Retained GEAK overlays are omitted from concurrency sweep and exported launch script #1435

A retained GEAK Python overlay can disappear after promotion. Reproduced on Hyperloom f89eedaeff3a4499d7f068394c05aff633bba841 with native CPU controls; correction is published in existing PR #1428 at 9686b45e75a00635edce040854ecf78a9faf6092, awaiting review and merge.

Two paths omit current_best.final_overlay:

• orchestrator/kernel/conc_sweep.py::_has_optimization checks only arguments and environment. An overlay-only current best is skipped as no_optimization_to_compare. With flags also present, _build_arm_grid creates an optimized variant without overlay_pythonpath, so it launches the stock module while labelling the arm optimized.
• The exported current_setting.sh is rendered from flags and environment without the retained overlay. Replaying that script therefore does not reconstruct the accepted Python import path.

The sweep reproduction runs the actual coordinator/config helpers and a real child Python interpreter with a small sitecustomize.py sentinel. All four pre-fix cases fail: overlay-only and overlay+flags, each with persistent-server and per-variant launch control. The correction carries the overlay only into the optimized arm and refuses to benchmark if the retained overlay is missing, inert or cannot be represented by the grid's existing path contract. 177 focused tests pass, two skip, including the four actual-child controls and four missing/inert/path rejection controls. The export correction is also published: 85 focused export tests pass, including actual Bash/Python child controls. The final combined source passes 20,743 tests, 43 skips, eight expected failures and 15 subtests, plus whole-repository Ruff, formatting, REUSE and whitespace checks.

These CPU controls use explicit test-only throughput fixtures. They establish lost deployment inputs, not a GPU speedup or historical incidence. Fresh Kimi GPU A/B is running separately; no positive GEAK adoption or production dashboard recovery is claimed.

There is also a distinct source-artifact limitation: a final_patch/final_launch_script reference survives in geak_result and product reporting, but no source deployment layer is added to current_best/build_env_spec. A source-only result selects the GEAK final-launcher fallback; a mixed source+flags result instead selects the ordinary flags/env grid. A path string does not prove that a patch was applied or that its bytes survive a fresh launch. The overlay correction does not claim to solve arbitrary source-patch replay.

Evidence: /home/sapmajum/work/geak_ci_delivery_20260907/composed_validation/geak_source_retention_audit/receipt.json; independent source review /home/sapmajum/work/geak_ci_delivery_20260907/native_source_retention_independent_review/review.json (SHA256 7c2218309dde5058bf84f2eac5c14838f9cc18a46e33764331c02004b60e1f54). All original failures and output files remain retained.

Final combined independent review: two actual-child controls verify that native rejection preserves the previous overlay through export and sweep, while fresh acceptance retains the candidate and fresh grading axes. Baseline remains stock. Receipt SHA256 2f5fce1e23441d4fa75ab3d827024b28cdb4a799780da7653b7792680a414463.
At the 06:11 UTC GitHub read, the new head's unit tests on Python 3.10/3.11, coverage, lint, packaging, documentation and security checks pass. The single-GPU CI smoke and its aggregate status remain pending.