"""Story/dialogue register of the FULL attention context per site.

For each " Name" site at global pos g, the model predicting g sees exactly
[cstart, g) with cstart = ((g-1)//1024)*1024 (train block geometry).
Measure marker presence in that prefix (and simple density stats).
"""
import json, os, sys
import numpy as np

REPO = "/mnt/nlpgpu-io1/data/maggie/explain-m0de-collapse/replicate"
sys.path.insert(0, REPO)
from frozen_token_loader import FrozenTokenStream

BASE = os.path.join(REPO, "out/alice_vs_megan")
SITES = os.path.join(BASE, "sites")
STATS = os.path.join(BASE, "stats")
BLOCK = 1024
EOS = 100257
FOCUS = ["Alice", "Mary", "Maria", "Julia", "Ann", "Megan", "Karen", "Sarah"]

from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained(
    os.path.join(REPO, "data/olmo3_prefix_hfstream_50b/tokenizer"), trust_remote_code=True)
enc1 = lambda s: tok(s, add_special_tokens=False)["input_ids"]

stream = FrozenTokenStream(os.path.join(REPO, "data/olmo3_prefix_hfstream_50b"))
sites = np.load(os.path.join(SITES, "sites_all.npz"))

# marker id sets
ids_once = set(enc1(" Once") + enc1("Once"))
ids_upon = set(enc1(" upon"))
ids_story = set(enc1(" story") + enc1(" stories") + enc1(" Story") + enc1(" tale") + enc1(" tales"))
ids_girl = set(enc1(" girl") + enc1(" girls") + enc1(" Girl"))
ids_named = set(enc1(" named"))
ids_she = set(enc1(" she") + enc1(" She") + enc1(" her") + enc1(" Her") + enc1(" hers"))
ids_said = set(enc1(" said") + enc1(" says") + enc1(" asked") + enc1(" replied"))
QUOTE_CHARS = set('"“”‘’')
quote_ids = set()
for i in range(100278):
    try:
        s = tok.decode([i])
    except Exception:
        continue
    if any(c in QUOTE_CHARS for c in s):
        quote_ids.add(i)

def mask(idset):
    m = np.zeros(100352, bool); m[list(idset)] = True; return m
M = {k: mask(v) for k, v in dict(once=ids_once, upon=ids_upon, story=ids_story,
                                 girl=ids_girl, named=ids_named, she=ids_she,
                                 said=ids_said, quote=quote_ids).items()}

out = {}
per_site = {}
for nm in FOCUS:
    p = np.sort(sites[f"pos_{nm.join(['sp_',''])}" if False else f"pos_sp_{nm}"])
    n = len(p)
    if n == 0: continue
    flags = {k: np.zeros(n, bool) for k in M}
    she_cnt = np.zeros(n, np.int32)
    plen = np.zeros(n, np.int32)
    # group by block to reuse reads
    cstart = ((p - 1) // BLOCK) * BLOCK
    uniq, inv = np.unique(cstart, return_inverse=True)
    for bi, cs in enumerate(uniq):
        idx = np.flatnonzero(inv == bi)
        end = int(p[idx].max())
        buf = stream.get_tokens(int(cs), end - int(cs) + 1)
        for i in idx:
            pref = buf[: int(p[i]) - int(cs)]
            # cut at last EOS: attention crosses docs, but register comes from the
            # CURRENT doc; compute both (whole prefix + since-doc-start)
            plen[i] = len(pref)
            for k, m in M.items():
                flags[k][i] = m[pref].any()
            she_cnt[i] = int(M["she"][pref].sum())
    fairy = flags["once"] & flags["upon"]
    story_reg = fairy | flags["story"] | (flags["girl"] & flags["she"])
    dialog = flags["quote"] | flags["said"]
    out[nm] = {
        "n": int(n),
        "prefix_len_mean": round(float(plen.mean()), 1),
        "pct_once_upon_prefix": round(float(fairy.mean()), 4),
        "pct_story_word": round(float(flags["story"].mean()), 4),
        "pct_girl_word": round(float(flags["girl"].mean()), 4),
        "pct_named_word": round(float(flags["named"].mean()), 4),
        "pct_she_her": round(float(flags["she"].mean()), 4),
        "she_per_prefix_mean": round(float(she_cnt.mean()), 2),
        "pct_quote": round(float(flags["quote"].mean()), 4),
        "pct_said": round(float(flags["said"].mean()), 4),
        "pct_story_register": round(float(story_reg.mean()), 4),
        "pct_dialog": round(float(dialog.mean()), 4),
        "n_story_register": int(story_reg.sum()),
        "n_dialog": int(dialog.sum()),
        "n_girl_and_she": int((flags["girl"] & flags["she"]).sum()),
    }
    per_site[nm] = {"pos": p, "story_reg": story_reg, "dialog": dialog,
                    "she": flags["she"], "girl": flags["girl"]}
    print(nm, json.dumps(out[nm]), flush=True)

np.savez_compressed(os.path.join(SITES, "block_register_flags.npz"),
                    **{f"{nm}_{k}": v for nm, d in per_site.items()
                       for k, v in d.items() if k != "pos"},
                    **{f"{nm}_pos": d["pos"] for nm, d in per_site.items()})
json.dump(out, open(os.path.join(STATS, "block_register.json"), "w"), indent=1)
print("written", os.path.join(STATS, "block_register.json"))