/* x86_64 Linux only. Learned proposal, not a trainer or semantic validator.
 * Reuse reviewed bounded self-slot I/O. Legacy main/model are unreachable and
 * eliminated by --gc-sections; empty WEIGHTS_PATH satisfies its incbin only. */
#define main ln_unused_legacy_main
#include "nest.c"
#undef main
#if !defined(__linux__) || !defined(__x86_64__) || __BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__
#error "patch_nest requires little-endian x86_64 Linux"
#endif
#ifndef PATCH_WEIGHTS_PATH
#error "PATCH_WEIGHTS_PATH required"
#endif
__asm__(".pushsection .rodata\n.balign 4\n.global lp_w0\n.global lp_w1\nlp_w0:\n.incbin \"" STR(PATCH_WEIGHTS_PATH) "\"\nlp_w1:\n.popsection\n");
extern unsigned char lp_w0[], lp_w1[];
#define LP_COUNT 107994
/* Schema order is an ABI, mirrored by pack_patch.py. */
typedef struct { const float *ce,*c1w,*c1b,*c2w,*c2b,*cfw,*cfb,*be,*b1w,*b1b,*b2w,*b2b,*fw,*fb,*pw,*pb,*nw,*nb,*rw,*rb; float threshold; } PatchModel;
static const float *lp_take(const unsigned char **p, const unsigned char *end, size_t n) {
    if ((size_t)(end-*p) < n*4) die("truncated patch weights");
    const float *v=(const float *)*p;
    for(size_t i=0;i<n;i++) if(!isfinite(v[i])) die("nonfinite patch weights");
    *p+=n*4; return v;
}
static PatchModel lp_load(void) {
    const unsigned char *p=lp_w0,*end=lp_w1;
    if(end-p<32 || memcmp(p,"LNPATCH1",8)) die("missing patch weights"); p+=8;
    if(rd_u32(&p,end)!=1 || rd_u32(&p,end)!=96 || rd_u32(&p,end)!=256 || rd_u32(&p,end)!=LP_COUNT) die("unsupported patch weight schema");
    PatchModel m={0}; memcpy(&m.threshold,p,4);p+=4;
    if(!isfinite(m.threshold) || fabsf(m.threshold-.55f)>1e-6f || rd_u32(&p,end)!=0) die("unsupported confidence policy");
#define T(field,n) m.field=lp_take(&p,end,n)
    T(ce,257*24);T(c1w,48*24*5);T(c1b,48);T(c2w,48*48*5);T(c2b,48);
    T(cfw,64*288);T(cfb,64);T(be,257*24);T(b1w,64*24*9);T(b1b,64);
    T(b2w,64*64*9);T(b2b,64);T(fw,64*128);T(fb,64);T(pw,64);T(pb,1);
    T(nw,128);T(nb,1);T(rw,8*64);T(rb,8);
#undef T
    if(p!=end) die("unexpected patch weight bytes");return m;
}
/* Channel-major, PyTorch cross-correlation convention and exact GELU. */
static void lp_conv(const float *x,float *y,int length,int ni,int no,int kernel,const float *w,const float *bias) {
    for(int o=0;o<no;o++) for(int t=0;t<length;t++) {
        float s=bias[o];
        for(int i=0;i<ni;i++) for(int k=0;k<kernel;k++) {
            int at=t+k-kernel/2;
            if(at>=0 && at<length) s+=x[i*length+at]*w[(o*ni+i)*kernel+k];
        }
        y[o*length+t]=.5f*s*(1.f+erff(s*.7071067811865475f));
    }
}
static void lp_embed(const unsigned char *bytes,int n,int length,const float *w,float *out) {
    for(int i=0;i<length;i++) for(int c=0;c<24;c++) out[c*length+i]=w[(i<n ? (int)bytes[i]+1:0)*24+c];
}
static float lp_dot(const float *x,const float *w,int n,float b) { for(int i=0;i<n;i++) b+=x[i]*w[i];return b; }
typedef struct {int position, replacement, proposed; float confidence; const char *status;} PatchResult;
static PatchResult lp_predict(const PatchModel *m,const char *request,const unsigned char *bytes,int n) {
    size_t size=strlen(request);
    if(!size || size>96 || n<1 || n>256) die("expected 1..96 command bytes and 1..256 binary bytes");
    float ce[24*96],c1[48*96],c2[48*96],pooled[288],c[64];
    float be[24*256],b1[64*256],b2[64*256],h[256*64],joint[128],scores[257];
    lp_embed((const unsigned char*)request,(int)size,96,m->ce,ce);
    lp_conv(ce,c1,96,24,48,5,m->c1w,m->c1b);lp_conv(c1,c2,96,48,48,5,m->c2w,m->c2b);
    for(int k=0;k<48;k++) for(int bin=0;bin<6;bin++) {
        float v=-INFINITY; for(int t=bin*16;t<(bin+1)*16;t++) if(c2[k*96+t]>v) v=c2[k*96+t]; pooled[k*6+bin]=v;
    }
    for(int k=0;k<64;k++) c[k]=tanhf(lp_dot(pooled,m->cfw+k*288,288,m->cfb[k]));
    lp_embed(bytes,n,256,m->be,be);lp_conv(be,b1,256,24,64,9,m->b1w,m->b1b);lp_conv(b1,b2,256,64,64,9,m->b2w,m->b2b);
    for(int t=0;t<256;t++) {
        for(int k=0;k<64;k++) {joint[k]=b2[k*256+t];joint[k+64]=c[k];}
        for(int k=0;k<64;k++) h[t*64+k]=tanhf(lp_dot(joint,m->fw+k*128,128,m->fb[k]));
        scores[t]=t<n ? lp_dot(h+t*64,m->pw,64,m->pb[0]):-1e9f;
    }
    for(int k=0;k<64;k++) { float v=-INFINITY;for(int t=0;t<256;t++) if(b2[k*256+t]>v) v=b2[k*256+t]; joint[k]=v; joint[k+64]=c[k]; }
    scores[256]=lp_dot(joint,m->nw,128,m->nb[0]);
    int pos=0;for(int t=0;t<257;t++) {if(!isfinite(scores[t])) die("nonfinite inference");if(scores[t]>scores[pos]) pos=t;}
    float sum=0;for(int t=0;t<257;t++) sum+=expf(scores[t]-scores[pos]);
    PatchResult r={pos,0,0,1.f/sum,"none"};
    if(pos==256) return r;
    for(int bit=0;bit<8;bit++) {
        float logit=lp_dot(h+pos*64,m->rw+bit*64,64,m->rb[bit]);
        if(!isfinite(logit)) die("nonfinite inference");
        if(logit>=0) r.replacement |= 1<<bit;
    }
    if(r.confidence<m->threshold || pos>=n) {r.status="uncertain";return r;}
    for(int bit=0;bit<8;bit++) {float v=lp_dot(h+pos*64,m->rw+bit*64,64,m->rb[bit]);float p=1.f/(1.f+expf(-fabsf(v)));if(p<r.confidence)r.confidence=p;}
    if(r.confidence<m->threshold) {r.status="uncertain_byte";return r;}
    if(r.replacement==bytes[pos]) {r.status="no_change";return r;}
    r.proposed=1;r.status="unvalidated_candidate";return r;
}
/* Publish with link(2), not overwrite-capable rename: even a destination
 * created concurrently cannot be replaced. The reused copier stages privately. */
static void lp_emit(const char *src, const char *out, const unsigned char *bytes, int n) {
    struct stat s;
    if (lstat(out, &s) == 0 || errno != ENOENT)
        die("output must be a new path");
    size_t size = strlen(out);
    char *tmp = malloc(size + sizeof ".lppublish.XXXXXX");
    if (!tmp) die("out of memory");
    memcpy(tmp, out, size);
    strcpy(tmp + size, ".lppublish.XXXXXX");
    int fd = mkstemp(tmp);
    if (fd < 0) die("cannot stage candidate");
    if (close(fd)) {
        unlink(tmp);
        free(tmp);
        die("cannot close staging file");
    }
    stage_executable(src, tmp, "rv32", bytes, n, 1, 0);
    int failed = link(tmp, out);
    unlink(tmp);
    free(tmp);
    if (failed) die("candidate publish failed; destination was not replaced");
}

#include "responses_http.h"

int main(int argc, char **argv) {
    PatchModel model = lp_load(); /* Fail closed even for init/dump. */
    if (argc == 3 && !strcmp(argv[1], "--serve"))
        return rh_serve(&model, argv[2]);
    /* Open the running inode, even if its original pathname was replaced or
     * unlinked. Both read_self and stage_executable open this same image. */
    const char *src = "/proc/self/exe";
    char arch[16];
    unsigned char bytes[256];
    if (argc == 5 && !strcmp(argv[1], "--init") && !strcmp(argv[3], "-o")) {
        size_t n;
        unsigned char *raw = read_file(argv[2], &n, 256);
        if (!n || n % 4) die("seed must be nonempty aligned raw RV32");
        lp_emit(src, argv[4], raw, (int)n);
        free(raw);
        return 0;
    }
    int n = read_self(src, arch, bytes);
    if (!n || n % 4 || strcmp(arch, "rv32"))
        die("initialize a bounded raw RV32 slot first");
    if (argc == 2 && !strcmp(argv[1], "--dump-self")) {
        if (fwrite(bytes, 1, (size_t)n, stdout) != (size_t)n)
            die("stdout write failed");
        return 0;
    }
    int inspect = argc == 3 && !strcmp(argv[1], "--inspect");
    int propose = argc == 5 && !strcmp(argv[1], "--propose") && !strcmp(argv[3], "-o");
    if (!inspect && !propose)
        die("usage: --serve PORT | --init raw.rv32 -o new | --inspect 'request' | --propose 'request' -o new | --dump-self");
    PatchResult r = lp_predict(&model, argv[2], bytes, n);
    printf("{\"position\":%d,\"replacement\":%d,\"confidence\":%.9g,\"proposed\":%s,\"status\":\"%s\"}\n",
           r.position, r.replacement, r.confidence, r.proposed ? "true" : "false", r.status);
    if (propose && r.proposed) {
        bytes[r.position] = (unsigned char)r.replacement;
        lp_emit(src, argv[4], bytes, n);
    }
    return propose && !r.proposed ? 3 : 0;
}