its-magic 0.1.2-9 → 0.1.2

package/template/scripts/caveman_compress_input.py

@@ -0,0 +1,903 @@
+#!/usr/bin/env python3
+"""US-0090 / DEC-0073 — Caveman-style input-side file compression (safe-mode v1).
+
+Opt-in, default-off, script-invoked operator tool. Reads scratchpad gating
+(`CAVEMAN_COMPRESS_INPUT=1` + non-empty `CAVEMAN_FILE_SCOPE`), refuses deny-
+listed paths, applies a strictly-idempotent line-level minifier, and writes
+sidecar originals under `docs/.caveman-originals/<relative/path>/<file>`
+BEFORE target mutation. Stdlib only.
+
+Reason-code vocabulary (DEC-0073 §7) — 9 codes in 3 families:
+
+  Gating     CAVEMAN_COMPRESS_MODE_DISABLED
+             CAVEMAN_COMPRESS_FLAG_CONFLICT
+  Scope      CAVEMAN_COMPRESS_SCOPE_EMPTY
+             CAVEMAN_COMPRESS_SCOPE_UNKNOWN_PROFILE
+             CAVEMAN_COMPRESS_SCOPE_VIOLATION
+  Integrity  CAVEMAN_COMPRESS_DENY_HIT
+             CAVEMAN_COMPRESS_NOT_IDEMPOTENT
+             CAVEMAN_COMPRESS_LITERAL_REGION_DAMAGED
+             CAVEMAN_COMPRESS_ORIGINAL_MISSING
+
+No post-write codes. Any addition requires a subsequent DEC revising §7.
+"""
+
+from __future__ import annotations
+
+import argparse
+import fnmatch
+import hashlib
+import json
+import os
+import re
+import sys
+from pathlib import Path, PurePosixPath
+
+# ---------------------------------------------------------------------------
+# Deny-list baseline (DEC-0073 §4.1). Any edit here requires a subsequent DEC.
+# `deny_list_version` in --report is the SHA-256 of sorted(DENY_BASELINE) as
+# canonical JSON; T-005 subtest asserts the hash is stable (drift detection).
+# ---------------------------------------------------------------------------
+
+DENY_BASELINE: tuple[str, ...] = (
+    # Secrets
+    ".env",
+    ".env.*",
+    "**/.env",
+    "**/.env.*",
+    # Intake evidence
+    "handoffs/intake_evidence/*.json",
+    "handoffs/intake_evidence/**/*.json",
+    # Canonical product authority
+    "docs/product/backlog.md",
+    "docs/product/acceptance.md",
+    # Canonical engineering authority
+    "docs/engineering/state.md",
+    "docs/engineering/decisions.md",
+    "decisions/DEC-*.md",
+    # Sprint lifecycle evidence
+    "sprints/*/plan-verify.json",
+    "sprints/*/uat.json",
+    "sprints/*/summary.md",
+    "sprints/*/release-findings.md",
+    "sprints/*/qa-findings.md",
+    "sprints/*/tasks.md",
+    "sprints/*/sprint.md",
+    # Publish / runtime / install surfaces
+    "package.json",
+    "package-lock.json",
+    "installer.sh",
+    "installer.ps1",
+    "installer.py",
+    "installer.js",
+    "installer.cmd",
+    "installer.bat",
+    ".github/workflows/*.yml",
+    ".cursor/hooks/*.py",
+    "bin/its-magic.js",
+    "packaging/homebrew/*.rb",
+    # Contract surfaces
+    ".cursor/rules/*.mdc",
+    ".cursor/commands/*.md",
+    ".cursor/skills/**/SKILL.md",
+    # Manifest / parity sources
+    "docs/engineering/context/installer-owned-paths.manifest",
+    "docs/engineering/release-targets.json",
+    "docs/engineering/token-cost-parity-manifest.md",
+    # Binaries (extension class)
+    "**/*.png",
+    "**/*.jpg",
+    "**/*.jpeg",
+    "**/*.gif",
+    "**/*.webp",
+    "**/*.pdf",
+    "**/*.zip",
+    "**/*.tar",
+    "**/*.tar.gz",
+    "**/*.tgz",
+    "**/*.ico",
+    "**/*.woff",
+    "**/*.woff2",
+    "**/*.ttf",
+    "**/*.eot",
+    "**/*.otf",
+    "**/*.mp3",
+    "**/*.mp4",
+    "**/*.mov",
+    "**/*.wav",
+    "**/*.bin",
+    "**/*.exe",
+    "**/*.dll",
+)
+
+# Patterns from .gitignore that contribute as additional denies (§4 step 2).
+GITIGNORE_SECRET_PREFIXES: tuple[str, ...] = (
+    ".env",
+    "*secret*",
+    "*credential*",
+    "*token*",
+    "*private*",
+)
+
+# Frozen v1 allow-list profile table (DEC-0073 §5.1).
+FROZEN_PROFILES: dict[str, tuple[str, ...]] = {
+    "docs-prose-only": (
+        "docs/user-guides/**/*.md",
+        "docs/engineering/runbook.md",
+        "docs/engineering/state-archive/**/*.md",
+        "handoffs/archive/*.md",
+    ),
+}
+
+SIDECAR_ROOT_REL = "docs/.caveman-originals"
+
+REASON_CODES_BY_FAMILY: dict[str, tuple[str, ...]] = {
+    "Gating": (
+        "CAVEMAN_COMPRESS_MODE_DISABLED",
+        "CAVEMAN_COMPRESS_FLAG_CONFLICT",
+    ),
+    "Scope": (
+        "CAVEMAN_COMPRESS_SCOPE_EMPTY",
+        "CAVEMAN_COMPRESS_SCOPE_UNKNOWN_PROFILE",
+        "CAVEMAN_COMPRESS_SCOPE_VIOLATION",
+    ),
+    "Integrity": (
+        "CAVEMAN_COMPRESS_DENY_HIT",
+        "CAVEMAN_COMPRESS_NOT_IDEMPOTENT",
+        "CAVEMAN_COMPRESS_LITERAL_REGION_DAMAGED",
+        "CAVEMAN_COMPRESS_ORIGINAL_MISSING",
+    ),
+}
+
+ALL_REASON_CODES: tuple[str, ...] = tuple(
+    code for codes in REASON_CODES_BY_FAMILY.values() for code in codes
+)
+
+VENDOR_INSTALL_LEAK_TOKEN = "npx skills add"
+
+FENCE_RE = re.compile(r"^(`{3,}|~{3,})")
+
+
+# ---------------------------------------------------------------------------
+# Utility: canonical JSON + hashing.
+# ---------------------------------------------------------------------------
+
+def canonical_json(payload: object) -> str:
+    return json.dumps(payload, sort_keys=True, separators=(",", ":"))
+
+
+def deny_list_version_hash(deny_entries: tuple[str, ...] = DENY_BASELINE) -> str:
+    canonical = canonical_json(sorted(deny_entries))
+    return hashlib.sha256(canonical.encode("utf-8")).hexdigest()
+
+
+# ---------------------------------------------------------------------------
+# Glob evaluation. Paths use forward slashes (POSIX) regardless of host OS
+# per DEC-0073 §5 (grammar requires '/').
+# ---------------------------------------------------------------------------
+
+def _normalize_rel(rel: str) -> str:
+    return str(PurePosixPath(rel.replace("\\", "/")))
+
+
+def glob_match(pattern: str, rel_path: str) -> bool:
+    rel_path = _normalize_rel(rel_path)
+    pattern = pattern.replace("\\", "/")
+    if "**" in pattern:
+        # Translate ** to match any number of path segments (including zero).
+        parts = pattern.split("**")
+        regex_parts: list[str] = []
+        for i, part in enumerate(parts):
+            regex_parts.append(fnmatch.translate(part).rstrip(r"\Z"))
+            if i < len(parts) - 1:
+                regex_parts.append(r".*")
+        rx = re.compile("".join(regex_parts) + r"\Z")
+        return rx.match(rel_path) is not None
+    return fnmatch.fnmatchcase(rel_path, pattern)
+
+
+def any_glob_match(patterns: tuple[str, ...] | list[str], rel_path: str) -> bool:
+    for p in patterns:
+        if glob_match(p, rel_path):
+            return True
+    return False
+
+
+# ---------------------------------------------------------------------------
+# Scratchpad reader (stdlib). Parses simple KEY=VALUE lines; ignores comments.
+# ---------------------------------------------------------------------------
+
+def read_scratchpad(path: Path) -> dict[str, str]:
+    values: dict[str, str] = {}
+    if not path.is_file():
+        return values
+    text = path.read_text(encoding="utf-8")
+    key_re = re.compile(r"^\s*([A-Z][A-Z0-9_]+)\s*=\s*(.*?)\s*$")
+    for raw in text.splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#") or line.startswith("<!--"):
+            continue
+        m = key_re.match(line)
+        if m:
+            values[m.group(1)] = m.group(2)
+    return values
+
+
+# ---------------------------------------------------------------------------
+# Scope grammar (DEC-0073 §5).
+# ---------------------------------------------------------------------------
+
+class ScopeParseError(Exception):
+    def __init__(self, code: str, detail: str) -> None:
+        self.code = code
+        self.detail = detail
+        super().__init__(f"{code}: {detail}")
+
+
+def parse_scope(raw: str) -> tuple[list[str], str | None]:
+    """Return (glob_patterns, profile_name_or_None).
+
+    Raises ScopeParseError on CAVEMAN_COMPRESS_SCOPE_EMPTY or
+    CAVEMAN_COMPRESS_SCOPE_UNKNOWN_PROFILE.
+    """
+    if raw is None:
+        raise ScopeParseError("CAVEMAN_COMPRESS_SCOPE_EMPTY", "scope unset")
+    trimmed = raw.strip()
+    if not trimmed or trimmed.startswith("#"):
+        raise ScopeParseError("CAVEMAN_COMPRESS_SCOPE_EMPTY", "scope empty")
+
+    profile_name: str | None = None
+    globs: list[str] = []
+
+    if ";" in trimmed or trimmed.startswith("profile:") or trimmed.startswith("globs:"):
+        # Hybrid form: profile:<name>;globs:<csv> (order-independent).
+        tokens = [t.strip() for t in trimmed.split(";") if t.strip()]
+        profile_seen = False
+        globs_seen = False
+        for tok in tokens:
+            if tok.startswith("profile:"):
+                if profile_seen:
+                    raise ScopeParseError(
+                        "CAVEMAN_COMPRESS_SCOPE_UNKNOWN_PROFILE",
+                        "second profile: token in scope (one profile per scope)",
+                    )
+                profile_seen = True
+                profile_name = tok[len("profile:"):].strip()
+            elif tok.startswith("globs:"):
+                globs_seen = True
+                for g in tok[len("globs:"):].split(","):
+                    g = g.strip()
+                    if g:
+                        globs.append(g)
+            else:
+                # Bare token inside a hybrid string: treat as unknown grammar.
+                raise ScopeParseError(
+                    "CAVEMAN_COMPRESS_SCOPE_UNKNOWN_PROFILE",
+                    f"unrecognized token in hybrid scope: {tok!r}",
+                )
+        if not profile_seen and not globs_seen:
+            raise ScopeParseError(
+                "CAVEMAN_COMPRESS_SCOPE_UNKNOWN_PROFILE",
+                "hybrid scope has neither profile: nor globs:",
+            )
+    elif "," in trimmed or "/" in trimmed or "*" in trimmed or trimmed.endswith(".md"):
+        # Raw CSV globs form.
+        for g in trimmed.split(","):
+            g = g.strip()
+            if g:
+                globs.append(g)
+    else:
+        # Named profile form.
+        profile_name = trimmed
+
+    if profile_name is not None:
+        if profile_name not in FROZEN_PROFILES:
+            raise ScopeParseError(
+                "CAVEMAN_COMPRESS_SCOPE_UNKNOWN_PROFILE",
+                f"unknown profile {profile_name!r}; v1 allows: "
+                f"{sorted(FROZEN_PROFILES.keys())}",
+            )
+        globs = list(FROZEN_PROFILES[profile_name]) + globs
+
+    if not globs:
+        raise ScopeParseError(
+            "CAVEMAN_COMPRESS_SCOPE_EMPTY", "resolved scope produced zero globs"
+        )
+    return globs, profile_name
+
+
+# ---------------------------------------------------------------------------
+# Deny-list resolution (DEC-0073 §4). Deny always wins.
+# ---------------------------------------------------------------------------
+
+def load_gitignore_secret_patterns(repo_root: Path) -> list[str]:
+    gi = repo_root / ".gitignore"
+    patterns: list[str] = []
+    if not gi.is_file():
+        return patterns
+    for raw in gi.read_text(encoding="utf-8").splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#") or line.startswith("!"):
+            continue
+        for prefix in GITIGNORE_SECRET_PREFIXES:
+            if fnmatch.fnmatchcase(line, prefix) or line.startswith(prefix.strip("*")):
+                patterns.append(line)
+                break
+    return patterns
+
+
+def load_cursorignore_overlay(repo_root: Path, enabled: bool) -> list[str]:
+    if not enabled:
+        return []
+    ci = repo_root / ".cursorignore"
+    if not ci.is_file():
+        return []
+    out: list[str] = []
+    for raw in ci.read_text(encoding="utf-8").splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#") or line.startswith("!"):
+            continue
+        out.append(line)
+    return out
+
+
+def resolve_deny_patterns(repo_root: Path, ingest_cursorignore: bool) -> list[str]:
+    patterns: list[str] = list(DENY_BASELINE)
+    patterns.extend(load_gitignore_secret_patterns(repo_root))
+    patterns.extend(load_cursorignore_overlay(repo_root, ingest_cursorignore))
+    return patterns
+
+
+def file_is_denied(rel_path: str, deny_patterns: list[str], repo_root: Path) -> bool:
+    if any_glob_match(deny_patterns, rel_path):
+        return True
+    # Vendor-install leak: refuse any file that contains the banned token.
+    target = repo_root / rel_path
+    if target.is_file():
+        try:
+            head = target.read_bytes()[:65536]
+        except OSError:
+            return False
+        if VENDOR_INSTALL_LEAK_TOKEN.encode("utf-8") in head:
+            return True
+    return False
+
+
+# ---------------------------------------------------------------------------
+# Safe-mode minifier (DEC-0073 §6). Strictly idempotent by construction.
+# Inside fenced code blocks (zone 1 of DEC-0072 §4): NO mutation of line
+# bodies; blank line runs inside fences are NOT collapsed (fence content is
+# literal).
+# ---------------------------------------------------------------------------
+
+def compress_safe_mode(text: str) -> str:
+    ends_with_newline = text.endswith("\n") or text.endswith("\r\n") or text.endswith("\r")
+    # Normalize line endings first (step 3).
+    unified = text.replace("\r\n", "\n").replace("\r", "\n")
+    lines = unified.split("\n")
+    if unified.endswith("\n"):
+        # split produced a trailing empty element for the final newline.
+        has_eof_newline = True
+        body_lines = lines[:-1]
+    else:
+        has_eof_newline = ends_with_newline
+        body_lines = lines
+
+    out: list[str] = []
+    in_fence = False
+    fence_marker: str | None = None
+    prev_blank_outside_fence = False
+    for raw in body_lines:
+        if in_fence:
+            out.append(raw)
+            stripped = raw.lstrip()
+            if fence_marker is not None and stripped.startswith(fence_marker[0] * 3):
+                mfence = FENCE_RE.match(stripped)
+                if mfence and len(mfence.group(1)) >= len(fence_marker):
+                    in_fence = False
+                    fence_marker = None
+            prev_blank_outside_fence = False
+            continue
+
+        mfence = FENCE_RE.match(raw.lstrip())
+        if mfence:
+            in_fence = True
+            fence_marker = mfence.group(1)
+            out.append(raw)
+            prev_blank_outside_fence = False
+            continue
+
+        trimmed = raw.rstrip()
+        if trimmed == "":
+            if prev_blank_outside_fence:
+                continue
+            out.append("")
+            prev_blank_outside_fence = True
+        else:
+            out.append(trimmed)
+            prev_blank_outside_fence = False
+
+    joined = "\n".join(out)
+    if has_eof_newline:
+        joined += "\n"
+    return joined
+
+
+# ---------------------------------------------------------------------------
+# Literal-region scan (DEC-0072 §4 reused verbatim — 9 zones).
+# Safe-mode only mutates blank-line counts and trailing whitespace on non-
+# fenced lines. We verify:
+#   (a) fenced code block content is byte-identical between input and output
+#       (zone 1);
+#   (b) non-blank, non-fence lines of input rstripped equal non-blank, non-
+#       fence lines of output rstripped in order (zones 2–9: paths, IDs,
+#       reason codes, checklists, contract markers, strict-proof tuple
+#       fields, isolation-evidence fields, git refs are line-scoped tokens
+#       that are preserved by rstrip + blank-run collapse).
+# ---------------------------------------------------------------------------
+
+def _extract_line_maps(text: str) -> tuple[list[tuple[int, int, str]], list[str]]:
+    """Return (fence_segments, non_fence_content_lines).
+
+    fence_segments: list of (start_line, end_line, body_bytes_joined_with_LF)
+    non_fence_content_lines: rstripped non-blank content lines, in order.
+    """
+    unified = text.replace("\r\n", "\n").replace("\r", "\n")
+    lines = unified.split("\n")
+    if unified.endswith("\n") and lines and lines[-1] == "":
+        lines = lines[:-1]
+    fences: list[tuple[int, int, str]] = []
+    content: list[str] = []
+    in_fence = False
+    fence_start = -1
+    fence_marker: str | None = None
+    fence_body: list[str] = []
+    for idx, raw in enumerate(lines):
+        mfence = FENCE_RE.match(raw.lstrip())
+        if in_fence:
+            fence_body.append(raw)
+            if (
+                mfence is not None
+                and fence_marker is not None
+                and len(mfence.group(1)) >= len(fence_marker)
+            ):
+                fences.append((fence_start, idx, "\n".join(fence_body)))
+                in_fence = False
+                fence_marker = None
+                fence_body = []
+        else:
+            if mfence is not None:
+                in_fence = True
+                fence_start = idx
+                fence_marker = mfence.group(1)
+                fence_body = [raw]
+                continue
+            stripped = raw.rstrip()
+            if stripped:
+                content.append(stripped)
+    if in_fence:
+        # Unterminated fence: treat body as fence for literal preservation.
+        fences.append((fence_start, len(lines) - 1, "\n".join(fence_body)))
+    return fences, content
+
+
+def literal_region_preserved(original: str, proposed: str) -> tuple[bool, str]:
+    orig_fences, orig_content = _extract_line_maps(original)
+    new_fences, new_content = _extract_line_maps(proposed)
+    if [f[2] for f in orig_fences] != [f[2] for f in new_fences]:
+        return False, "fenced-code zone mutated"
+    if orig_content != new_content:
+        return False, "non-blank content-line tokens differ"
+    return True, ""
+
+
+# ---------------------------------------------------------------------------
+# File discovery.
+# ---------------------------------------------------------------------------
+
+def enumerate_scope(
+    repo_root: Path, allow_globs: list[str], deny_patterns: list[str]
+) -> tuple[list[str], list[str]]:
+    """Return (allowed_rel_paths, denied_rel_paths_under_allow)."""
+    matched: set[str] = set()
+    denied_under_allow: list[str] = []
+    for pat in allow_globs:
+        # Use python's recursive glob for `**`.
+        # repo_root.glob interprets `**` when included in the pattern via `**/`.
+        for p in repo_root.glob(pat):
+            if p.is_file():
+                rel = str(p.relative_to(repo_root)).replace("\\", "/")
+                matched.add(rel)
+    allowed: list[str] = []
+    for rel in sorted(matched):
+        if file_is_denied(rel, deny_patterns, repo_root):
+            denied_under_allow.append(rel)
+        else:
+            allowed.append(rel)
+    return allowed, denied_under_allow
+
+
+# ---------------------------------------------------------------------------
+# Atomic write helpers.
+# ---------------------------------------------------------------------------
+
+def atomic_write(target: Path, data: bytes) -> None:
+    target.parent.mkdir(parents=True, exist_ok=True)
+    tmp = target.with_suffix(target.suffix + ".caveman.tmp")
+    try:
+        tmp.write_bytes(data)
+        os.replace(tmp, target)
+    except Exception:
+        try:
+            if tmp.exists():
+                tmp.unlink()
+        except OSError:
+            pass
+        raise
+
+
+def sidecar_for(repo_root: Path, rel_path: str) -> Path:
+    return repo_root / SIDECAR_ROOT_REL / rel_path
+
+
+# ---------------------------------------------------------------------------
+# Fail-closed helpers.
+# ---------------------------------------------------------------------------
+
+def fail_closed(code: str, detail: str = "") -> int:
+    msg = f"REASON_CODE={code}"
+    if detail:
+        msg += f" detail={detail}"
+    print(msg, file=sys.stderr)
+    return 2
+
+
+# ---------------------------------------------------------------------------
+# Report builder.
+# ---------------------------------------------------------------------------
+
+def build_report(
+    *,
+    deny_version: str,
+    scope_resolved: list[str],
+    files_considered: list[str],
+    files_eligible: list[str],
+    files_would_write: list[str],
+    violations: list[dict[str, str]],
+    idempotency_check: dict[str, object],
+) -> dict[str, object]:
+    return {
+        "deny_list_version": deny_version,
+        "scope_resolved": scope_resolved,
+        "files_considered": files_considered,
+        "files_eligible": files_eligible,
+        "files_would_write": files_would_write,
+        "violations": violations,
+        "idempotency_check": idempotency_check,
+        "reason_codes_vocabulary": {
+            family: list(codes) for family, codes in REASON_CODES_BY_FAMILY.items()
+        },
+    }
+
+
+# ---------------------------------------------------------------------------
+# CLI.
+# ---------------------------------------------------------------------------
+
+def build_parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(
+        prog="caveman_compress_input.py",
+        description=(
+            "US-0090 / DEC-0073 safe-mode input-side compressor. "
+            "Default off. Opt-in via scratchpad. Deny always wins over allow. "
+            "Sidecar originals under docs/.caveman-originals/."
+        ),
+        add_help=True,
+        allow_abbrev=False,
+    )
+    p.add_argument(
+        "--dry-run", action="store_true",
+        help="(default) inventory + diff summary to stdout; no mutation.",
+    )
+    p.add_argument(
+        "--write", action="store_true",
+        help="Perform sidecar + target mutation on eligible files.",
+    )
+    p.add_argument(
+        "--verify-originals", action="store_true",
+        help="Walk sidecar tree and verify every sidecar has a target (and vice versa).",
+    )
+    p.add_argument(
+        "--report", action="store_true",
+        help="Emit JSON report on stdout (incompatible with --write).",
+    )
+    p.add_argument(
+        "--repo", type=Path, default=Path(__file__).resolve().parent.parent,
+        help="Repository root (default: parent of scripts/).",
+    )
+    return p
+
+
+def parse_args(argv: list[str]) -> tuple[argparse.Namespace, list[str]]:
+    parser = build_parser()
+    args, unknown = parser.parse_known_args(argv)
+    return args, unknown
+
+
+def detect_flag_conflict(
+    args: argparse.Namespace, unknown: list[str]
+) -> str | None:
+    if unknown:
+        return f"unknown flag(s): {unknown}"
+    if args.dry_run and args.write:
+        return "--dry-run with --write"
+    if args.write and args.verify_originals:
+        return "--write with --verify-originals"
+    if args.write and args.report:
+        return "--write with --report"
+    return None
+
+
+def verify_originals(repo_root: Path, deny_patterns: list[str]) -> tuple[bool, list[dict[str, str]]]:
+    sidecar_root = repo_root / SIDECAR_ROOT_REL
+    violations: list[dict[str, str]] = []
+    if not sidecar_root.is_dir():
+        return True, violations
+    for sidecar in sidecar_root.rglob("*"):
+        if not sidecar.is_file():
+            continue
+        rel_from_root = sidecar.relative_to(sidecar_root)
+        if rel_from_root.name == ".gitkeep":
+            continue
+        rel = str(rel_from_root).replace("\\", "/")
+        target = repo_root / rel
+        if not target.is_file():
+            violations.append({
+                "code": "CAVEMAN_COMPRESS_ORIGINAL_MISSING",
+                "path": rel,
+                "stage": "verify-originals",
+                "detail": "sidecar has no target",
+            })
+    return len(violations) == 0, violations
+
+
+def run(argv: list[str]) -> int:
+    args, unknown = parse_args(argv)
+
+    conflict = detect_flag_conflict(args, unknown)
+    if conflict:
+        return fail_closed("CAVEMAN_COMPRESS_FLAG_CONFLICT", conflict)
+
+    repo_root: Path = args.repo
+
+    # Read scratchpad (for gating + CAVEMAN_COMPRESS_INGEST_CURSORIGNORE overlay flag).
+    scratchpad_path = repo_root / ".cursor" / "scratchpad.md"
+    scratchpad = read_scratchpad(scratchpad_path)
+    mode = scratchpad.get("CAVEMAN_COMPRESS_INPUT", "0").strip()
+    scope_raw = scratchpad.get("CAVEMAN_FILE_SCOPE", "").strip()
+    ingest_overlay = scratchpad.get("CAVEMAN_COMPRESS_INGEST_CURSORIGNORE", "0").strip()
+    ingest_overlay_bool = ingest_overlay == "1"
+
+    deny_patterns = resolve_deny_patterns(repo_root, ingest_overlay_bool)
+    deny_version = deny_list_version_hash(DENY_BASELINE)
+
+    # Resolve scope regardless of mode so --report and --dry-run can narrate.
+    parse_err: ScopeParseError | None = None
+    allow_globs: list[str] = []
+    profile_name: str | None = None
+    if scope_raw:
+        try:
+            allow_globs, profile_name = parse_scope(scope_raw)
+        except ScopeParseError as e:
+            parse_err = e
+
+    # --verify-originals path.
+    if args.verify_originals:
+        ok, viols = verify_originals(repo_root, deny_patterns)
+        if args.report:
+            report = build_report(
+                deny_version=deny_version,
+                scope_resolved=allow_globs,
+                files_considered=[],
+                files_eligible=[],
+                files_would_write=[],
+                violations=viols,
+                idempotency_check={"status": "skipped (verify-originals mode)"},
+            )
+            print(canonical_json(report))
+        if not ok:
+            return fail_closed("CAVEMAN_COMPRESS_ORIGINAL_MISSING",
+                               f"{len(viols)} orphan(s)")
+        return 0
+
+    # Gating (applies to --dry-run + --write; --report alone may still narrate).
+    mutation_mode = args.write
+
+    if mutation_mode:
+        if mode != "1":
+            return fail_closed("CAVEMAN_COMPRESS_MODE_DISABLED",
+                               "CAVEMAN_COMPRESS_INPUT != 1")
+        if parse_err is not None:
+            return fail_closed(parse_err.code, parse_err.detail)
+        if not scope_raw:
+            return fail_closed("CAVEMAN_COMPRESS_SCOPE_EMPTY",
+                               "CAVEMAN_FILE_SCOPE empty")
+
+    # Non-write paths: gracefully report but do not touch files.
+    if parse_err is not None and (args.report or args.dry_run or not (
+        args.write or args.verify_originals
+    )):
+        # For pure report / pure dry-run, surface the scope issue but keep
+        # exit 0 so operators can inspect via --report.
+        if args.report:
+            report = build_report(
+                deny_version=deny_version,
+                scope_resolved=[],
+                files_considered=[],
+                files_eligible=[],
+                files_would_write=[],
+                violations=[{
+                    "code": parse_err.code,
+                    "path": "(scope)",
+                    "stage": "pre-write",
+                    "detail": parse_err.detail,
+                }],
+                idempotency_check={"status": "skipped (scope unresolved)"},
+            )
+            print(canonical_json(report))
+            return 0
+        return fail_closed(parse_err.code, parse_err.detail)
+
+    # Enumerate allowed and deny-intersection.
+    files_eligible: list[str] = []
+    files_considered: list[str] = []
+    files_would_write: list[str] = []
+    violations: list[dict[str, str]] = []
+
+    if allow_globs:
+        allowed, denied_under_allow = enumerate_scope(
+            repo_root, allow_globs, deny_patterns
+        )
+        files_considered = allowed + denied_under_allow
+        for rel in denied_under_allow:
+            violations.append({
+                "code": "CAVEMAN_COMPRESS_DENY_HIT",
+                "path": rel,
+                "stage": "pre-write",
+                "detail": "file matched allow-list but intersects deny baseline",
+            })
+        files_eligible = allowed
+
+    # Safe-mode idempotency self-check on a canonical fixture string — stable
+    # identity check for --report (AC-6 by construction).
+    fixture = "a\n\n\n\nb\n   trailing   \n"
+    once = compress_safe_mode(fixture)
+    twice = compress_safe_mode(once)
+    idempotency_ok = once == twice
+    idempotency_status: dict[str, object] = {
+        "status": "ok" if idempotency_ok else "drift",
+        "algorithm": "safe-mode-line-collapse-trim-lf",
+        "fixture_byte_stable": idempotency_ok,
+    }
+    if not idempotency_ok:
+        violations.append({
+            "code": "CAVEMAN_COMPRESS_NOT_IDEMPOTENT",
+            "path": "(self-check)",
+            "stage": "during-write",
+            "detail": "compress(compress(fixture)) != compress(fixture)",
+        })
+
+    # --write path: actual mutation under atomic sidecar-first order.
+    if mutation_mode:
+        for rel in files_eligible:
+            target = repo_root / rel
+            try:
+                original_bytes = target.read_bytes()
+            except OSError as exc:
+                violations.append({
+                    "code": "CAVEMAN_COMPRESS_DENY_HIT",
+                    "path": rel,
+                    "stage": "pre-write",
+                    "detail": f"unreadable: {exc}",
+                })
+                continue
+            try:
+                original_text = original_bytes.decode("utf-8")
+            except UnicodeDecodeError:
+                violations.append({
+                    "code": "CAVEMAN_COMPRESS_DENY_HIT",
+                    "path": rel,
+                    "stage": "pre-write",
+                    "detail": "non-utf8 bytes (binary refuse)",
+                })
+                continue
+            proposed = compress_safe_mode(original_text)
+
+            ok, why = literal_region_preserved(original_text, proposed)
+            if not ok:
+                violations.append({
+                    "code": "CAVEMAN_COMPRESS_LITERAL_REGION_DAMAGED",
+                    "path": rel,
+                    "stage": "during-write",
+                    "detail": why,
+                })
+                continue
+
+            # Idempotency re-check on real file.
+            if compress_safe_mode(proposed) != proposed:
+                violations.append({
+                    "code": "CAVEMAN_COMPRESS_NOT_IDEMPOTENT",
+                    "path": rel,
+                    "stage": "during-write",
+                    "detail": "second compression differs",
+                })
+                continue
+
+            if proposed == original_text:
+                files_would_write.append(rel)
+                continue  # no mutation required; skip sidecar churn
+
+            # Sidecar first, then target (atomic temp+replace on both).
+            sidecar = sidecar_for(repo_root, rel)
+            try:
+                atomic_write(sidecar, original_bytes)
+                atomic_write(target, proposed.encode("utf-8"))
+            except OSError as exc:
+                violations.append({
+                    "code": "CAVEMAN_COMPRESS_DENY_HIT",
+                    "path": rel,
+                    "stage": "pre-write",
+                    "detail": f"atomic write failed: {exc}",
+                })
+                continue
+            files_would_write.append(rel)
+    else:
+        # Dry-run: compute what would write, do not mutate.
+        for rel in files_eligible:
+            target = repo_root / rel
+            try:
+                original_text = target.read_text(encoding="utf-8")
+            except (OSError, UnicodeDecodeError):
+                continue
+            proposed = compress_safe_mode(original_text)
+            if proposed != original_text:
+                files_would_write.append(rel)
+
+    report = build_report(
+        deny_version=deny_version,
+        scope_resolved=allow_globs,
+        files_considered=files_considered,
+        files_eligible=files_eligible,
+        files_would_write=files_would_write,
+        violations=violations,
+        idempotency_check=idempotency_status,
+    )
+    if args.report:
+        print(canonical_json(report))
+
+    if violations:
+        first = violations[0]
+        return fail_closed(first["code"], first.get("detail", ""))
+
+    if not args.report:
+        # Default human-readable dry-run narration (one line per file).
+        if files_would_write:
+            print(f"caveman-compress dry-run: {len(files_would_write)} file(s) "
+                  f"would be rewritten")
+            for rel in files_would_write:
+                print(f"  - {rel}")
+        else:
+            print("caveman-compress dry-run: no changes")
+        print(f"deny_list_version={deny_version}")
+        if profile_name:
+            print(f"profile={profile_name}")
+
+    return 0
+
+
+def main() -> int:
+    return run(sys.argv[1:])
+
+
+if __name__ == "__main__":
+    sys.exit(main())