Fix knock-in pricing in the cost filter, and restore negative intervention costs

[VonAlphaBisZulu]

Three defects on the same path. Each is invisible alone, because a knockout-only problem never touches any of them. The first causes silently wrong results with ordinary positive costs and is independent of the other two.

1. filter_sd_maxcost charged a knock-in its knockout cost

networktools.filter_sd_maxcost priced an intervention by dict membership, kocost first:

(kocost[k] if k in kocost else kicost.get(k, 0)) if v != 0 else 0

SDProblem prices it the other way round — knock-in wins over knockout (strainDesignProblem.py:147,149-151, where ko_cost[i] is nan-masked wherever ki_cost[i] is set). A reaction present in both dicts is therefore re-priced after the fact and the design is discarded.

The overlap is the normal case, not an edge case: compute_strain_designs.py:476-477 defaults uncmp_ko_cost to every reaction id whenever the caller passes only ki_cost.

This is not specific to negative costs. On a four-reaction network with elementary flux vectors {R1,R2,R4} and {R1,R3,R4}, using ko_cost={'R2': 50}, ki_cost={all: 1}, max_cost=3, all positive:

	designs
before	1 — [['R1','R3','R4']]
after	2 — both

The MILP finds {R1,R2,R4} at cost 3; the filter re-prices it at 52 and drops it. Gene-based problems escape this because uncmp_ko_cost starts empty there, which is why no existing test catches it.

Fixed by dispatching on the sign of the design value (+1 knock-in, −1 knockout), the value semantics the function already documents. Where ki_cost is empty the new expression is identical to the old term for term, so knockout behaviour cannot change by construction.

The same expression also zipped the filtered designs against the unfiltered cost list when stamping '**cost**', misaligning the labels and the resulting sort whenever the filter removed anything. Fixed alongside, since correct stamping is required once anything is filtered.

2. cont_MILP made verify_sd reject everything at a negative budget

cont_MILP keeps the three leading rows that constrain z alone. Stripped of their z columns they read 0 <= max_cost — vacuous for a non-negative budget, but false for a negative one, so every design failed verification. Their right-hand sides are now neutralised in the continuous copy. The rows are kept rather than dropped, because z_map_constr_ineq columns are aligned to row positions.

3. Row 0 forbade a negative total cost

Row 0 asserted total cost >= 0. That is implied whenever costs are non-negative, and z is binary so the objective is bounded without it. Its only effect was to make negative costs infeasible by construction. Left open; the row itself is kept so row indices stay put.

What this restores

Negative intervention costs as a way of forcing a reaction into every solution — give it a large negative cost and set max_cost so that omitting it is unaffordable, while the budget still bounds the rest:

cost(R2)	max_cost	result
1	3	both EFVs
−100	−97	{R1,R2,R4} only — R2 forced, ≤3 other units
−100	−98	{R1,R2,R4} only
−100	0	{R1,R2,R4} only, unbounded rest

Validation

• e_coli_core MCS (growth >= 0.001, max_cost 3): identical 353 designs, sizes {1:19, 2:111, 3:223}, compared as sets of frozensets — 0 missing, 0 extra
• Test suite: 368 passed, 2 skipped (performance tests deselected)
• The positive-cost regression above verified against unmodified main

Not addressed, flagged deliberately

• The deeper root cause of the overlap is compute_strain_designs.py:476-477 defaulting ko_cost to all reactions. Fixing it there would shrink the MILP's knockout candidate set and change ordinary behaviour, so it is left alone.
• strainDesignSolutions.py:219-243 resolves the same overlap by summing both dicts, so a reaction in both is double-counted in reported costs. Reporting only; does not affect design sets.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RFtof9nZXFvCNoXz19po8C