#!/usr/bin/env python3
"""verdicts.py — the ONE verdict-verification implementation (protocol 2).

Both consumers run the SAME checks over a verdict record, so the gate
and the round-time validator can never drift (the W25 disposition):

  * check_join.py (the gate) — rejects rule-quote / evidence claims the
    tree does not reproduce; a rejected verdict returns to the missing
    list and the run is INCOMPLETE until redone;
  * batch_judge.py (round time) — the same checks run early so the
    model can correct before the gate refuses.

A verdict record:

    {"adr", "verdict", "rule_quote", "basis",
     "evidence": [{"query" | "query_ref", "in",
                   "expect": "some" | "none",
                   "matches": [{"at": "file:line", "text"}], "note"}],
     "dormant_mechanism"?}

The checks:
  * rule_quote is required and must be a whitespace-normalized
    substring of the ADR's ## Rule section (protocol 2 — the rule is
    the yardstick, not the whole file; W11);
  * every evidence record resolves to a replayable query — `call
    <name>` or a profile query_ref — and is RE-EXECUTED against the
    tree; expect:"some" must reproduce every listed match (at + text);
    expect:"none" must have zero structural matches AND zero textual
    hits;
  * on replayable platforms a `partial` needs >=1 evidence record and
    a `not_verifiable` must name the searches it ran.

Stdlib + tsq (tree-sitter) when an engine is available; without one,
evidence degrades to the quote-level fallback and says so.
"""

import os
import re

VERDICTS = ["conformant", "partial", "unwitnessed", "not_verifiable"]
CALL_RE = re.compile(r"^call\s+(\S+)$")
RULE_SECTION = re.compile(r"^#+\s*Rule\s*$(.*?)(?=^#+\s|\Z)", re.M | re.S)


def norm_ws(s):
    return re.sub(r"\s+", " ", s).strip()


def rule_scope(adr_text):
    """The ## Rule section's bytes, or None when the ADR has none."""
    m = RULE_SECTION.search(adr_text or "")
    return m.group(1) if m else None


class Verifier:
    """repo + profile queries + a tsq module (None = no engine: evidence
    falls back to quote-level verification, reported in notes)."""

    def __init__(self, repo, profile_queries=None, tsq_mod=None):
        self.repo = repo
        self.profile_queries = profile_queries or {}
        self.tsq = tsq_mod

    # ── query resolution ─────────────────────────────────────────────

    def resolve_query(self, ev):
        """An evidence record's replayable query, or None. Citable
        forms: `call <name>` (the vetted built-in), a profile-ratified
        `query_ref` (kind call or scm), or a CONSTRUCTED tree-sitter
        query (`query_scm`) — the retrieval net the judge cast:
        presence citations replay as-is; absence citations additionally
        require positive controls (run_controls)."""
        if ev.get("query_ref"):
            named = self.profile_queries.get(ev["query_ref"])
            if isinstance(named, dict) and named.get("kind") == "call":
                return {"name": named["name"], "in": named.get("in") or ev.get("in")}
            if isinstance(named, dict) and named.get("kind") == "scm":
                return {"scm": named["scm"], "in": named.get("in") or ev.get("in")}
            return None
        if ev.get("query_scm"):
            return {"scm": ev["query_scm"], "in": ev.get("in")}
        m = CALL_RE.match((ev.get("query") or "").strip())
        return {"name": m.group(1), "in": ev.get("in")} if m else None

    # ── replay ───────────────────────────────────────────────────────

    def replay(self, q):
        """Re-execute a query (call or scm). [{'at','text'}] — or None
        (no engine). A malformed constructed query raises ValueError."""
        if self.tsq is None:
            return None
        eng = self.tsq
        lang = eng.LANGS["elixir"]
        prefixes = [q["in"]] if q.get("in") else []
        out = []
        parser = eng.tree_sitter.Parser(lang)
        if q.get("scm"):
            try:
                eng.tree_sitter.Query(lang, q["scm"])
            except Exception as e:
                raise ValueError(f"constructed query does not compile: {e}")
        for rel in eng.code_files(self.repo, prefixes):
            try:
                src = open(os.path.join(self.repo, rel), "rb").read()
            except OSError:
                continue
            root = parser.parse(src).root_node
            nodes = (
                eng.run_scm(lang, src, root, q["scm"])
                if q.get("scm")
                else eng.find_calls(src, root, q["name"])
            )
            for node in nodes:
                out.append(
                    {
                        "at": f"{rel}:{node.start_point[0] + 1}",
                        "text": src[node.start_byte : node.end_byte].decode(errors="replace"),
                    }
                )
        return out

    def run_controls(self, ev):
        """Positive controls for a constructed-query ABSENCE citation:
        each control is a code fragment plus the match count the query
        must find in it. 'The net caught nothing' is citable only when
        the net has proven teeth. Returns problems."""
        problems = []
        controls = ev.get("controls") or []
        if not controls:
            return [
                "an absence citation from a constructed query requires >=1 "
                "positive control ({source, matches}) proving the query "
                "matches what it claims to search for"
            ]
        eng = self.tsq
        lang = eng.LANGS["elixir"]
        parser = eng.tree_sitter.Parser(lang)
        any_teeth = False
        for j, c in enumerate(controls):
            src = (c.get("source") or "").encode()
            root = parser.parse(src).root_node
            try:
                nodes = eng.run_scm(lang, src, root, ev["query_scm"])
            except Exception as e:
                problems.append(f"controls[{j}]: query failed on control: {e}")
                continue
            want = c.get("matches")
            if want is not None and len(nodes) != want:
                problems.append(
                    f"controls[{j}]: declared {want} match(es), query found {len(nodes)}"
                )
            if len(nodes) >= 1:
                any_teeth = True
        if not any_teeth:
            problems.append(
                "no control demonstrates the query matching anything — "
                "the net has no proven teeth; the absence is not citable"
            )
        return problems

    def textual_hits(self, name, prefix):
        if self.tsq is None:
            return 0
        return self.tsq.textual_hits(self.repo, [prefix] if prefix else [], name)

    # ── the checks ───────────────────────────────────────────────────

    def verify_evidence(self, v):
        """Replay a verdict's evidence records.

        Returns (problems, notes): problems reject the verdict; notes
        are degradations to report (engine unavailable)."""
        problems, notes = [], []
        for i, ev in enumerate(v.get("evidence") or []):
            if not isinstance(ev, dict):
                problems.append(f"evidence[{i}]: not a record")
                continue
            q = self.resolve_query(ev)
            if q is None:
                problems.append(
                    f"evidence[{i}] is not a replayable query "
                    "(use `call <name>`, a profile query_ref, or a constructed `query_scm`)"
                )
                continue
            try:
                found = self.replay(q)
            except ValueError as e:
                problems.append(f"evidence[{i}]: {e}")
                continue
            if found is None:
                # no engine here. Presence claims degrade to the quote-level
                # fallback; a constructed ABSENCE cannot be verified at all
                # without replay — reject rather than trust.
                if ev.get("expect") == "none" and q.get("scm"):
                    problems.append(
                        f"evidence[{i}]: constructed absence citation cannot be "
                        "verified without the replay engine (tree-sitter)"
                    )
                    continue
                notes.append(
                    "evidence replay unavailable (tree-sitter not installed) "
                    "— quote-level fallback"
                )
                for m in ev.get("matches") or []:
                    rel = (m.get("at") or "").split(":")[0]
                    try:
                        body = open(
                            os.path.join(self.repo, rel), encoding="utf-8", errors="replace"
                        ).read()
                    except OSError:
                        body = ""
                    if not m.get("text") or norm_ws(m["text"]) not in norm_ws(body):
                        problems.append(f"evidence[{i}]: fallback quote failed at {m.get('at')}")
                continue
            if ev.get("expect") == "none":
                if found:
                    problems.append(
                        f"evidence[{i}]: expect none but {len(found)} structural "
                        f"match(es) — first {found[0]['at']}"
                    )
                elif q.get("scm"):
                    problems += [f"evidence[{i}]: {p}" for p in self.run_controls(ev)]
                else:
                    hits = self.textual_hits(q["name"], q.get("in"))
                    if hits:
                        problems.append(
                            f"evidence[{i}]: 0 structural matches but {hits} textual "
                            f"hit(s) for '{q['name']}' — inspect before citing absence"
                        )
            else:
                have = {m["at"]: m["text"] for m in found}
                if not (ev.get("matches") or []):
                    problems.append(f"evidence[{i}]: expect some but no matches recorded")
                for m in ev.get("matches") or []:
                    at = (m or {}).get("at")
                    if at not in have:
                        problems.append(f"evidence[{i}]: claimed match {at} not reproduced")
                    elif m.get("text") and norm_ws(m["text"]) not in norm_ws(have[at]):
                        problems.append(f"evidence[{i}]: text at {at} does not match the tree")
        return problems, notes

    def requirements(self, v, replayable):
        """The hard requirements (replayable platforms only): a partial
        must point at code; a not_verifiable must name its searches."""
        if not replayable:
            return []
        ev = v.get("evidence") or []
        if v.get("verdict") == "partial" and not ev:
            return ["a partial verdict requires >=1 replayable evidence record"]
        if v.get("verdict") == "not_verifiable" and not ev:
            return ["a not_verifiable verdict must name the queries it ran"]
        return []

    def validate(self, scope_text, rec, replayable=True):
        """All problems for one verdict record against the ADR's Rule
        scope. Returns (problems, notes)."""
        problems, notes = [], []
        quote = rec.get("rule_quote") or ""
        if not quote.strip():
            problems.append("missing rule_quote — quote the rule's load-bearing sentence VERBATIM")
        elif norm_ws(quote) not in norm_ws(scope_text or ""):
            problems.append(
                "rule_quote does not verify against the ADR's Rule section — quote VERBATIM"
            )
        ev_problems, ev_notes = self.verify_evidence(rec)
        problems += ev_problems
        notes += ev_notes
        problems += self.requirements(rec, replayable)
        return problems, notes