[Maya-Kid]

Equal RRF scores inherited the randomized iteration order of a HashMap. With disjoint lexical/vector hits, repeating the same query could change both the evidence order and which candidate survives top_k; the shared helper also feeds concept mapping selection.

Break equal scores by ascending ID after descending score. Regressions cover ties across every cutoff (including zero), reversed channel order, and equal totals formed from different rank contributions. Both new tests failed before the fix.

Validation on this branch:

• cargo +1.98.1 fmt --all --check
• cargo +1.98.1 clippy --locked --workspace --all-targets -- -D warnings
• cargo +1.98.1 test --locked --workspace: 1,411 passed, with dedicated PostgreSQL 16/pgvector and both database/PDF requirements enabled; 6 existing opt-in tests remain ignored
• cargo +1.98.1 build --locked --workspace
• Unchanged frontend at the same base: frozen pnpm 10.2.1 install, 228 tests, and production build passed

[WaylandYang]

Thanks @Maya-Kid. Equal RRF scores took the order of a freshly seeded HashMap, so the same query could keep different candidates at the cutoff. Breaking the tie by id after the score is the right fix, and the test at every limit pins it. Landing it.