[initial-d]

Existing corpus-window experiments did not establish a need for distinct learned memory rows. Add source-pinned official MEMIT MultiCounterFact editing experiments on Qwen3-0.6B and Qwen3.5-4B, with original decoder replay, masked-gradient checks, three hash capacities, independently trained controls, and calibration-selected inverted-index BM25.

The completed 1,024-edit grid includes 36 trained models and four base/retrieval evaluations across three seeds. Archived candidate scores do not show a stable large-table gain; BM25 has the highest observed quality. Four separately registered frozen-backbone probes retain the locality/generalization tradeoff. Both actual trained native bundles independently reopen with checked next-token logits byte-exact, but both table and complete bundle bytes expand.

The full 20,877-edit grid remains in progress. A final cache log encountered Transformers set-valued loading information after saving the artifacts; preserve the nonzero exits and add explicit hash/coverage-validated resume instead of recomputing completed work. Model, data, training schedule and scoring protocol remain fixed. This PR does not claim editing-method SOTA, large-table necessity, native serving superiority, or generation-metric coverage. The precision audit discloses the installed FLA solver's internal TF32 path.

Validation: 118 local tests passed, with 124 dependency skips (242 discovered), including corruption/coverage/protocol resume checks; real-weight preflights on both models; CPU/CUDA replay fixtures; dependency-free recomputation of all 40 main/4 exploratory scalar reports and RAG calibration selection; optional NumPy reproduction of all 12 paired bootstrap intervals; complete exported-row and 16 actual question-prefix checks per model, including independent native reopen. Exact commands, source snapshots, hashes, costs and failure logs are archived. CI verifies the completed scalar evidence.