"""Reproducible finite checks and reduced-model experiments, v2.

No result from this script is molecular or laboratory evidence. A conventional
joint sensing/control design is included and intentionally ties the same policy.
"""
from pathlib import Path
import argparse, json, csv, sys, platform
from dataclasses import replace
import numpy as np
from scipy.linalg import eigh
from contracts import (laplacian,kron,relative_spectrum,assemble_responses,
                       conductance,energy_weights,weighted_response_bounds)
from contract_genome import capsule,compile_grid
from contract_growth import Config,manufacture
from recovery import recoverability,posterior,joint_value

ROOT=Path(__file__).resolve().parents[1]


def theorem_checks():
    rng=np.random.default_rng(845117)
    min_slack=1.; comp_max=0.; trials=[]
    for run in range(300):
        # Overlapping four-port modules. Shared nodes are the only interactions.
        count=2+run%18; nodes=2*count+2
        local=[]; nominal=[]; maps=[]; deviations=[]
        for j in range(count):
            # Actual module has one interior node. No target stores its coordinate.
            edge=np.array([[a,b] for a in range(5) for b in range(a+1,5)])
            w=np.exp(rng.normal(0,.8,len(edge)))
            # Shared modulation makes module errors strongly correlated.
            shared=.055*np.sin(run)
            ratio=1+shared+rng.uniform(-.04,.04,len(edge))
            D0=kron(laplacian(5,edge,w),[0,1,2,3])
            D=kron(laplacian(5,edge,w*ratio),[0,1,2,3])
            nominal.append(D0); local.append(D); maps.append(np.arange(2*j,2*j+4))
            deviations.append(max(abs(relative_spectrum(D,D0)-1)))
        target=assemble_responses(nominal,maps,nodes,[0,1,nodes-2,nodes-1])
        actual=assemble_responses(local,maps,nodes,[0,1,nodes-2,nodes-1])
        deviation=max(abs(relative_spectrum(actual,target)-1))
        eta=max(deviations); min_slack=min(min_slack,eta-deviation)
        comp_max=max(comp_max,deviation)
        assert deviation<=eta+2e-12
        trials.append({'modules':count,'largest_local_error':eta,'global_error':deviation})
    # Single-probe blindness: deleting the edge between equal-voltage ports.
    D0=laplacian(3,[[0,1],[0,2],[1,2]],[1,1,1]).toarray()
    D1=laplacian(3,[[0,1],[0,2],[1,2]],[1,1,0]).toarray()
    u=np.array([1.,0.,0.]); v=np.array([0.,1.,0.])
    blind={'first_probe_current_difference':float(np.linalg.norm((D1-D0)@u)),
           'second_probe_current_difference':float(np.linalg.norm((D1-D0)@v)),
           'relative_spectrum':relative_spectrum(D1,D0).tolist()}
    # General noncommuting quadratic evidence-control complementarity.
    errs=[]; synerg=[]
    for _ in range(400):
        A=rng.normal(size=(5,5)); S=A@A.T+.2*np.eye(5)
        O=rng.normal(size=(3,5)); noise=.4*np.eye(3)
        Sp=posterior(S,O,noise); Om=S-Sp
        B=rng.normal(size=(5,3)); U=rng.normal(size=(5,2))
        B2=np.c_[B,U]
        Aobs=rng.normal(size=(5,5)); dOm=Aobs@Aobs.T
        def val(b,o):return recoverability(S,o,b)['dividend']
        synergy=val(B2,Om+dOm)-val(B2,Om)-val(B,Om+dOm)+val(B,Om)
        synerg.append(synergy)
        mu=rng.normal(size=5)
        ctrl=np.linalg.solve(B.T@B+np.eye(3),B.T@mu)
        direct=.5*np.linalg.norm(mu-B@ctrl)**2+.5*np.dot(ctrl,ctrl)
        W=B@B.T
        formula=.5*mu@np.linalg.solve(W+np.eye(5),mu)
        errs.append(abs(direct-formula))
    assert min(synerg)>-1e-12
    # Weighted deterministic response bounds with independent and adversarial damage.
    net=compile_grid(capsule(9,2)); N=len(net['parent']); w=energy_weights(N,net['edges'],net['target'],net['left'],net['right'])
    G0=conductance(N,net['edges'],net['target'],net['left'],net['right'])
    worst_slack=1.
    weighted=[]
    for _ in range(250):
        ratio=np.exp(rng.normal(-.2,.4,len(w)))
        lo,hi=weighted_response_bounds(w,ratio)
        true=conductance(N,net['edges'],net['target']*ratio,net['left'],net['right'])/G0
        assert lo-1e-12<=true<=hi+1e-12
        worst_slack=min(worst_slack,true-lo,hi-true)
        weighted.append([lo,true,hi])
    return {'composition_trials':300,'composition_min_slack':min_slack,
            'composition_largest_global_error':comp_max,'composition_examples':trials,
            'blind_probe_counterexample':blind,'quadratic_trials':400,
            'quadratic_max_absolute_error':max(errs),'minimum_complementarity':min(synerg),
            'weighted_trials':250,'weighted_min_slack':worst_slack,
            'power_participation':float(1/np.dot(w,w)),'weighted_examples':weighted}


def run(reps):
    (ROOT/'results').mkdir(exist_ok=True)
    methods=['open_loop','blueprint','blind_reserve','sensor_only','early_seal',
             'contract','conventional_joint','biased_reference']
    allrows=[]; summaries=[]
    for dim,n in [(2,8),(3,5)]:
        for method in methods:
            rows=[]
            for k in range(reps):
                cfg=Config(dim=dim,n=n,seed=20260919+1000*dim+k)
                row,snapshot=manufacture(cfg,method)
                rows.append(row); allrows.append(row)
                if k==0 and method in ('contract','open_loop','early_seal'):
                    np.savez_compressed(ROOT/'results'/f'{dim}d_{method}.npz',**snapshot)
            summary={'dimension':dim,'method':method,'replicates':reps,
                'complete_count':sum(r['complete'] for r in rows),
                'functional_yield_count':sum(r['completed_function_ok'] for r in rows),
                'all_contract_count':sum(r['all_local_contracts'] for r in rows),
                'false_certificate_count':sum(r['false_certificate'] for r in rows)}
            for field in ['functional_ratio','total_time','additional_material','nominal_material',
                          'inspection_samples','maturation_rounds','blocked_unaccepted_modules',
                          'local_contract_fraction','measurement_radius']:
                vals=[r[field] for r in rows]
                summary[field+'_mean']=float(np.mean(vals))
                summary[field+'_sd']=float(np.std(vals,ddof=1)) if len(vals)>1 else 0.
            summaries.append(summary)
            print(dim,method,summary['functional_yield_count'],reps,round(summary['functional_ratio_mean'],5),flush=True)
    result={'replicates':reps,'total_runs':len(allrows),'summaries':summaries,
            'methods':methods,'measure': 'completed object within 8% of target two-terminal conductance',
            'python':platform.python_version(),'numpy':np.__version__,
            'scope':'finite reduced module simulation; bounded ideal actuator and calibrated terminal reference'}
    (ROOT/'results/runs.json').write_text(json.dumps(allrows,indent=2))
    (ROOT/'results/summary.json').write_text(json.dumps(result,indent=2))
    with (ROOT/'results/summary.csv').open('w') as f:
        writer=csv.DictWriter(f,fieldnames=summaries[0].keys()); writer.writeheader(); writer.writerows(summaries)
    checks=theorem_checks()
    (ROOT/'results/theorem_checks.json').write_text(json.dumps(checks,indent=2))
    # Closure drift deliberately violates the claimed 1% envelope: negative control.
    stressed=[]
    for k in range(reps):
        cfg=Config(seed=707400+k,seal_actual_bound=.18)
        r,_=manufacture(cfg,'contract'); stressed.append(r)
    (ROOT/'results/seal_stress.json').write_text(json.dumps(stressed,indent=2))
    print('theorem_checks:',{k:v for k,v in checks.items() if not isinstance(v,list)},flush=True)
    print('seal_stress',sum(r['false_certificate'] for r in stressed),'false certificates;',
          sum(r['completed_function_ok'] for r in stressed),'completed function passes',flush=True)


if __name__=='__main__':
    p=argparse.ArgumentParser();p.add_argument('--reps',type=int,default=32)
    run(p.parse_args().reps)