --release left every shipped applied test green (449 passed; only the review's own oracle red). Closing it needs a test asserting the rewritten diagram, not the debug invariant.

Measured in G5-T1's falsifying review of 9181c1a.

[tsondru]

Re-scoped 2026-08-31 (owner call at the G5-T6 STOP): the filed premise no longer holds at main 27be055 — the identical adjacent_runs perturbation now reddens 6 tests in debug and 4 in release (Gate.churn 223→219 / 1032→1027, Score.divergent 15→14, Counts.divergent 15→14; measured in the G5-T6 probe, /tmp/audit-reports/G5-T6/close-out.md). Release coverage exists but is incidental corpus-statistic coverage. The task proceeds as: add the narrow, targeted adjacency-contract oracle anyway, with the measured 6/4 baseline as the premise.

[tsondru]

PR #389 (squash on main). Prod diff: 0 (+92 tests/smc_nf_regression.rs, +7 CHANGELOG, ±1 SMC-NF-RECONCILIATION.md). Narrow adjacency-contract oracle, red in both profiles under the b1 == a2 deletion via its own assert_eq in release; falsification table in the PR body. Null recorded there: a first-shared-layer-only decision is not distinguished by this fixture.