ABAP-Indexing/summarize/chunks.py

"""로직 조각 후처리 — LLM 이 준 조각을 코드와 대조해 확정한다 (docs/logic-chunk-design.md).

원칙: 사실은 파서가, 해석은 LLM 이.
- 줄 번호: first_line 앵커로 검증·보정, 실패하면 버림
- 테이블 · 호출: 조각 코드 범위를 파서로 다시 돌려 채움 (LLM 값은 쓰지 않음)
- 해시: 조각 코드 sha256
"""
from __future__ import annotations

import hashlib
import re
from dataclasses import dataclass, field

from parser.refs import extract_refs
from parser.statements import split_statements

from .schemas import CHUNK_KINDS, LogicChunk

MIN_LINES = 1
_WS = re.compile(r"\s+")
_SQL_HEAD = re.compile(
    r"^\s*(SELECT|OPEN\s+CURSOR|INSERT|UPDATE|MODIFY|DELETE|COMMIT\s+WORK|ROLLBACK\s+WORK|"
    r"CALL\s+FUNCTION|CALL\s+TRANSACTION|CALL\s+METHOD|SUBMIT|PERFORM|AUTHORITY-CHECK|"
    r"LOOP\s+AT|COLLECT|MESSAGE)\b",
    re.I,
)


def _norm(s: str) -> str:
    return _WS.sub(" ", (s or "").strip()).lower()


@dataclass
class ResolvedChunk:
    seq: int
    line_start: int
    line_end: int
    code: str
    code_hash: str
    kind: str
    purpose_ko: str
    purpose_en: str
    keywords_ko: list[str]
    keywords_en: list[str]
    sap_objects: list[str]
    tables_read: list[str] = field(default_factory=list)
    tables_write: list[str] = field(default_factory=list)
    calls: list[str] = field(default_factory=list)
    confidence: float = 0.5


def numbered_code(lines: list[str], line_start: int, line_end: int) -> str:
    """include 전체 줄 목록에서 [line_start, line_end] 를 '  12| code' 형식으로."""
    out = []
    for no in range(line_start, line_end + 1):
        if 1 <= no <= len(lines):
            out.append(f"{no:5d}| {lines[no - 1]}")
    return "\n".join(out)


def parser_hints(lines: list[str], line_start: int, line_end: int, limit: int = 40) -> list[str]:
    """DB 접근 · 호출 · 검증 문장이 시작되는 줄 — LLM 이 자를 후보 지점 힌트."""
    hints = []
    for no in range(line_start, line_end + 1):
        if 1 <= no <= len(lines):
            m = _SQL_HEAD.match(lines[no - 1])
            if m:
                hints.append(f"L{no} {lines[no - 1].strip()[:80]}")
                if len(hints) >= limit:
                    break
    return hints


def _find_anchor(lines: list[str], anchor: str, lo: int, hi: int) -> int | None:
    """[lo, hi] 안에서 anchor(정규화) 와 같은 줄 번호. 없으면 접두 일치."""
    a = _norm(anchor)
    if not a:
        return None
    for no in range(lo, hi + 1):
        if _norm(lines[no - 1]) == a:
            return no
    head = a[:20]
    if len(head) >= 8:
        for no in range(lo, hi + 1):
            if _norm(lines[no - 1]).startswith(head):
                return no
    return None


def resolve_chunks(raw: list[LogicChunk], lines: list[str], unit_start: int, unit_end: int,
                   include: str, known_symbols: set[str]) -> tuple[list[ResolvedChunk], list[str]]:
    """LLM 조각 → 검증·보정·사실 채움. 반환: (확정 조각(줄 순), 버린 사유 목록)."""
    dropped: list[str] = []
    resolved: list[ResolvedChunk] = []
    seen_ranges: set[tuple[int, int]] = set()

    for c in raw:
        s, e = int(c.line_start), int(c.line_end)
        if e < s:
            s, e = e, s
        length = e - s

        # 1) 앵커 검증 — line_start 줄이 first_line 과 다르면 unit 안에서 찾아 보정
        if c.first_line:
            ok = unit_start <= s <= unit_end and _norm(lines[s - 1]) == _norm(c.first_line) \
                if 1 <= s <= len(lines) else False
            if not ok:
                found = _find_anchor(lines, c.first_line, unit_start, unit_end)
                if found is None:
                    dropped.append(f"L{s}-L{e}: first_line 을 unit 안에서 찾지 못함 ({c.first_line[:40]!r})")
                    continue
                s, e = found, found + length

        # 2) 범위 클램프
        if s < unit_start or s > unit_end:
            dropped.append(f"L{s}-L{e}: unit 범위(L{unit_start}-L{unit_end}) 밖")
            continue
        e = min(e, unit_end)
        if e - s + 1 < MIN_LINES:
            dropped.append(f"L{s}-L{e}: 너무 짧음")
            continue
        if (s, e) in seen_ranges:
            dropped.append(f"L{s}-L{e}: 중복 범위")
            continue
        seen_ranges.add((s, e))

        # 3) 코드 · 해시 · 파서 사실
        code = "\n".join(lines[s - 1 : e])
        stmts, _ = split_statements(code, include)
        refs = extract_refs(stmts, list(range(len(stmts))), set(), known_symbols)
        kind = c.kind if c.kind in CHUNK_KINDS else "other"
        resolved.append(ResolvedChunk(
            seq=0, line_start=s, line_end=e, code=code,
            code_hash=hashlib.sha256(code.encode("utf-8")).hexdigest(),
            kind=kind, purpose_ko=c.purpose_ko.strip(), purpose_en=c.purpose_en.strip(),
            keywords_ko=list(dict.fromkeys(k.strip() for k in c.keywords_ko if k.strip()))[:15],
            keywords_en=list(dict.fromkeys(k.strip() for k in c.keywords_en if k.strip()))[:15],
            sap_objects=list(dict.fromkeys(k.strip().upper() for k in c.sap_objects if k.strip()))[:20],
            tables_read=refs.tables_read, tables_write=refs.tables_write,
            calls=[x["target"] for x in refs.calls][:20],
            confidence=max(0.0, min(1.0, float(c.confidence))),
        ))

    resolved.sort(key=lambda r: (r.line_start, r.line_end))
    for i, r in enumerate(resolved, 1):
        r.seq = i
    return resolved, dropped


def covered_lines(chunks: list[ResolvedChunk]) -> int:
    covered: set[int] = set()
    for c in chunks:
        covered.update(range(c.line_start, c.line_end + 1))
    return len(covered)