Leg C — imperfections and faults on error seeds

The error budget

error_budget.element rows are the vocabulary of ErrorStudy.add_error:

{"element": [{"pattern": "QUAD_*", "parameter": "dx", "sigma": 0.10, "kind": "Quadrupole"},
             {"pattern": "FMAP_*", "parameter": "voltage_rel", "sigma": 0.01, "kind": "cavity"},
             {"pattern": "FMAP_*", "parameter": "phase_offset", "sigma": 1.0, "kind": "cavity"}],
 "beam":    [{"parameter": "centroid_x", "sigma": 0.1}, {"parameter": "emit_nx_rel", "sigma": 0.05}]}

(sigma for Gaussian at cutoff 3σ, half_width for uniform; kind gates a name glob by element type — mandatory on decks where solenoids and cavities share the FMAP_* prefix). The budget may also live in its own file next to the spec ("error_budget": "error_budget.json"); it is inlined when the spec is loaded, so the campaign folder is self-contained. At creation the campaign runs the engine's audit and refuses a budget with a selector that matches nothing, a warn-skipped or inert parameter, an unknown distribution, or overlapping selectors — the misalignment-study driver's policy, so a malformed budget never silently draws nothing.

Draws as overrides

Each seed is drawn once by the engine's own _apply_errors / _apply_beam_errors (ErrorStudy.draw_overrides) and transported as @index.attr overrides to a scan-pool run, so the seeds run in the process pool instead of the serial run loop; the draws are saved in legs/imperfections/draws/seed_NNNN.json and travel with an exported job. With correction.enabled the orbit is corrected on the errored copy and the steerer kicks join the overrides (the envelope reading backend by default, a per-seed BPM-noise stream). correction.pairing says how steerers and BPMs are paired: cards (the default) goes through the deck's ADJUST_STEERER cards (run_correction_from_lattice) — a deck without cards then corrects nothing and the seed row says none; auto takes every magnetic Steerer against every BPM marker (auto_paired_correction: one-to-one with the nearest downstream BPM when the counts match, otherwise a global SVD over all BPMs), the same apply_correction the card driver calls, so a deck whose cards pair each steerer with its next BPM gives the same kicks either way; without cards there is no per-steerer vmax clip. A correction pass costs tracking runs of the whole deck: the SVD measures one response column per steerer and plane (two runs per steerer plus the reading), the one-to-one method probes every pair on its own (five runs per pair), so on a long linac n_iter is the knob to watch. A draw that lands on an attribute the design element does not carry (an alignment offset on a marker) is reported as untransportable rather than dropped.

Faults on seeds

The top seeds.faults_on_seeds.top_n fault cases are replayed on the first n_seeds error seeds, before and after compensation (the compensator settings found on the design lattice), and judged by the criticality rule against that seed's own run; the robust fraction per case is the share of seeds on which the fault stays non-critical — the answer to "does the compensation found on paper survive the machine's imperfections".

Outputs

seeds.csv (per seed: transmission, exit energy, emittances, sizes, loss power, the correction's status, the orbit rms over every BPM before the correction and after its last pass with the number of passes, number of draws; the control row is the nominal lattice with the same beam), faults_on_seeds.csv, robustness.csv, and the figures seeds_transmission.png, seeds_emittance.png, faults_on_seeds_robustness.png.