modbison/backend/app/services/chunk_codec.py

"""Python port of the chunk binary/JSON codec, `docs/architecture/world.md` §3.

This is the server-side half of the "single shared format, two
implementations" contract world.md §5.2 and
`world_persistence_impl.md` §5.5/§9 call for. The Godot-side implementation
is `scripts/chunk_data_decoder.gd`; this module must decode anything it
produces and produce bytes it can decode.

IMPORTANT DEVIATION (see the "Known cross-codec divergence" note at the
bottom of this file): the current `scripts/chunk_data_decoder.gd` does not
actually implement IEEE-754 float32 for the `pos`/`rot`/`scale`/`bounds`
fields -- it performs a lossy numeric `int`/`float` cast instead of a bit
reinterpretation, which both truncates fractional values and does not
write IEEE-754 bits at all. This module implements the format as
*specified* (real `<f` IEEE-754 float32, matching every other field in the
spec and matching what any conformant decoder would assume). See the note
for what this means for cross-codec byte parity today.

Nothing here talks to a database or the FastAPI app; it is a pure,
side-effect-free codec so it can be golden-file tested in isolation.
"""

from __future__ import annotations

import hashlib
import json
import struct
from dataclasses import dataclass, field

MAGIC = b"AGPC"
HEADER_SIZE = 48
CURRENT_VERSION = 1

SECTION_META = 0x01
SECTION_TILE = 0x02
SECTION_OBJECT = 0x03
SECTION_FLAGS = 0x04
SECTION_REVISION_META = 0x05

CATEGORY_FIELDS = (
    "violence", "sexual", "language", "substance",
    "gambling", "horror", "hate", "custom",
)


class ChunkCodecError(Exception):
    """Malformed bytes -- the payload isn't shaped like a chunk at all."""


class ChunkValidationError(Exception):
    """A structurally-valid chunk failed a content/business-rule check.

    ``reason`` is the machine-readable code the API surface (world_persistence_impl.md
    §8.1) puts in the 422 response body's ``reason`` field, e.g.
    ``"checksum_mismatch"``, ``"layer_dims"``, ``"object_out_of_bounds"``.
    """

    def __init__(self, reason: str, message: str | None = None) -> None:
        self.reason = reason
        super().__init__(message or reason)


# ───────────────────────── Logical model ─────────────────────────

@dataclass
class TileDef:
    id: int
    asset_id: str = ""
    solid: bool = False
    flags: int = 0
    variants: int = 1


@dataclass
class TileLayer:
    name: str
    layer_id: int = 0
    width: int = 0
    height: int = 0
    depth: int = 1
    palette: list[TileDef] = field(default_factory=list)
    tiles: list[int] = field(default_factory=list)  # flattened, row-major


@dataclass
class ContentFlags:
    violence: int = 0
    sexual: int = 0
    language: int = 0
    substance: int = 0
    gambling: int = 0
    horror: int = 0
    hate: int = 0
    custom: int = 0
    mature: bool = False
    sources: list[str] = field(default_factory=list)

    def to_dict(self) -> dict:
        return {
            "categories": {name: getattr(self, name) for name in CATEGORY_FIELDS},
            "mature": self.mature,
            "sources": list(self.sources),
        }

    @classmethod
    def from_dict(cls, d: dict) -> "ContentFlags":
        cats = d.get("categories", {})
        return cls(
            **{name: int(cats.get(name, 0)) for name in CATEGORY_FIELDS},
            mature=bool(d.get("mature", False)),
            sources=list(d.get("sources", [])),
        )


@dataclass
class WorldObject:
    id: str
    type: str
    asset_id: str
    pos: tuple[float, float, float] = (0.0, 0.0, 0.0)
    rot: tuple[float, float, float, float] = (0.0, 0.0, 0.0, 1.0)
    scale: tuple[float, float, float] = (1.0, 1.0, 1.0)
    dynamic: bool = False
    properties: dict = field(default_factory=dict)
    flags: ContentFlags = field(default_factory=ContentFlags)


@dataclass
class ChunkBounds:
    """Wire representation is min-corner + size (matching the Godot AABB
    the reference encoder builds from), not min/max as world.md §3.1's
    logical-model prose describes -- both are fully recoverable from each
    other; this is just the on-the-wire convention."""

    min: tuple[float, float, float] = (0.0, 0.0, 0.0)
    size: tuple[float, float, float] = (0.0, 0.0, 0.0)

    @property
    def max(self) -> tuple[float, float, float]:
        return tuple(a + b for a, b in zip(self.min, self.size))


@dataclass
class ChunkPayload:
    version: int = CURRENT_VERSION
    world_id: str = ""
    room_id: str = ""
    coord: tuple[int, int, int] = (0, 0, 0)
    bounds: ChunkBounds = field(default_factory=ChunkBounds)
    tile_layers: list[TileLayer] = field(default_factory=list)
    objects: list[WorldObject] = field(default_factory=list)
    flags: ContentFlags = field(default_factory=ContentFlags)
    created_at: int = 0
    updated_at: int = 0
    revision: int = 0
    author: str = ""


def build_key(world_id: str | int, room_slug: str, cx: int, cy: int, cz: int) -> str:
    """world_persistence_impl.md §4.5: ``"{world.id}:{room.slug}:{cx}:{cy}:{cz}"``."""
    return f"{world_id}:{room_slug}:{cx}:{cy}:{cz}"


# ───────────────────────── Binary primitives ─────────────────────────
# All integers little-endian, matching world.md §3.6.3. Floats are packed
# as real IEEE-754 float32 -- see the module docstring's deviation note.

def _read_u16(data: bytes, off: int) -> int:
    return struct.unpack_from("<H", data, off)[0]


def _read_u32(data: bytes, off: int) -> int:
    return struct.unpack_from("<I", data, off)[0]


def _read_i32(data: bytes, off: int) -> int:
    return struct.unpack_from("<i", data, off)[0]


def _read_i64(data: bytes, off: int) -> int:
    return struct.unpack_from("<q", data, off)[0]


def _read_f32(data: bytes, off: int) -> float:
    return struct.unpack_from("<f", data, off)[0]


def _w_u16(v: int) -> bytes:
    return struct.pack("<H", v & 0xFFFF)


def _w_u32(v: int) -> bytes:
    return struct.pack("<I", v & 0xFFFFFFFF)


def _w_i32(v: int) -> bytes:
    return struct.pack("<i", v)


def _w_i64(v: int) -> bytes:
    return struct.pack("<q", v)


def _w_f32(v: float) -> bytes:
    return struct.pack("<f", v)


def _read_str(data: bytes, off: int) -> tuple[str, int]:
    """Length-prefixed UTF-8 string. Returns (value, next_offset)."""
    n = _read_u16(data, off)
    off += 2
    s = data[off:off + n].decode("utf-8")
    return s, off + n


def _w_str(s: str) -> bytes:
    b = s.encode("utf-8")
    if len(b) > 0xFFFF:
        raise ChunkCodecError(f"string field too long to encode ({len(b)} bytes)")
    return _w_u16(len(b)) + b


def checksum(payload_body: bytes) -> str:
    """SHA-256 hex digest of the payload body (world.md §3.6.3)."""
    return hashlib.sha256(payload_body).hexdigest()


# ───────────────────────── Section codecs ─────────────────────────

def _encode_metadata(payload: ChunkPayload) -> bytes:
    out = bytearray()
    out += _w_str(payload.world_id)
    out += _w_str(payload.room_id)
    out += _w_i32(payload.coord[0])
    out += _w_i32(payload.coord[1])
    out += _w_i32(payload.coord[2])
    out += _w_f32(payload.bounds.min[0])
    out += _w_f32(payload.bounds.min[1])
    out += _w_f32(payload.bounds.min[2])
    out += _w_f32(payload.bounds.size[0])
    out += _w_f32(payload.bounds.size[1])
    out += _w_f32(payload.bounds.size[2])
    return bytes(out)


def _decode_metadata(data: bytes, result: ChunkPayload) -> None:
    off = 0
    result.world_id, off = _read_str(data, off)
    result.room_id, off = _read_str(data, off)
    coord = (_read_i32(data, off), _read_i32(data, off + 4), _read_i32(data, off + 8))
    off += 12
    min_pos = (_read_f32(data, off), _read_f32(data, off + 4), _read_f32(data, off + 8))
    size = (_read_f32(data, off + 12), _read_f32(data, off + 16), _read_f32(data, off + 20))
    result.coord = coord
    result.bounds = ChunkBounds(min=min_pos, size=size)


def _encode_tile_layer(layer: TileLayer) -> bytes:
    out = bytearray()
    out += _w_str(layer.name)
    out += _w_u16(layer.layer_id)
    out += _w_u16(layer.width)
    out += _w_u16(layer.height)
    out += _w_u16(layer.depth)
    out += _w_u16(len(layer.palette))
    for td in layer.palette:
        out += _w_u16(td.id)
        out += _w_str(td.asset_id)
        out += bytes([1 if td.solid else 0])
        out += _w_u32(td.flags)
        out += bytes([td.variants & 0xFF])
    out += _w_u32(len(layer.tiles))
    for tile_id in layer.tiles:
        out += _w_u16(tile_id)
    return bytes(out)


def _decode_tile_layer(data: bytes) -> TileLayer:
    off = 0
    name, off = _read_str(data, off)
    layer_id = _read_u16(data, off); off += 2
    width = _read_u16(data, off); off += 2
    height = _read_u16(data, off); off += 2
    depth = _read_u16(data, off); off += 2
    palette_count = _read_u16(data, off); off += 2
    palette: list[TileDef] = []
    for _ in range(palette_count):
        tid = _read_u16(data, off); off += 2
        asset_id, off = _read_str(data, off)
        solid = data[off] != 0; off += 1
        flags = _read_u32(data, off); off += 4
        variants = data[off]; off += 1
        palette.append(TileDef(id=tid, asset_id=asset_id, solid=solid, flags=flags, variants=variants))
    tile_count = _read_u32(data, off); off += 4
    tiles: list[int] = []
    for _ in range(tile_count):
        tiles.append(_read_u16(data, off))
        off += 2
    return TileLayer(name=name, layer_id=layer_id, width=width, height=height,
                      depth=depth, palette=palette, tiles=tiles)


def _encode_flags(flags: ContentFlags) -> bytes:
    out = bytearray()
    for name in CATEGORY_FIELDS:
        out.append(getattr(flags, name) & 0xFF)
    out.append(1 if flags.mature else 0)
    out += _w_u16(len(flags.sources))
    for src in flags.sources:
        out += _w_str(src)
    return bytes(out)


def _decode_flags(data: bytes) -> ContentFlags:
    off = 0
    values = {}
    for name in CATEGORY_FIELDS:
        values[name] = data[off]
        off += 1
    mature = data[off] != 0
    off += 1
    source_count = _read_u16(data, off)
    off += 2
    sources = []
    for _ in range(source_count):
        s, off = _read_str(data, off)
        sources.append(s)
    return ContentFlags(**values, mature=mature, sources=sources)


def _encode_object(obj: WorldObject) -> bytes:
    out = bytearray()
    out += _w_str(obj.id)
    out += _w_str(obj.type)
    out += _w_str(obj.asset_id)
    for v in obj.pos:
        out += _w_f32(v)
    for v in obj.rot:
        out += _w_f32(v)
    for v in obj.scale:
        out += _w_f32(v)
    out.append(1 if obj.dynamic else 0)
    # Canonical JSON: keys sorted lexicographically (world.md §3.6.3).
    props_json = json.dumps(obj.properties, sort_keys=True, separators=(",", ":"))
    out += _w_str(props_json)
    out += _encode_flags(obj.flags)
    return bytes(out)


def _decode_object(data: bytes) -> WorldObject:
    off = 0
    oid, off = _read_str(data, off)
    otype, off = _read_str(data, off)
    asset_id, off = _read_str(data, off)
    pos = (_read_f32(data, off), _read_f32(data, off + 4), _read_f32(data, off + 8))
    off += 12
    rot = (
        _read_f32(data, off), _read_f32(data, off + 4),
        _read_f32(data, off + 8), _read_f32(data, off + 12),
    )
    off += 16
    scale = (_read_f32(data, off), _read_f32(data, off + 4), _read_f32(data, off + 8))
    off += 12
    dynamic = data[off] != 0
    off += 1
    props_str, off = _read_str(data, off)
    try:
        properties = json.loads(props_str) if props_str else {}
    except json.JSONDecodeError as exc:
        raise ChunkCodecError(f"object {oid!r}: invalid properties JSON") from exc
    obj_flags = _decode_flags(data[off:])
    return WorldObject(id=oid, type=otype, asset_id=asset_id, pos=pos, rot=rot,
                        scale=scale, dynamic=dynamic, properties=properties, flags=obj_flags)


def _encode_revision_meta(payload: ChunkPayload) -> bytes:
    out = bytearray()
    out += _w_i64(payload.created_at)
    out += _w_i64(payload.updated_at)
    out += _w_u32(payload.revision)
    out += _w_str(payload.author)
    return bytes(out)


def _decode_revision_meta(data: bytes, result: ChunkPayload) -> None:
    off = 0
    result.created_at = _read_i64(data, off); off += 8
    result.updated_at = _read_i64(data, off); off += 8
    result.revision = _read_u32(data, off); off += 4
    result.author, off = _read_str(data, off)


def _wrap_section(section_id: int, data: bytes) -> bytes:
    return bytes([section_id]) + _w_u32(len(data)) + data


# ───────────────────────── Top-level encode/decode ─────────────────────────

def encode_binary(payload: ChunkPayload) -> bytes:
    """Encode a ChunkPayload to the world.md §3.6 wire format."""
    sections = bytearray()
    sections += _wrap_section(SECTION_META, _encode_metadata(payload))
    for layer in payload.tile_layers:
        sections += _wrap_section(SECTION_TILE, _encode_tile_layer(layer))
    for obj in payload.objects:
        sections += _wrap_section(SECTION_OBJECT, _encode_object(obj))
    sections += _wrap_section(SECTION_FLAGS, _encode_flags(payload.flags))
    sections += _wrap_section(SECTION_REVISION_META, _encode_revision_meta(payload))

    payload_body = bytes(sections)
    digest = hashlib.sha256(payload_body).digest()  # 32 raw bytes

    header = bytearray()
    header += MAGIC
    header += _w_u16(CURRENT_VERSION)
    header += _w_u16(HEADER_SIZE - 8)  # header_len: bytes of header after this field (spec-correct)
    header += _w_u32(0)  # flags (reserved)
    header += _w_u32(len(payload_body))
    header += digest

    return bytes(header) + payload_body


def decode_binary(data: bytes) -> ChunkPayload:
    """Decode a world.md §3.6 binary chunk. Raises ChunkCodecError on
    structurally malformed input (short buffer, bad magic, unsupported
    version, truncated payload)."""
    if len(data) < HEADER_SIZE:
        raise ChunkCodecError(f"data too short for header ({len(data)} bytes)")

    if data[0:4] != MAGIC:
        raise ChunkCodecError("bad magic bytes")

    version = _read_u16(data, 4)
    if version < 1 or version > CURRENT_VERSION:
        raise ChunkCodecError(f"unsupported chunk format version {version}")

    # header_len (offset 6) is intentionally not used for offset math --
    # both this decoder and the Godot one always slice at the fixed
    # HEADER_SIZE, matching the reference implementation's actual (if
    # spec-inconsistent) behaviour.
    payload_len = _read_u32(data, 12)
    # bytes 16:48 are the header's advisory checksum -- decode_and_validate
    # re-reads it directly and recomputes against the parsed payload body.

    if len(data) - HEADER_SIZE != payload_len:
        raise ChunkCodecError(
            f"declared payload_len ({payload_len}) does not match actual "
            f"remaining bytes ({len(data) - HEADER_SIZE})"
        )

    payload_body = data[HEADER_SIZE:HEADER_SIZE + payload_len]

    result = ChunkPayload(version=version)

    offset = 0
    while offset < len(payload_body):
        if offset + 5 > len(payload_body):
            raise ChunkCodecError("truncated section header")
        section_id = payload_body[offset]
        section_len = _read_u32(payload_body, offset + 1)
        data_start = offset + 5
        data_end = data_start + section_len
        if data_end > len(payload_body):
            raise ChunkCodecError("truncated section body")
        section_data = payload_body[data_start:data_end]

        if section_id == SECTION_META:
            _decode_metadata(section_data, result)
        elif section_id == SECTION_TILE:
            result.tile_layers.append(_decode_tile_layer(section_data))
        elif section_id == SECTION_OBJECT:
            result.objects.append(_decode_object(section_data))
        elif section_id == SECTION_FLAGS:
            result.flags = _decode_flags(section_data)
        elif section_id == SECTION_REVISION_META:
            _decode_revision_meta(section_data, result)
        # Unknown section ids are skipped (forward-compat), matching the
        # Godot decoder's match-with-no-default behaviour.

        offset = data_end

    return result


# ───────────────────────── JSON encoding (world.md §3.5) ─────────────────────────
# Derived, never stored (world_persistence_impl.md §8.1) -- produced here
# from a decoded ChunkPayload for the ``Accept: application/json`` path.

def to_json_dict(payload: ChunkPayload, *, key: str, chunk_checksum: str) -> dict:
    return {
        "version": payload.version,
        "key": key,
        "world_id": payload.world_id,
        "room_id": payload.room_id,
        "coord": {"cx": payload.coord[0], "cy": payload.coord[1], "cz": payload.coord[2]},
        "bounds": {
            "min_x": payload.bounds.min[0], "min_y": payload.bounds.min[1], "min_z": payload.bounds.min[2],
            "max_x": payload.bounds.max[0], "max_y": payload.bounds.max[1], "max_z": payload.bounds.max[2],
        },
        "tiles": [
            {
                "name": layer.name,
                "layer_id": layer.layer_id,
                "width": layer.width,
                "height": layer.height,
                "depth": layer.depth,
                "palette": [
                    {"id": td.id, "asset_id": td.asset_id, "solid": td.solid,
                     "flags": td.flags, "variants": td.variants}
                    for td in layer.palette
                ],
                "tiles": list(layer.tiles),
            }
            for layer in payload.tile_layers
        ],
        "objects": [
            {
                "id": obj.id, "type": obj.type, "asset_id": obj.asset_id,
                "transform": {
                    "pos": list(obj.pos), "rot": list(obj.rot), "scale": list(obj.scale),
                },
                "properties": obj.properties,
                "dynamic": obj.dynamic,
                "flags": obj.flags.to_dict(),
            }
            for obj in payload.objects
        ],
        "flags": payload.flags.to_dict(),
        "meta": {
            "created_at": payload.created_at, "updated_at": payload.updated_at,
            "revision": payload.revision, "author": payload.author,
        },
        "checksum": chunk_checksum,
    }


# ───────────────────────── Validation (world_persistence_impl.md §5.5) ─────────────────────────

# Not precisely quantified by the design doc beyond "length limits" --
# these are this implementation's conservative defaults (world.md §7).
MAX_STRING_FIELD_LEN = 256
MAX_PROPERTIES_JSON_LEN = 16_384
MAX_SOURCES_PER_FLAGSET = 32


def _check_string_lengths(obj: WorldObject) -> None:
    for field_name, value in (("id", obj.id), ("type", obj.type), ("asset_id", obj.asset_id)):
        if len(value) > MAX_STRING_FIELD_LEN:
            raise ChunkValidationError(
                "string_field_too_long", f"object.{field_name} exceeds {MAX_STRING_FIELD_LEN} chars"
            )
    props_json = json.dumps(obj.properties)
    if len(props_json) > MAX_PROPERTIES_JSON_LEN:
        raise ChunkValidationError(
            "string_field_too_long", f"object {obj.id!r} properties exceed {MAX_PROPERTIES_JSON_LEN} bytes"
        )
    if len(obj.flags.sources) > MAX_SOURCES_PER_FLAGSET:
        raise ChunkValidationError("string_field_too_long", f"object {obj.id!r} has too many flag sources")


def decode_and_validate(
    data: bytes,
    *,
    expected_key: str,
    grid_size: tuple[int, int, int],
    max_objects: int,
) -> ChunkPayload:
    """Decode + fully validate a client-submitted chunk payload.

    This is the boundary world_persistence_impl.md §5.5 requires: "the
    backend must not trust the client". Raises ``ChunkCodecError`` for
    structurally malformed bytes (mapped to a 422 by the router with a
    generic reason) or ``ChunkValidationError`` with a specific ``reason``
    code for a well-formed-but-invalid payload.
    """
    payload = decode_binary(data)  # raises ChunkCodecError for structural problems first

    # Checksum is advisory in the header (world.md §7); the value that
    # matters is recomputed from the actual payload body we just parsed.
    payload_body = data[HEADER_SIZE:]
    recomputed = checksum(payload_body)
    header_checksum_hex = data[16:48].hex()
    if header_checksum_hex != recomputed:
        raise ChunkValidationError(
            "checksum_mismatch",
            "header checksum does not match the recomputed SHA-256 of the payload body",
        )

    actual_key = build_key(payload.world_id, payload.room_id, *payload.coord)
    if actual_key != expected_key:
        raise ChunkValidationError(
            "chunk_key_mismatch",
            f"payload key {actual_key!r} does not match URL {expected_key!r}",
        )

    for layer in payload.tile_layers:
        if (layer.width, layer.height, layer.depth) != tuple(grid_size):
            raise ChunkValidationError(
                "layer_dims",
                f"tile layer {layer.name!r} dims {(layer.width, layer.height, layer.depth)} "
                f"!= room grid {tuple(grid_size)}",
            )
        valid_ids = {0} | {td.id for td in layer.palette}
        for tile_id in layer.tiles:
            if tile_id not in valid_ids:
                raise ChunkValidationError(
                    "tile_out_of_palette",
                    f"tile layer {layer.name!r} references palette id {tile_id} not in its palette",
                )
        if len(layer.tiles) != layer.width * layer.height * layer.depth:
            raise ChunkValidationError(
                "layer_dims",
                f"tile layer {layer.name!r} has {len(layer.tiles)} cells, expected "
                f"{layer.width * layer.height * layer.depth}",
            )

    if len(payload.objects) > max_objects:
        raise ChunkValidationError(
            "too_many_objects", f"{len(payload.objects)} objects exceeds max {max_objects}"
        )

    bmin, bmax = payload.bounds.min, payload.bounds.max
    for obj in payload.objects:
        _check_string_lengths(obj)
        if not all(bmin[i] <= obj.pos[i] <= bmax[i] for i in range(3)):
            raise ChunkValidationError(
                "object_out_of_bounds",
                f"object {obj.id!r} position {obj.pos} outside chunk bounds "
                f"[{bmin}, {bmax}]",
            )

    if len(payload.author) > MAX_STRING_FIELD_LEN:
        raise ChunkValidationError("string_field_too_long", "meta.author too long")

    return payload


# ───────────────────────── Known cross-codec divergence ─────────────────────────
#
# `scripts/chunk_data_decoder.gd`'s `_read_float32`/`_write_float32` do:
#
#     static func _write_float32(val: float) -> PackedByteArray:
#         var bits := val as int
#         return _write_uint32(bits)
#
#     static func _read_float32(data: PackedByteArray, offset: int) -> float:
#         var bits := _read_uint32(data, offset)
#         return bits as float
#
# `as` between int and float in GDScript is a numeric conversion, not a bit
# reinterpretation (verified empirically against Godot 4.6.2 headless:
# `1.5 as int` -> `1`, `1 as float` -> `1.0`; the IEEE-754 bit pattern of
# 1.5f is 0x3FC00000 = 1069547520, not 1). So the shipped client codec
# truncates every float32 field (bounds, object pos/rot/scale) to its
# integer part and stores that integer as a raw int32 -- not as float32
# bits at all. This is a real bug in already-merged, client-side code
# (scripts/ is out of scope for this task) that breaks the "byte-stable
# across two implementations" guarantee world.md §3.6.3 states.
#
# This module implements the *specified* format (real IEEE-754 float32)
# rather than replicating that bug, because world.md §3.6 is normative and
# says "float32" for these fields, and mirroring the bug would mean the
# backend also corrupts every non-integer transform/bounds value it
# stores. Practical effect until the client-side bug is fixed: payloads
# with integer-only floats will decode identically on both sides; any
# payload with a fractional position/rotation/scale/bounds value will not
# be byte-identical (or numerically identical) between this module and the
# current `scripts/chunk_data_decoder.gd`. This is flagged for the
# client/editor follow-on task, not fixed here.