lyndonkl/wc-building-block-crossover

wc-building-block-crossover — recombine modules, then repair

Implements the recombination step of evolution-protocol.md against the module map in building-blocks.md. The discipline that defines this skill, straight from the Evolution document: a squad is not a flat list of independent slots; it is modules with internal linkage, and crossover must operate at module boundaries or it destroys the very things that make the parents good. "Excellent planning + Excellent memory without rediscovering both independently" — here: the best captain core and the best clean-sheet stack, fused.

Inputs

• The elite candidate set (from fitness selection), each with its block tags and per-block quality.
• Live constraints from tournament-state.md: budget (group $100m / KO $105m), nation cap (3 group, current-phase cap KO), phase.

Workflow

- [ ] 1. For each block slot, rank the elites by that block's quality (per-block fitness contribution)
- [ ] 2. Compose offspring by harvesting the top block from different parents (try a few combinations)
- [ ] 3. REPAIR each offspring to feasibility — cheapest block first, never gut a value block
- [ ] 4. Drop any offspring whose parents' blocks are fundamentally incompatible (repair would break a block)
- [ ] 5. Record block lineage per offspring; emit the feasible offspring

Step 1–2: Cross at block boundaries

For a squad, the blocks are BB1 Captain Core, BB2 Clean-Sheet Spine, BB3 Mid-Engine, BB4 Differential Pod, BB5 Enabler Bench (BB6 Fixture/Timing is the constraint layer, enforced in repair). For each block, identify which elite has the best version (highest contribution to fitness for that module — best captain core, best stack, sharpest pod). Compose 2–4 candidate offspring by mixing: e.g.

offspring_1 = BB1(A1) + BB2(A4) + BB3(A5) + BB4(A2) + BB5(cheapest-feasible)
offspring_2 = BB1(A6) + BB2(A3) + BB3(A5) + BB4(A4) + BB5(...)

Harvest whole blocks, not players. Do not pluck one defender out of a stack — the stack's value is its correlation; take it or leave it. For a matchday plan, cross MB2 captain-ladder from one parent, MB3 bench-order from another, MB1 XI / MB4 sub-triggers / MB5 chip likewise — keeping each block internally intact.

Step 3: Repair to feasibility (priority order)

Recombined blocks will usually break BB6 constraints. Repair in this order, preserving value-bearing blocks:

1. Formation. If no valid XI (1 GK; 3–5 DEF; 2–5 MID; 1–3 FWD) can be fielded from the 15, the recombination is structurally broken → re-pick the offending block from the next-best parent rather than patching individual players.
2. Nation cap. If a nation exceeds the cap, swap the lowest-xEV offender, preferring to move a BB5 enabler or BB4 differential over a BB1/BB2 value piece.
3. Budget. If over budget, release funds from BB5 first (downgrade the cheapest enabler), then BB4 (cheaper differential), and only touch BB1/BB2/BB3 if unavoidable — and if so, prefer downgrading one BB1 premium over breaking the BB2 stack (correlation is harder to rebuild than a single captain option).
4. Matchday spread / dead-minutes bench. Fix last; it's the cheapest lever.

Step 4: Drop incompatible offspring

Repair invariant: never satisfy a constraint by destroying a block. If, say, the only way to afford BB1(A1)'s captain core is to break BB2(A4)'s stack, the two parents are incompatible — drop that offspring and report it rather than emit a mangled squad. An incompatible pairing is information (those two genotypes don't fuse this round), not something to force.

Step 5: Lineage + emit

For each surviving offspring record:

offspring_id: <id>
blocks:
  BB1_captain_core:   { from: A1, players: [...] }
  BB2_clean_sheet:    { from: A4, players: [...] }
  BB3_mid_engine:     { from: A5, players: [...] }
  BB4_differential:   { from: A2, players: [...] }
  BB5_enabler_bench:  { from: repair, players: [...] }
repair_log: [ "downgraded enabler X→Y for $1.5m", "swapped nation-clash bench Z" ]
feasible: true

Lineage feeds the board ("A1 core + A4 stack + A2 pod") and the archetype scoreboard (which genotypes' blocks keep getting chosen). Hand the offspring to mutation, then diversity.

Guardrails

• Whole blocks only. No individual-player crossover — that's the failure mode this skill exists to prevent.
• Repair never guts a block. Cheapest-block-first; drop incompatible offspring rather than mangle.
• Always emit lineage + repair log. Unexplained offspring can't be presented or learned from.
• Produce a variance-diverse offspring set where the parents allow it (a safe fusion and a brave fusion) — the diversity skill checks this next, but don't pre-collapse it here.