code/eval_general_phrase.py

"""Self-checking evaluation for the generic noisy-romaji rescue.

This is deliberately *generative*: recovery cases are produced by applying
corruption operators to clean compositions of known dictionary/colloquial
units, so the suite proves the mechanism generalizes rather than memorizing one
string. It also asserts the confidence gate *abstains* on heavy/ambiguous noise.

Corruptions are split by whether they preserve the intended *reading*:

* reading-preserving (style flip wapuro<->Hepburn, IME small-tsu spelling) are
  hard-asserted to recover exactly -- these are the realistic, high-value wins;
* reading-altering (triardupling, geminate drop, vowel inflation) are reported
  only, because a corrupted form may legitimately collide with another real word
  (e.g. ``itan`` -> 異端 vs ``ittan`` -> 一旦) and forcing one answer would be the
  very overfitting we are avoiding.

Run:

    python src/eval_general_phrase.py

Exits non-zero if any hard assertion fails.
"""

import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parent))

from normalization import normalize_input
from romaji_kana import load_general_lexicon
from general_phrase import (
    build_general_phrase_index,
    canon_style,
    canonicalize_romaji_variants,
    general_phrase_rescue,
    _collapse_runs,
    _geminate_sokuon,
)

DEFAULT_GENERAL_LEXICON = "artifacts/lexicon/general_reading_lexicon.json"

PASS = []
FAIL = []


def check(name, cond, detail=""):
    (PASS if cond else FAIL).append(name)
    mark = "ok  " if cond else "FAIL"
    line = f"[{mark}] {name}"
    if detail:
        line += f"  ::  {detail}"
    print(line)


def report(name, value):
    print(f"[rpt ] {name}  ::  {value}")


# --------------------------------------------------------------------------- #
# Corruption operators (deterministic; no RNG)
# --------------------------------------------------------------------------- #

def op_style_flip(s):
    # wapuro -> Hepburn-ish flips that canon_style must absorb (reading-preserving)
    for a, b in (("sy", "sh"), ("ty", "ch"), ("tu", "tsu"), ("hu", "fu"), ("zy", "j")):
        if a in s:
            return s.replace(a, b, 1)
    return s


def op_sokuon_ime(s):
    # a geminate consonant -> IME small-tsu spelling (reading-preserving)
    for i in range(len(s) - 1):
        if s[i] == s[i + 1] and s[i] in "bcdfghjkmpqrstvwyz":
            return s[:i] + "xtu" + s[i + 1:]
    return s


def op_run(s):
    # triple the first consonant/glide (may alter reading near geminates)
    for i, ch in enumerate(s):
        if ch in "ybcdfghjkmpqrstvwz":
            return s[:i] + ch * 3 + s[i + 1:]
    return s


def op_drop_geminate(s):
    for i in range(len(s) - 1):
        if s[i] == s[i + 1] and s[i] in "bcdfghjkmpqrstvwyz":
            return s[:i] + s[i + 1:]
    return s


def op_long_vowel_inflate(s):
    for i, ch in enumerate(s):
        if ch in "aiueo":
            return s[:i] + ch + ch + s[i:]
    return s


# --------------------------------------------------------------------------- #

def main():
    g = load_general_lexicon(DEFAULT_GENERAL_LEXICON)
    index = build_general_phrase_index(g)

    print("=== canonicalizer unit tests ===")
    check("canon sh->sy", canon_style("shudan") == "syudan", canon_style("shudan"))
    check("canon sy stable", canon_style("syudan") == "syudan")
    check("canon shuuryou->syuuryou", canon_style("shuuryou") == "syuuryou", canon_style("shuuryou"))
    check("canon tsu->tu", canon_style("tsunami") == "tunami", canon_style("tsunami"))
    check("canon chi->ti", canon_style("chizu") == "tizu", canon_style("chizu"))
    check("canon fu->hu", canon_style("fujisan") == "huzisan", canon_style("fujisan"))
    check("canon ji->zi", canon_style("jisho") == "zisyo", canon_style("jisho"))
    check("sokuon moxtute->motte", _geminate_sokuon("moxtute") == "motte", _geminate_sokuon("moxtute"))
    check("sokuon ixtukai->ikkai", _geminate_sokuon("ixtukai") == "ikkai", _geminate_sokuon("ixtukai"))
    check("sokuon kixtute->kitte", _geminate_sokuon("kixtute") == "kitte", _geminate_sokuon("kixtute"))
    check("run syyyuu->syuu", _collapse_runs("syyyuu") == "syuu", _collapse_runs("syyyuu"))
    check("run aaaa->aa (vowel keeps 2)", _collapse_runs("aaaa") == "aa", _collapse_runs("aaaa"))
    check("run uu stable", _collapse_runs("uu") == "uu")
    variants = canonicalize_romaji_variants(
        "imamoxtutewruusyudannowataraitannsyyyuuryoudemoiiyo"
    )
    check("variants fold motte+syuu",
          any("motte" in v and "syuu" in v for v in variants),
          " | ".join(variants))
    check("variants repair tewruu only as additive candidate",
          any("imamotteru" in v for v in variants),
          " | ".join(variants))

    print("\n=== recovery: clean compositions of known units (HARD) ===")
    compositions = [
        (["ima", "motteru"], "今持ってる"),
        (["shudan", "no", "hani", "nara"], "手段の範囲なら"),
        (["ittan", "syuuryou", "demo", "iiyo"], "一旦終了でもいいよ"),
        (["ima", "motteru", "shudan", "no", "hani", "nara",
          "ittan", "syuuryou", "demo", "iiyo"],
         "今持ってる手段の範囲なら一旦終了でもいいよ"),
    ]
    clean_inputs = []
    for keys, expected in compositions:
        clean = "".join(keys)
        clean_inputs.append((clean, expected))
        res = general_phrase_rescue(clean, index)
        out = res[0] if res else None
        check(f"clean recover: {clean}", out == expected, f"got={out!r} want={expected!r}")

    print("\n=== recovery: reading-preserving corruptions (HARD) ===")
    safe_ops = [("style_flip", op_style_flip), ("sokuon_ime", op_sokuon_ime)]
    for clean, expected in clean_inputs:
        for opname, op in safe_ops:
            corrupt = op(clean)
            if corrupt == clean:
                continue
            res = general_phrase_rescue(corrupt, index)
            out = res[0] if res else None
            check(f"{opname}: {corrupt}", out == expected, f"got={out!r} want={expected!r}")

    print("\n=== recovery: reading-altering corruptions (REPORT ONLY) ===")
    risky_ops = [("run_triple", op_run), ("drop_geminate", op_drop_geminate),
                 ("long_vowel", op_long_vowel_inflate)]
    for clean, expected in clean_inputs:
        for opname, op in risky_ops:
            corrupt = op(clean)
            if corrupt == clean:
                continue
            res = general_phrase_rescue(corrupt, index)
            report(f"{opname}: {corrupt}", res[0] if res else None)

    print("\n=== negative gate: heavy / ambiguous noise must abstain (HARD) ===")
    negatives = [
        "xqzkwbvfjpmnlrtxqz",
        "zzzzzzzzzzzzzzzz",
        "qwlkjghfdspqwlkjghfds",
    ]
    for neg in negatives:
        res = general_phrase_rescue(neg, index)
        check(f"abstain: {neg}", res is None, f"got={res[0] if res else None!r}")

    print("\n=== contamination guard: target intent recovers when typed closer (HARD) ===")
    # The target sentence IS recoverable once the *severe* corruptions are typed
    # closer to canonical: moxtuteru (IME small-tsu) -> 持ってる, haninara -> 範囲なら,
    # ittan -> 一旦. Proves the route is general, not a single-string fit.
    closer = "imamoxtuterusyudannohaninaraittansyuuryoudemoiiyo"
    res = general_phrase_rescue(closer, index)
    out = res[0] if res else None
    want = "今持ってる手段の範囲なら一旦終了でもいいよ"
    check("closer-typing recovers full phrase", out == want, f"got={out!r}")

    print("\n=== recovery: mixed case / separators with reusable noise (HARD) ===")
    noisy = "IMAMO-XTUteWRUU SYUDANNO-HANINARA ITANNSYYYUURYOUDEMOIIYO"
    res = general_phrase_rescue(normalize_input(noisy), index)
    out = res[0] if res else None
    check("case/space/hyphen + tewruu + itannsyyyuu recovers", out == want, f"got={out!r}")

    print("\n=== safety guard: real target case must abstain (HARD) ===")
    target = "imamoxtutewruusyudannowataraitannsyyyuuryoudemoiiyo"
    res = general_phrase_rescue(target, index)
    check("target with watara ambiguity abstains", res is None, f"got={res[0] if res else None!r}")
    if res:
        m = res[1]
        report("target meta",
               f"anchor={m['anchor_ratio']:.2f} fill={m['fill_ratio']:.2f} "
               f"drops={m['drops']} cost/char={m['cost_per_char']:.3f}")

    print(f"\n==== {len(PASS)} passed, {len(FAIL)} failed ====")
    if FAIL:
        print("FAILURES:")
        for f in FAIL:
            print("  -", f)
        raise SystemExit(1)


if __name__ == "__main__":
    main()