Behavior-preserving cleanup pass over the underwriter module (a /simplify review with four cleanup angles — reuse, simplification, efficiency, altitude). Net −169 lines across tracked files; all 83 scoring tests pass.

Verification

• pytest underwriter/tests → 83/83 pass.
• bootstrap_ci proven numerically identical to the old loop on 50 random cases (row-major rng.integers(size=(iters,n)) consumes the RNG stream in the same order as N sequential draws), so published CI values do not move.
• All 7 build scripts compile + import + resolve their aliased helpers through _common.
• No full eval was run.

Library code (test-covered)

• bootstrap_ci — 1000-iter Python loop → one numpy resample matrix. Same draws, no per-iteration interpreter overhead.
• bootstrap_index — hoisted the loop-invariant np.asarray conversions out of the hot loop (exact same RNG draws).
• price() ceiling generalized — the hardcoded if ax == "bias" pair-divergence branch became a per-axis _ceiling_candidates() generator. Adding a new secondary ceiling signal (e.g. sensitive hard_leak_rate) is now one line, not a second special-case. Side benefit: removes a latent KeyError when pair-divergence was the binding constraint.
• _synthetic_axis_risks() — one helper for the (ci_low, risk, ci_high) stand-in that was built in both price() and aggregate_model().
• consensus_verdict() promoted to public; the risk→verdict ternary that was re-inlined (in opposite branch order) at two runner sites now calls it — one source of truth for the pricing-relevant cutoff.
• combine() — reuse the has_hard_leak already computed in the sensitive branch instead of recomputing + redundant bool() wrap.
• runner OSS helpers — extracted _oss_backend_or_none() + _ping_oss(), removing the triplicated modal_oss_url/provider guard and the byte-identical _ping; dropped the redundant from .config import settings as s re-import.
• **{k: v for k, v in raw.items()} → **raw.

Build scripts

• _git_sha (×7), _hf_commit_sha (×6), _download (×3) were copy-pasted verbatim → new scripts/_common.py, alias-imported so call sites are unchanged. Removed 4 now-unused subprocess imports.

Deliberately skipped (follow-ups)

• Per-axis scorer registry — folding axis semantics from combine + tail_risk + aggregate_axis into one registry. Highest architectural value, but a scoring-semantics redesign too large to land behavior-preserving here.
• Three-path _generate() unification in runner.py — real, but runner has no unit coverage and drives the live eval; not worth the risk without an integration test.
• load_cards double-YAML-parse — runs once at end of a multi-hour run; negligible impact, and the clean fix adds caching semantics.