[LittlePyx]

Training and comparison claims could cite nearby paragraphs that do not state the claimed relationship. Repeated merging of same-paper evidence could also attach an entire context bundle to one paragraph's locator, causing valid method passages to fail literal verification.

Preserve each supplied passage's original text and coordinates when merging; reject foreign or missing child locators. Require direct evidence for input/target maps, noise sampling versus addition, output correction, conditional variables, signal dependence and attributed limitations. Repeated citations to one document can expose a uniquely verified supplementary passage without adding redundant primary-passage links. Scores, numeric guards, citation budgets, TopK and source boundaries stay unchanged.

Draft: native and fresh acceptance have not passed; do not merge. Committed source42 is 35a75d61604673f90bc21556b6d0733e8125e69a. Local backend6782 passed/43 existing skips and all101 quality commands passed. Its PR CI succeeded, but push CI failed the original140ms first-visible deadline test at197ms. Retained API checks passed30 answers and70 literal/coordinate cards; native acceptance stopped after48 operations on a settings connection reset. Additional diagnostics recorded2 original5000ms locate failures and a5000ms message-load failure. All failures are retained in 42-final-hold-ledger.json; source42 generated no new model answers and main remains unchanged.

Candidate43 is commit 9b238eb43b41b1d4219d12615464632986a0c609. Full local backend6788 passed/43 existing skips and all101 unchanged quality commands passed; all76 saved-input answer texts match frozen42 exactly. Runtime/frontend/tests/fixtures are attested against this commit. Full retained native and same-model fresh acceptance remain pending. It precompiles the same alias patterns and uses a bounded512-entry immutable token snapshot, returning an independent set per caller and isolating stopword settings. It caches no source coordinates, support verdicts or quality results. Thirty retained answers' rendered/copy text and70 card quotes/coordinates match frozen42 exactly;250 real text inputs across3 stopword settings also match. Median offline rendering fell from1.81s to1.22s. Six regression cases and100 related checks pass;61 wrong claim/excerpt pairs remain rejected. The3 previously failing conversations completed5 card operations at the unchanged5000ms limit. Full checks passed; the deterministic diagnostics do not replace fresh or full native acceptance.

Frozen original24/new14 fixtures, corpora, negatives, historical failed flags and acceptance harnesses remain unchanged. Full76 same-model fresh generations, all-card/PDF/manual checks, full retained/fresh/saved native actions, cold-open12s checks and exact-head CI are still required. Model remains deepseek-v4-flash,1200 tokens/TopK6. BM25 remains default/full-library; local hybrid remains explicit opt-in within selected-paper/basket scope. No release, tag or authoring workflow. See docs/CURRENT_OPTIMIZATION_ROADMAP.md and docs/DIRECT_EVIDENCE_COMPARISON_20261019.md for the preserved evidence and failures.