v3 paper: skeleton, risks write-up, artifacts

[nikolay-e]

Skeleton (abstract leads with the nontrivial number, not 0.919 — v2 proved why):

1. Intro — the mixer is under-calibrated (v2), RRF rejected (repo data), "why does fusion break" → calibration thesis.
2. Related work — + agentic line (RepoGraph / LocAgent / CodexGraph) as context for locate mode; SPLADE-Code for the lexical channel; HKPR/diffusions; DPP in retrieval.
3. Why naive fusion breaks — mode scales formally, RRF/PIT empirics from the corpus, incommensurability argument. The negative-results section is the paper's distinguishing part.
4. Method — 4.1 calibration + fused objective; 4.2 HKPR as generalisation (EGO/PPR limits, hk-relax, determinism); 4.3 DPP-MAP and the partition-matroid link; 4.4 sequential + discovery-union.
5. Protocol — dcbench (annotation protocol already in repo), splits, pre-registration, reuse disclosure, v5 config + v4 bridge.
6. Results — nontrivial first; per-role table; HKPR t-curve; precision/recall front (EGO, PPR, fused, fused+DPP); coverage decomposition; latency.
7. Limitations — knapsack-DPP without guarantee; one calibration manifest across 15+ languages; dcbench repo diversity; sequential's double cost.
8. Appendix — symbol-to-code map (as v2), ablations, grids, Appendix E code-drift record.

Risks with their write-up: HKPR flat → structural result + curve; calibration not transferring → rank-only fallback; DPP drops recall → front not point; "third run on the same data" → fresh holdout + disclosure; code drift → tag paper-v3 + Q-freeze.

Trilogy arc: v1 problem + PPR; v2 retention/retrieval decomposition + mixer diagnosis; v3 failure mechanism of naive fusion and its cure.