v0.1.6

surge-dispatch release: SCUC LP tightening, large-network
performance, and a fix for unbounded thermal-slack relaxations. GO C3
adapter gains diagnostic knobs for isolating large-network bottlenecks.

Fixed

• Bounded thermal-slack relaxation in SCUC. Per-branch thermal
  slack columns (branch_lower_slack, branch_upper_slack) were
  allocated with col_upper = +∞, letting the LP relaxation
  hallucinate unbounded virtual capacity on degenerate networks.
  col_upper is now capped at 10× rating; slack rates are unchanged,
  so the economic tradeoff is preserved. On 1576-bus D1 s003 this
  takes the SCUC MIP from time_limit at 3637s with a −$1.7e14 dual
  bound to optimal at 66s with a 1.92% gap; commitment decisions
  unchanged.

Added

• Sparse reserve-product participation in SCUC. Reserve LP columns
  are now emitted only for (product, resource) pairs that can
  qualify under some commitment state AND have a nonzero offer
  capacity in some period. Applies to both generators and
  dispatchable loads. On 617-bus D2 the pre-presolve LP shrinks by
  ~97k columns.
• Consumer-level DL reserve aggregation. Dispatchable loads that
  share a reserve_group (the GO C3 pattern of price-decomposed
  consumer blocks) now share a single reserve variable per product,
  bounded by total offer and coupled to total served. Removes a
  spurious per-block pro-rata constraint that was over-restricting
  consumer reserve when block served-levels were uneven.
• Sparse reserve row families. Cross-headroom / cross-footroom,
  shared-limit, and energy-coupling rows are now emitted only for
  participating resources. On 73-bus D3 s303 pre-presolve rows drop
  37% and nonzeros 17%.
• SW0 branch-binary strip. When allow_branch_switching=false,
  branch_commitment/startup/shutdown columns and their
  state-evolution rows are omitted from the LP entirely instead of
  allocated and pinned. On 617-bus D2 this removes ~123k cols and
  ~82k rows up-front.
• GO C3 SCUC diagnostic knobs on GoC3Policy / DispatchRuntime:
  • scuc_disable_bus_power_balance — drop per-bus KCL rows and
    pb_* slack cols; replace with a single system-balance row plus
    a post-solve DC-PF theta repair before N-1 screening. Defaults to
    true for GO C3. On 6049-bus D1 s015 this takes SCUC from
    unsolved at 300s to optimal in 8s.
  • scuc_copperplate — zero the power-balance penalty so per-bus
    rows become trivially satisfied via free slack (for isolating
    whether MIP cost lives in UC or in network coupling).
  • scuc_firm_bus_balance_slacks, scuc_firm_branch_thermal_slacks,
    disable_scuc_thermal_limits — per-family slack-firming probes.

Performance (surge-dispatch)

• O(N²) hoists in attach_keyed_period_views. Branch / flowgate
  shadow-price lookups and zonal reserve participant matching no
  longer re-scan the network per period. On 4224-bus D1 s014 this
  function drops from 108s to sub-second.
• Hoisted network.bus_index_map() rebuilds out of the
  build_capacity_logic_reserve_rows zone loops. Per-product,
  per-period cost drops ~1000× on 4224-bus D1 s014.
• Cached zonal participant sets on ActiveZonalRequirement,
  eliminating an O(N_bus) HashMap rebuild per DL per zonal
  requirement per period in SCUC bounds construction.