Epic #260 brought the MLX backend to FITTING parity (sharing, Newton, all pdf families), but the non-fitting surface is still absent and raises NotImplementedError or simply does not exist: keep_best (the #51 best-iterate safeguard — late Newton overshoots were its original motivation), save/load persistence (pdtype/n_kurt_done state is currently unpersistable), outlier rejection (do_reject), transform/source extraction, and the LLt/MIR diagnostic outputs.
Filed so the gap is tracked before #157/#161 (LLt/MIR work) are scoped — both currently hard-code a two-backend assumption because MLX lacks the substrate. Scoping decision (which pieces, one epic vs individual issues) is deliberately left open for the lead/user.
Found during the epic #260 backlog triage. Related: #260, #157, #161, #51.
Converted to an epic per the 2026-08-23 decision (user-facing first). Phases: 1 transform + save/load (#287), 2 keep_best (#288), 3 rejection + LLt/MIR (#289). Runs via the epic-dev workflow on user go-ahead; phase 3 depends on the #157 LLt decision landing in torch/numpy first.
Epic #260 brought the MLX backend to FITTING parity (sharing, Newton, all pdf families), but the non-fitting surface is still absent and raises NotImplementedError or simply does not exist: keep_best (the #51 best-iterate safeguard — late Newton overshoots were its original motivation), save/load persistence (pdtype/n_kurt_done state is currently unpersistable), outlier rejection (do_reject), transform/source extraction, and the LLt/MIR diagnostic outputs.
Filed so the gap is tracked before #157/#161 (LLt/MIR work) are scoped — both currently hard-code a two-backend assumption because MLX lacks the substrate. Scoping decision (which pieces, one epic vs individual issues) is deliberately left open for the lead/user.
Found during the epic #260 backlog triage. Related: #260, #157, #161, #51.
Converted to an epic per the 2026-08-23 decision (user-facing first). Phases: 1 transform + save/load (#287), 2 keep_best (#288), 3 rejection + LLt/MIR (#289). Runs via the epic-dev workflow on user go-ahead; phase 3 depends on the #157 LLt decision landing in torch/numpy first.