[acoliver]

Recreated after the llxprt account suspension removed original PR #292. Head 0b7ee85593d0997062bae056742e0aea2f038a0e on restored branch branch4-issue220-edit-eval, based on the shared #252 repair branch.

Contents:

• Evaluator scripts, findings documentation, examples and tests for the shell-enabled ordinary/recovery x current/explicit edit-guidance investigation.
• Four faithful successful trials with preserved bytes outside the target function; 33 historical tests passed; both guidance arms succeeded; four defective runtime controls failed the intended witness. Findings are explicitly bounded: one fixed-order sample per cell does not establish a causal guidance benefit, and earlier shell-disabled no-edit failures remain failures.
• Integrated main and the pinned Parity verification can recursively run the enclosing workspace and leave orphan test trees #252 shared repair; ToolConfig and PID-readiness fixes included without weaker assertions.
• Exact-head verification: 13 Python evaluator runs, fmt (workspace+xtask), full workspace all-target/all-feature clippy/tests/build, xtask tests, all numeric exits 0; 792 library tests plus integration/examples.

Remaining before ready: final shared repair integration (bc839f2/29905a23), invalidated-gate verification, and no redundant initial feature review. Do not merge.

[acoliver]

The feature-only current-main replacement is now #320, exact dbd7b6c03a5e46f6497ffbfef089a9d8738dbfd4, stacked on updated #308. All13 unique evaluation/docs/example/test paths and original source/model/outcome qualifications are mapped and retained; no historical experiment or redundant initial review was repeated. Original branch/head remains intact.

Closing this old conflicted delivery in favor of #320 avoids force rewriting its evidence history or importing inherited stale repairs. #320 remains draft while final package/native acceptance, evidence-delivery limits and fresh exact-head CI complete. Closure is consolidation, not merge approval.