research: measured Nession #1258 before/after handler locality experiment

[BestNathan]

Refs BestNathan/nession#1258.

Completed measurements

• Frozen Nession baseline 6b7dcff69275b1e69ad6f8484240a41436db4cc5 and accepted staging treatment 696892d0cb003a507dde5f9b7087014db8f92e18.
• Reproducible deterministic source retrieval study: four known versioned Unit identities plus runtime-owned relay.end negative control; raw checksums archived in workflow #38051474658 and observed results in docs/topics/agent-native-repository-architecture/research/experiments/nession-server-handler-locality-2026-10/RESULTS.md.
• Four versioned operations: selected implementation source bytes 1,313,540 → 29,416, conservative file closure 1,313,540 → 386,714; localized tests found for 2/4 tasks. All five owners correctly identified by known canonical identity.
• Cross-unit concurrent worktree merge from Nession Quality Gate also recorded.
• Existing frozen benchmark integrity gate repaired by explicitly fetching the three immutable research commits B/C/oracle from Nession before validating; no integrity checks disabled.

Methodological limits

This was not an AI coding-agent run. No token, context window, model calls, elapsed navigation, repair-loop, or software productivity effects have been measured. The result supports narrower known-Unit source retrieval only; findings remain bounded and explicitly document the relay.end exception. Further controlled agent A/B tasks are needed for any AI-productivity claim.

CI

ci, AI Native Benchmark Integrity, and Nession Handler Locality Study all successful on head 1737b60cae22815777f2d01031e3a005e37a9bc1.