Bill knock-ins once, and stop them doubling as knockout candidates

[VonAlphaBisZulu]

Follow-up to #79. Two remaining consequences of a reaction sitting in both cost dictionaries at once. Neither can arise for a knockout-only problem, which is why the existing tests never reached them.

1. Interventions were billed once per dictionary

SDSolutions._compute_costs_and_bounds walked each cost dictionary in turn and added whatever it found, so a reaction present in both was charged twice:

	reported	true
{R1,R2,R4}, all ki_cost 1	6.0	3

It now selects the dictionary by the sign of the entry — positive for an addition, negative for a removal — matching how SDProblem prices interventions and how filter_sd_maxcost selects a cost.

Design sets are untouched; only reported costs change, and only where the dictionaries overlapped.

2. Knock-in candidates were also offered as knockouts

compute_strain_designs defaulted ko_cost to every reaction in the model whenever the caller gave no knockout costs — including reactions the caller had just named as knock-in candidates. Knock-ins override knockouts in the MILP (SDProblem masks ko_cost wherever ki_cost is set), so those entries described a state that could never be solved.

The visible consequence: a SUPPRESS problem with every reaction named addable still answered with knockouts nobody asked for.

knockouts ['R4']  ·  knockouts ['R1']

The default now skips the named knock-ins. Reactions the caller did not name stay knockable, so mixed problems behave exactly as before — tests/test_05_straindesign.py::test_mcs passes ki_cost={'R2': 1} and expects knockout designs on the other reactions, and that still holds.

For the record, the stronger variant — emptying ko_cost entirely when knock-ins are present — was tried and broke 25 parametrisations of that test, so it is deliberately not what this does.

Regression tests

Four new tests in tests/test_05_straindesign.py, parametrised over every installed solver:

• test_efv_enumeration_via_knockins — a PROTECT module with all reactions addable enumerates minimal supports, i.e. elementary flux vectors
• test_negative_cost_forces_reaction_into_every_design — a large negative cost forces a reaction in while the budget still bounds the rest
• test_intervention_costs_are_not_counted_twice
• test_knockin_candidates_are_not_also_knockout_candidates

12 of the 16 fail without the changes in this PR. The 4 that pass either way guard the fixes already merged in #79.

Validation

• e_coli_core MCS (growth >= 0.001, max_cost 3): identical 353 designs, compared as sets
• Test suite: 384 passed, 2 skipped
• Heavy design gates (--medium --large, incl. test_ecoli_core_mcs_455 and test_iml1515_mcs_393) run locally — reported below once complete

🤖 Generated with Claude Code

https://claude.ai/code/session_01RFtof9nZXFvCNoXz19po8C

[VonAlphaBisZulu]

Heavy design gates run locally (these are behind --medium/--large and skipped by CI):

15 passed in 185.6s
  test_ecoli_core_mcs_455   [cplex, gurobi]   PASSED
  test_weak_mcs_wgcp        [cplex, gurobi]   PASSED
  test_weak_optknock / robustknock / optcouple [cplex, gurobi]  PASSED
  test_imlcore_mcs_ethanol  [cplex, gurobi]   PASSED
  test_imlcore_mcs_growth   [cplex, gurobi]   PASSED
  test_iml1515_mcs_393      [gurobi]          PASSED  (164.1s)

Both the 455 and 393 design gates hold on both solvers.