FINAL4 loop-index smoke regression

The source strings in rs/tests/loop_indices.rs were checked byte-for-byte against the existing FINAL4 installed-smoke sources. The historical archive and smoke directory were not modified.

• release_helper.tri: c5d2e49afbfc5896403eba9fcca4ea84c076678a6c543111d7e44022d301db67
• imported-nox.tri: dfbade31bff4ef28259c974f43213567c04a33ee3a58377c0e4e69670cfba0e1

cargo test --release --locked -p joy-rs --test loop_indices -- --test-threads=1 passed 2 tests, zero failed/ignored/Rust warnings, in 15.68 seconds. Log: /tmp/joy-final4-loop-indices-proofs.log.

The exact fixture produced [803] from [3,5] and [809] from [4,5] in both source profiles. All four cases generated and verified public execution certificates and genuine Triton ZK execution proofs, including serialization roundtrips. Changes to either public input or public output were rejected. The fixture has no secret input; the ZK backend hides its execution witness.

The independent scope regression also generated four public certificates for nested loops with the same index name, an absorbed conditional local shadow, and early returns: flag0 returns7, flag1 returns109 in both profiles. A nearer runtime index binding correctly fails compilation rather than inheriting a loop/module constant.

Read-only inspection found no additional defect in the changed Scope logic: bind clears constant metadata, lookup_constant stops at the nearest binding including None, loop frames restore outer bindings, and sealed absorbed frames clear names, constant metadata, and types together. Both ordinary-transform and returning-loop paths use bind_loop_index; indices outside U32 are refused. This establishes the bounded/unrolled index fix, not support for arbitrary runtime-index array access on nox.