-- Voxel world mode: turn a map's tile layer into one static 3D mesh.
--
-- The scene description comes from Structures.lua, which -- 3dSen-style --
-- detects each connected drawn thing on the map and picks its model:
--
--   flat      ground / water / void: a single quad.
--   top art   ledges, roofs (profile-authored): a box with the art on its
--             TOP face; partial side bands crop the art (a 6px ledge face
--             is the bottom of the lip drawing).
--   volume    walls, buildings, tree lines: each column rises to the
--             structure's REAL drawn height (Structures measures it,
--             repeat-aware and region-consistent -- a 6-row house is 48px,
--             a 40-row border forest is rows of 16px trees). The south
--             face folds the full artwork upright, 8px band by band, band
--             k sampling the map row k tiles north; the top wears the
--             structure's top rows.
--   object    small props with a silhouette (plants, signs, lone trees):
--             per-pixel voxel prisms prebuilt by Structures, standing on
--             synthesized ground -- this mesher just emits their quads.
--             Round trees arrive as STAMPS (a shared hull template plus a
--             cell offset) and expand here, straight into the vertex
--             stream, so no map retains per-cell copies of its forests.
--
-- Side faces are never stretched: all sides are 8px bands with the art
-- tiled per band and cropped at partial bands.
--
-- Texturing samples the TILESET ATLAS, not a rendered copy of the map. The
-- atlas is 128x48; a map-space canvas covering the biggest routes would be
-- ~5 MB each with up to five live at once (connected maps), which is real
-- memory on the mobile targets. Sampling the atlas costs 24 KB, and costs
-- nothing in fidelity because TerrainAtlas hands back the same atlas
-- TileRenderer draws with -- including the fully recolored one RED++
-- bakes -- so terrain color comes through untouched.
--
-- BUILDS ARE ASYNCHRONOUS. A frame never blocks on meshing: VoxelScene
-- requests what it wants to draw, request() queues a build job, and
-- pump() -- called once a frame from the pipeline's update -- advances
-- the queue inside a few-millisecond budget (BuildBudget suspends the
-- job's coroutine mid-loop when the slice is spent). Until a mesh lands
-- the scene simply draws without it: the engine's flat path while the
-- current map has nothing, the body-only variant while the full one (the
-- border ring) is still cooking, neighbours popping in as they finish.
-- The synchronous get() remains for probes and tests.
--
-- Meshes are cached per map id and EVICTED down to the live set (current
-- map + connected neighbours) whenever that set changes -- setLive()
-- releases far maps' GPU meshes and their Structures analysis, which is
-- what used to grow the heap by gigabytes over a cross-region trek.

-- the mod namespace (see main.lua): V.require loads a sibling module
local V = ...

local Assets = require("src.render.Assets")
local Structures = V.require("Structures")
local TileShape = V.require("TileShape")
local Voxel3D = V.require("Voxel3D")
local Budget = V.require("BuildBudget")

local ffi = nil
do
  local ok, mod = pcall(require, "ffi")
  if ok then ffi = mod end
end

local ChunkMesher = {}

-- Ring of border blocks meshed around the body, matching the width
-- TileRenderer draws so the two modes end at the same place.
local RING = 3

-- A sliver of a texel, to keep a quad's sampling inside its own tile.
-- Without any inset the perspective rasteriser lands on a NEIGHBOURING
-- tile's texel along the shared edge and stitches bright seams across the
-- whole map.
--
-- It has to be a sliver and not, as it first was, half a texel. A tile is
-- 8 texels of art across 8 world pixels -- one texel per pixel exactly --
-- and insetting the uv by half a texel at each end squeezes that art into
-- a 7-texel sample range while the quad still covers 8 world pixels. The
-- art then advances 7/8 of a texel per pixel: boundaries drift off the
-- pixel grid, one art pixel gets sampled twice and another never at all.
-- Nothing showed it until the voxel wireframe drew the grid those pixels
-- were supposed to be sitting on. Interpolation error is nowhere near a
-- fiftieth of a texel, so this is as safe against bleed and costs 0.25% of
-- a pixel of drift across a whole tile.
local INSET = 0.02

-- The south face of a volume is the artwork itself, so it draws at full
-- brightness; its top face darkens a touch so the plateau behind a
-- standing drawing reads as depth rather than repeating the same art at
-- the same energy.
local VOLUME_TOP_SHADE = 0.85

local cache = {}     -- map id -> { full = mesh|false, body = ..., grass = ... }
local gen = {}       -- map id -> generation, bumped by invalidate/evict

-- Horizontal neighbours: tile step, face direction id (see Voxel3D).
local SIDES = {
  { 1, 0, 1 },    -- +X east
  { -1, 0, 2 },   -- -X west
  { 0, 1, 5 },    -- +Z south
  { 0, -1, 6 },   -- -Z north
}

local function keyOf(tx, ty)
  return (ty + 64) * 4096 + (tx + 64)
end

-- ------------------------------------------------------------ vertex sinks

-- A sink accepts quads (4 corners, 4 uv pairs, flat or per-corner shade)
-- and finishes into a drawable mesh. The TABLE sink reproduces the
-- historical pure-Lua output -- geometry() returns its arrays for the
-- headless suite. The FFI sink packs the same six floats per vertex
-- straight into one growing native buffer, unindexed (v1 v2 v3 v1 v3 v4),
-- skipping ~a million short-lived Lua tables per route and LOVE's slow
-- table-by-table vertex upload.

local function newTableSink()
  local verts, indices, quads = {}, {}, 0
  return {
    push = function(c, uv, shade)
      local flat = type(shade) ~= "table"
      for i = 1, 4 do
        local cc, t = c[i], uv[i]
        verts[#verts + 1] = { cc[1], cc[2], cc[3], t[1], t[2],
                              flat and shade or shade[i] }
      end
      Voxel3D.pushQuad(indices, quads)
      quads = quads + 1
    end,
    results = function()
      return verts, indices, quads
    end,
    finish = function()
      return Voxel3D.newMesh(verts, indices)
    end,
  }
end

local TRI_ORDER = { 1, 2, 3, 1, 3, 4 }

local function newFfiSink()
  local cap = 4096 * 6
  local buf = ffi.new("float[?]", cap * 6)
  local n = 0
  local sink
  sink = {
    push = function(c, uv, shade)
      if n + 6 > cap then
        local grown = ffi.new("float[?]", cap * 2 * 6)
        ffi.copy(grown, buf, n * 6 * 4)
        buf, cap = grown, cap * 2
      end
      local flat = type(shade) ~= "table"
      local base = n * 6
      for k = 1, 6 do
        local i = TRI_ORDER[k]
        local cc, t = c[i], uv[i]
        buf[base] = cc[1]
        buf[base + 1] = cc[2]
        buf[base + 2] = cc[3]
        buf[base + 3] = t[1]
        buf[base + 4] = t[2]
        buf[base + 5] = flat and shade or shade[i]
        base = base + 6
      end
      n = n + 6
    end,
    finish = function()
      if n == 0 then return nil end
      -- upload in slices with budget ticks between: a route-sized mesh
      -- is ~10-20MB and one atomic setVertices was the last remaining
      -- frame spike. The mesh is not cached (so never drawn) until the
      -- whole upload lands, and LuaJIT yields fine across pcall.
      local ok, mesh = pcall(function()
        local m = love.graphics.newMesh(Voxel3D.FORMAT, n,
                                        "triangles", "static")
        local CHUNK = 65536              -- vertices per slice (~1.5MB)
        local i = 0
        while i < n do
          local count = math.min(CHUNK, n - i)
          local bytes = count * 6 * 4
          local data = love.data.newByteData(bytes)
          ffi.copy(data:getFFIPointer(), buf + i * 6, bytes)
          m:setVertices(data, i + 1)
          data:release()
          i = i + count
          Budget.check()
        end
        return m
      end)
      return ok and mesh or nil
    end,
  }
  return sink
end

local function newSink()
  if ffi and love and love.data and love.data.newByteData
     and love.graphics and love.graphics.newMesh then
    return newFfiSink()
  end
  return newTableSink()
end

-- -------------------------------------------------------------- geometry

-- Emit the raw geometry for `map` into `sink`. `bodyOnly` skips the
-- border ring -- the shape the 2D path's drawMapOnly has always had: a
-- neighbour map contributes its body, and only the CURRENT map supplies
-- the ring around the view.
--
-- `masks` (full variant only) lists rectangles, in this map's world
-- pixels, where connected neighbour BODIES sit: ring geometry inside them
-- is suppressed. The 2D renderer never needed this because it painted
-- neighbour bodies OVER the ring; with a depth buffer the ring's standing
-- trees would rise straight through the neighbour's flat ground -- cross
-- into Route 1 and a wall of border trees sprouts over Pallet.
--
-- Kept free of any GPU call so it can be exercised headless -- the
-- geometry is the part with the interesting invariants, and a suite that
-- needed a real GL context to check them would never run in CI.
local function runGeometry(map, bodyOnly, masks, sink)
  local push = sink.push
  local tileset = map.tileset
  local S = Structures.forMap(map)
  local perRow = tileset.tilesPerRow or 16
  local atlasW = tileset.imageWidth or (perRow * 8)
  local atlasH = tileset.imageHeight or 48

  local function heightAt(tx, ty)
    local k = keyOf(tx, ty)
    if S.skip[k] then return 0 end
    local run = S.runs[k]
    if run then return run.h end
    local s = S.shapeAt[k]
    return s and s.h or 0
  end

  -- one atlas-rect UV, optionally cropped to art rows [vTop, vBot] of 8
  local function uvRect(tile, vTop, vBot)
    local ax = (tile % perRow) * 8
    local ay = math.floor(tile / perRow) * 8
    local vi = math.min(INSET, (vBot - vTop) / 4)
    return (ax + INSET) / atlasW, (ax + 8 - INSET) / atlasW,
           (ay + vTop + vi) / atlasH, (ay + vBot - vi) / atlasH
  end

  -- ------------------------------------------------------ ambient occlusion
  --
  -- Ambient light is what reaches a surface from the sky at large, so it is
  -- blocked by how much geometry crowds a point rather than by where the
  -- sun happens to be -- which makes it the exact complement of the shadow
  -- pass, and the reason both are worth having. The shadow map draws the
  -- long directional shadow a building throws; this draws the dark seam in
  -- every corner the sky cannot see into, at every scale finer than a
  -- shadow map texel.
  --
  -- Baked per vertex, the classic voxel way: each corner counts the
  -- neighbours that crowd it and steps down once per neighbour, and the
  -- rasteriser interpolates the steps into a smooth falloff. Costs exactly
  -- nothing at draw time, and it is resolution-independent -- a screen
  -- space pass would blur across the pixel grid this whole mode is built
  -- to keep crisp.
  --
  -- (What was here before was a one-directional contact shadow keyed to a
  -- sun in the northwest: two neighbours, one corner, top faces only.)

  -- Intensity. Both terms below are DARKENING amounts rather than
  -- multipliers, so this one number scales the whole effect: 1.0 is the
  -- barely-there first cut, and everything is expressed against it.
  local AO_STRENGTH = 2.4
  local AO_STEP = 0.09 * AO_STRENGTH      -- per crowding neighbour, max 3
  local AO_EDGE = 1 - 0.14 * AO_STRENGTH  -- creases / corners on a face
  local AO_GROUND = 0.12 * AO_STRENGTH    -- a prop's contact with the floor
  local AO_RISE = 6                       -- px over which the floor lets go
  local AO_FLOOR = 0.25                   -- never let a vertex reach black

  -- Both sinks copy a per-corner shade straight out into the vertex stream
  -- and keep no reference, so these two scratch rows are reused for every
  -- quad on the map rather than allocating a table per face -- a route
  -- builds a few hundred thousand of them.
  local aoTop = { 0, 0, 0, 0 }
  local aoSide = { 0, 0, 0, 0 }

  -- A top face's four corners, each occluded by the three cells that touch
  -- it: two edge neighbours and the diagonal between them.
  local function aoShades(tx, ty, h, shade)
    local n = heightAt(tx, ty - 1) > h
    local s = heightAt(tx, ty + 1) > h
    local e = heightAt(tx + 1, ty) > h
    local w = heightAt(tx - 1, ty) > h
    local nw = heightAt(tx - 1, ty - 1) > h
    local ne = heightAt(tx + 1, ty - 1) > h
    local sw = heightAt(tx - 1, ty + 1) > h
    local se = heightAt(tx + 1, ty + 1) > h
    if not (n or s or e or w or nw or ne or sw or se) then return shade end
    local function corner(a, b, d)
      local k = 0
      if a then k = k + 1 end
      if b then k = k + 1 end
      -- a diagonal wedged behind both of its edges adds nothing: the
      -- corner is already as enclosed as it can get, and counting it
      -- again is what turns an ordinary inside corner black
      if d and not (a and b) then k = k + 1 end
      -- floored, so cranking AO_STRENGTH deepens the seams instead of
      -- punching holes of pure black through the world
      return shade * math.max(AO_FLOOR, 1 - AO_STEP * k)
    end
    -- corners in topQuad order: NW, NE, SE, SW
    aoTop[1], aoTop[2] = corner(n, w, nw), corner(n, e, ne)
    aoTop[3], aoTop[4] = corner(s, e, se), corner(s, w, sw)
    return aoTop
  end

  -- The same idea on an upright face, where the crowding is of two kinds:
  -- the CREASE it rises out of (the band sitting on the ground, or on
  -- whatever lower neighbour exposed the face) and the INSIDE CORNERS
  -- where the columns flanking it stand proud of the band. `hl`/`hr` are
  -- those flanking heights in FACE order -- left then right as seen from
  -- outside, per LATERAL below -- so the shades line up with sideQuad's
  -- corners without the caller thinking about compass directions.
  local LATERAL = {
    [1] = { 0, 1, 0, -1 },    -- east face:  left south, right north
    [2] = { 0, -1, 0, 1 },    -- west face:  left north, right south
    [5] = { -1, 0, 1, 0 },    -- south face: left west,  right east
    [6] = { 1, 0, -1, 0 },    -- north face: left east,  right west
  }
  -- Ground contact for the prebuilt prop quads -- the per-pixel plants,
  -- signs and lone trees, and the round-tree stamps. Those arrive from
  -- Structures already finished, so the neighbour counting above has no
  -- columns to count. What it CAN say is that the ground plane itself
  -- blocks half the sky, so the closer a voxel sits to it the less ambient
  -- light reaches it -- which is what plants a prop on the floor instead
  -- of leaving it looking pasted over the top.
  local aoProp = { 0, 0, 0, 0 }
  local function groundShades(c, shade)
    if type(shade) == "table" then return shade end
    local y1, y2, y3, y4 = c[1][2], c[2][2], c[3][2], c[4][2]
    if math.min(y1, y2, y3, y4) >= AO_RISE then return shade end
    for i = 1, 4 do
      local t = c[i][2] / AO_RISE
      aoProp[i] = shade * (t >= 1 and 1 or (1 - AO_GROUND * (1 - t)))
    end
    return aoProp
  end

  local AO_CORNER = math.max(AO_FLOOR, AO_EDGE * AO_EDGE)  -- crease AND flank
  local function sideShades(hl, hr, y0, y1, crease, shade)
    if not (crease or hl > y0 or hr > y0) then return shade end
    -- corners run bottom-left, bottom-right, top-right, top-left
    local base = crease and AO_EDGE or 1
    aoSide[1] = shade * (hl > y0 and (crease and AO_CORNER or AO_EDGE) or base)
    aoSide[2] = shade * (hr > y0 and (crease and AO_CORNER or AO_EDGE) or base)
    aoSide[3] = shade * (hr > y1 and AO_EDGE or 1)
    aoSide[4] = shade * (hl > y1 and AO_EDGE or 1)
    return aoSide
  end

  local function topQuad(x0, z0, h, tile, shade)
    local u0, u1, v0, v1 = uvRect(tile, 0, 8)
    push({ { x0, h, z0 }, { x0 + 8, h, z0 },
           { x0 + 8, h, z0 + 8 }, { x0, h, z0 + 8 } },
         { { u0, v0 }, { u1, v0 }, { u1, v1 }, { u0, v1 } },
         aoShades(x0 / 8, z0 / 8, h, shade))
  end

  -- vertical quad for face direction `d` of the tile column at (x0, z0),
  -- spanning heights [y0, y1] and showing art rows [vTop, vBot] of `tile`.
  -- Corners run bottom-left, bottom-right, top-right, top-left as seen
  -- from outside; u follows +X on the north/south faces so a door or sign
  -- never draws mirrored.
  local function sideQuad(d, x0, z0, y0, y1, tile, vTop, vBot, shade)
    local x1, z1 = x0 + 8, z0 + 8
    local c
    if d == 5 then                                       -- south, at z1
      c = { { x0, y0, z1 }, { x1, y0, z1 }, { x1, y1, z1 }, { x0, y1, z1 } }
    elseif d == 6 then                                   -- north, at z0
      c = { { x1, y0, z0 }, { x0, y0, z0 }, { x0, y1, z0 }, { x1, y1, z0 } }
    elseif d == 1 then                                   -- east, at x1
      c = { { x1, y0, z1 }, { x1, y0, z0 }, { x1, y1, z0 }, { x1, y1, z1 } }
    else                                                 -- west, at x0
      c = { { x0, y0, z0 }, { x0, y0, z1 }, { x0, y1, z1 }, { x0, y1, z0 } }
    end
    local u0, u1, v0, v1 = uvRect(tile, vTop, vBot)
    push(c, { { u0, v1 }, { u1, v1 }, { u1, v0 }, { u0, v0 } }, shade)
  end

  local def = map.def
  local tw, th = def.width * 4, def.height * 4         -- map size in tiles
  local r = bodyOnly and 0 or RING * 4

  -- true when the (ring) position lies under a connected neighbour's body
  local function masked(px0, pz0, px1, pz1)
    if not masks then return false end
    for _, mk in ipairs(masks) do
      if px1 > mk[1] and px0 < mk[3] and pz1 > mk[2] and pz0 < mk[4] then
        return true
      end
    end
    return false
  end

  -- The inclusive variant for OBJECT quads: a quad TOUCHING a neighbour
  -- body counts as under it. The old test took the quad's center with
  -- strict bounds, and a quad whose center sat exactly on the body's
  -- edge line escaped the mask -- stringing stray pixel fragments of
  -- otherwise-dropped border trees along every map seam.
  local function maskedClosed(px0, pz0, px1, pz1)
    if not masks then return false end
    for _, mk in ipairs(masks) do
      if px1 >= mk[1] and px0 <= mk[3] and pz1 >= mk[2] and pz0 <= mk[4] then
        return true
      end
    end
    return false
  end

  for ty = -r, th + r - 1 do
    for tx = -r, tw + r - 1 do
      Budget.tick()
      local k = keyOf(tx, ty)
      local s, tile = S.shapeAt[k], S.tileAt[k]
      local inBody = tx >= 0 and ty >= 0 and tx < tw and ty < th
      if not inBody and masked(tx * 8, ty * 8, tx * 8 + 8, ty * 8 + 8) then
        s = nil
      end

      if s and S.skip[k] then
        -- an object stands here; paint its synthesized ground and let the
        -- prebuilt prism quads (appended below) carry the art
        if S.ground[k] then topQuad(tx * 8, ty * 8, 0, S.ground[k], 1) end
      elseif s then
        local run = S.runs[k]
        local h = run and run.h or s.h
        local x0, z0 = tx * 8, ty * 8

        -- top face. A roofed volume gets a GABLE segment: the roof rises
        -- from the facade top at the south eave to a ridge across the
        -- footprint's middle, then falls back to the facade at the north
        -- edge -- so the far side sits LOW. (The first cut was a shed
        -- plane rising all the way north, which turns a building into a
        -- ramp.) The south slope wears the structure's roof rows (ridge
        -- art at the ridge, eaves art at the eave); the back slope
        -- mirrors them. Exposed east/west flanks hip: their outer edge
        -- drops toward the eave, rounding the drawn corner tiles into 45
        -- degree corners. Flat-topped volumes wear their top rows;
        -- everything else its own art.
        if run and run.rise > 0 then
          local mid = run.extent / 2
          local function gableH(d)     -- d = rows north of the south eave
            local t = d <= mid and d / mid or (run.extent - d) / (run.extent - mid)
            return run.h + run.rise * math.max(0, math.min(1, t))
          end
          local d0 = run.front - ty                -- rows from the south edge
          local hS = gableH(d0)
          local hN = gableH(d0 + 1)
          -- art by proximity to the ridge, mirrored over the back
          local rel = 1 - math.abs(d0 + 0.5 - mid) / math.max(mid, 0.5)
          local idx = math.min(run.roofRows - 1,
                               math.floor((1 - rel) * run.roofRows))
          local roofTile = map:tileAt(tx, run.north + idx)
          local swY, seY, neY, nwY = hS, hS, hN, hN
          if heightAt(tx - 1, ty) < run.h then     -- west flank: hip
            swY = math.max(run.h, hS - 8)
            nwY = math.max(run.h, hN - 8)
          end
          if heightAt(tx + 1, ty) < run.h then     -- east flank: hip
            seY = math.max(run.h, hS - 8)
            neY = math.max(run.h, hN - 8)
          end
          local u0, u1, v0, v1 = uvRect(roofTile, 0, 8)
          push({ { x0, swY, z0 + 8 }, { x0 + 8, seY, z0 + 8 },
                 { x0 + 8, neY, z0 }, { x0, nwY, z0 } },
               { { u0, v1 }, { u1, v1 }, { u1, v0 }, { u0, v0 } }, 0.95)
        elseif run then
          local m = math.min(2, run.extent)
          local topTile = map:tileAt(tx, run.north + ((ty - run.north) % m))
          topQuad(x0, z0, h, topTile, VOLUME_TOP_SHADE)
        else
          local topTile = tile
          if s.art == "upright" and s.authored then
            -- Top art for a pinned box.  A furniture drawing is top-view
            -- rows over floor(h/8) face-on rows the fold stands upright;
            -- a face row's top would repeat its front art lying flat, so
            -- it wears the nearest row above the face block instead --
            -- the drawn tabletop (and whatever sits on it) stays on top,
            -- and a fully-folded structure (wall, desk) tops with its
            -- northmost row.
            local north, front = ty, ty
            while ty - north < 6 do
              local bs = S.shapeAt[keyOf(tx, north - 1)]
              if bs and bs.authored and bs.class == s.class then
                north = north - 1
              else
                break
              end
            end
            while front - ty < 6 do
              local bs = S.shapeAt[keyOf(tx, front + 1)]
              if bs and bs.authored and bs.class == s.class then
                front = front + 1
              else
                break
              end
            end
            local row = math.min(ty, front - math.floor(h / 8))
            if row < north then
              -- the whole run folded onto the face: top with the drawn
              -- row just above it when that row is furniture too (a
              -- bookcase wearing its shelf-top trim), else with the
              -- run's own top row
              local above = S.shapeAt[keyOf(tx, north - 1)]
              row = (above and above.authored and above.art == "upright")
                    and (north - 1) or north
            end
            topTile = S.tileAt[keyOf(tx, row)]
          end
          topQuad(x0, z0, h, topTile,
                  s.art == "upright" and VOLUME_TOP_SHADE or 1)
        end

        -- sides: 8px bands wherever the neighbour is lower. Band k spans
        -- heights [8k, 8k+8) and shows one full tile of art; a partial
        -- band crops the art rows to match, so nothing ever stretches.
        for _, side in ipairs(SIDES) do
          local nh = heightAt(tx + side[1], ty + side[2])
          if nh < h then
            local d = side[3]
            -- the columns flanking this face, for the inside-corner term:
            -- fixed for the whole face, so they are read once rather than
            -- once per 8px band
            local lat = LATERAL[d]
            local hl = lat and heightAt(tx + lat[1], ty + lat[2]) or 0
            local hr = lat and heightAt(tx + lat[3], ty + lat[4]) or 0
            for band = math.floor(nh / 8), math.ceil(h / 8) - 1 do
              local y0 = math.max(nh, band * 8)
              local y1 = math.min(h, band * 8 + 8)
              if y1 > y0 then
                local src, shade = tile, Voxel3D.FACE_SHADE[d]
                if run then
                  -- fold the structure's artwork up this face: band k
                  -- samples the map row k tiles north of the structure's
                  -- front, clamped to its extent. The south face is the
                  -- drawing itself (full brightness); the other sides wear
                  -- the same rows darkened, so a building's flank matches
                  -- its face instead of smearing one tile
                  if d == 6 then
                    src = map:tileAt(tx, math.min(run.front,
                                                  run.north + band))
                  else
                    src = map:tileAt(tx, math.max(run.north,
                                                  run.front - band))
                  end
                  if d == 5 then shade = 1 end
                elseif s.art == "upright" then
                  -- profile-authored upright (a pinned wall or furniture
                  -- box): fold the drawing up the face, band 0 the
                  -- structure's southmost same-class row and higher bands
                  -- the rows north of it, repeating past the top.  The
                  -- south face is the drawing itself (full brightness);
                  -- flanks and back wear the same front stack darkened, so
                  -- a desk's side matches its face instead of smearing a
                  -- different jumble per row.
                  if d == 5 then shade = 1 end
                  local front = ty
                  while front < ty + 6 do
                    local fs2 = S.shapeAt[keyOf(tx, front + 1)]
                    if fs2 and fs2.authored and fs2.class == s.class then
                      front = front + 1
                    else
                      break
                    end
                  end
                  local fk = keyOf(tx, front - band)
                  local fs = S.shapeAt[fk]
                  if fs and fs.authored and fs.class == s.class then
                    src = S.tileAt[fk]
                  end
                end
                sideQuad(d, x0, z0, y0, y1, src,
                         (band * 8 + 8) - y1, (band * 8 + 8) - y0,
                         sideShades(hl, hr, y0, y1, y0 <= nh, shade))
              end
            end
          end
        end
      end
    end
  end

  -- Prebuilt quads from Structures (per-pixel voxel props, lathed
  -- columns) plus the round-tree stamps expanded in place. Keep rules,
  -- by the quad's own extent:
  --   body-only   the quad must overlap the OPEN body interval -- a
  --               neighbour's ring props must not march past its edge
  --               into this map, and a quad lying exactly ON the edge
  --               plane would z-fight the map that owns that plane.
  --   full        anything overlapping the body stays whole (props that
  --               straddle the edge no longer shed their outer half);
  --               pure ring quads drop when they touch a neighbour body
  --               (maskedClosed), which is what strings of seam pixels
  --               were: fragments of dropped border trees whose centers
  --               sat exactly on the boundary line.
  local bw, bh = tw * 8, th * 8
  local function keepQuad(x0, z0, x1, z1)
    local overBody = x1 > 0 and x0 < bw and z1 > 0 and z0 < bh
    if bodyOnly then return overBody end
    return overBody or not maskedClosed(x0, z0, x1, z1)
  end

  local scUV = { { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 } }
  local function quadUV(q)
    if q.uv then return q.uv end
    for i = 1, 4 do
      scUV[i][1], scUV[i][2] = q.u, q.v
    end
    return scUV
  end

  for _, q in ipairs(S.objectQuads) do
    Budget.tick()
    local x0 = math.min(q[1][1], q[2][1], q[3][1], q[4][1])
    local x1 = math.max(q[1][1], q[2][1], q[3][1], q[4][1])
    local z0 = math.min(q[1][3], q[2][3], q[3][3], q[4][3])
    local z1 = math.max(q[1][3], q[2][3], q[3][3], q[4][3])
    if keepQuad(x0, z0, x1, z1) then
      push({ q[1], q[2], q[3], q[4] }, quadUV(q), groundShades(q, q.shade))
    end
  end

  -- true when the rect sits entirely inside one neighbour-body rect
  local function containedInMask(x0, z0, x1, z1)
    if not masks then return false end
    for _, mk in ipairs(masks) do
      if x0 >= mk[1] and x1 <= mk[3] and z0 >= mk[2] and z1 <= mk[4] then
        return true
      end
    end
    return false
  end

  -- round-tree stamps: the shared hull template translated per cell,
  -- through reusable scratch corners so expansion allocates nothing.
  -- A hull spans at most its own 16px cell, so one rect test usually
  -- answers for the whole stamp: strictly interior stamps keep every
  -- quad, ring stamps buried under a neighbour body (or, body-only, ring
  -- stamps full stop) skip without touching their quads. Only stamps
  -- crossing a boundary walk quad by quad.
  local sc = { { 0, 0, 0 }, { 0, 0, 0 }, { 0, 0, 0 }, { 0, 0, 0 } }
  for _, st in ipairs(S.roundStamps or {}) do
    local mx, mz = st.mx, st.mz
    local sx0, sz0, sx1, sz1 = mx - 8, mz - 8, mx + 8, mz + 8
    local interior = sx0 > 0 and sx1 < bw and sz0 > 0 and sz1 < bh
    local overBody = sx1 > 0 and sx0 < bw and sz1 > 0 and sz0 < bh
    local keepAll, skipAll
    if bodyOnly then
      keepAll = interior
      skipAll = not overBody
    else
      keepAll = interior or not maskedClosed(sx0, sz0, sx1, sz1)
      skipAll = not overBody and containedInMask(sx0, sz0, sx1, sz1)
    end
    if not skipAll then
      for _, q in ipairs(st.quads) do
        Budget.tick()
        for i = 1, 4 do
          local c, s2 = q[i], sc[i]
          s2[1] = c[1] + mx
          s2[2] = c[2]
          s2[3] = c[3] + mz
        end
        local ok = keepAll
        if not ok then
          local x0 = math.min(sc[1][1], sc[2][1], sc[3][1], sc[4][1])
          local x1 = math.max(sc[1][1], sc[2][1], sc[3][1], sc[4][1])
          local z0 = math.min(sc[1][3], sc[2][3], sc[3][3], sc[4][3])
          local z1 = math.max(sc[1][3], sc[2][3], sc[3][3], sc[4][3])
          ok = keepQuad(x0, z0, x1, z1)
        end
        if ok then
          push(sc, quadUV(q), groundShades(sc, q.shade))
        end
      end
    end
  end
end

-- The raw geometry for `map`: (vertex list, triangle index list, quad
-- count). Synchronous and GPU-free -- the headless suite and the probes
-- exercise the invariants through this.
function ChunkMesher.geometry(map, bodyOnly, masks)
  local sink = newTableSink()
  runGeometry(map, bodyOnly, masks, sink)
  return sink.results()
end

-- Build the mesh for `map` synchronously. Returns nil when there is
-- nothing to draw or meshes are unavailable (headless).
function ChunkMesher.build(map, bodyOnly, masks)
  local sink = newSink()
  runGeometry(map, bodyOnly, masks, sink)
  return sink.finish()
end

-- The tall-grass rows as their own mesh: VoxelScene draws it AFTER the
-- characters so the southern row of a grass cell still overdraws a
-- walker's feet (characters stamp over terrain, Gen 1 style, so ordinary
-- terrain could never do this).
local function buildGrassMesh(map)
  local S = Structures.forMap(map)
  if #S.grassQuads == 0 then return nil end
  local verts, indices, n = {}, {}, 0
  for _, q in ipairs(S.grassQuads) do
    for i = 1, 4 do
      local c = q[i]
      local uv = q.uv and q.uv[i] or { q.u, q.v }
      verts[#verts + 1] = { c[1], c[2], c[3], uv[1], uv[2], q.shade }
    end
    Voxel3D.pushQuad(indices, n)
    n = n + 1
  end
  return Voxel3D.newMesh(verts, indices)
end

-- ------------------------------------------------------------- the cache

local function entry(id)
  local c = cache[id]
  if not c then
    c = {}
    cache[id] = c
  end
  return c
end

local function releaseEntry(c)
  for _, slot in ipairs({ "full", "body", "grass" }) do
    local mesh = c[slot]
    if mesh and mesh.release then pcall(mesh.release, mesh) end
    c[slot] = nil
  end
end

-- ---------------------------------------------------------- async builds

local jobs = {}       -- FIFO of pending jobs
local jobIndex = {}   -- "id:slot" -> job

local clock = (love and love.timer and love.timer.getTime) or os.clock

local function jobKey(id, slot)
  return id .. ":" .. slot
end

local function finishJob(job, ok, err)
  jobIndex[jobKey(job.id, job.slot)] = nil
  for i, j in ipairs(jobs) do
    if j == job then
      table.remove(jobs, i)
      break
    end
  end
  if not ok then
    -- name the reason: in a real session a lost build is a black map
    print("[warn] voxel mesh build failed for " .. tostring(job.id)
          .. ": " .. tostring(err))
    if (gen[job.id] or 0) == job.gen then
      entry(job.id)[job.slot] = false
    end
  end
end

-- A build only lands if the map's generation still matches the one the
-- job was queued under -- invalidate/evict bump it to cancel in-flight
-- work whose inputs went stale.
local function runJob(job)
  local map = job.map
  local c = entry(job.id)
  if c.grass == nil then
    local okG, grass = pcall(buildGrassMesh, map)
    if (gen[job.id] or 0) ~= job.gen then return end
    c.grass = (okG and grass) or false
  end
  local sink = newSink()
  runGeometry(map, job.slot == "body", job.masks, sink)
  local mesh = sink.finish()
  if (gen[job.id] or 0) ~= job.gen then return end
  c[job.slot] = mesh or false
end

-- Queue a build unless the slot is already cached or queued. Returns the
-- cached mesh when there is one (false-cached misses return nil).
-- `urgent` marks the current map's meshes: pump() gives those a bigger
-- slice and runs them before neighbour jobs.
function ChunkMesher.request(map, bodyOnly, masks, urgent)
  local slot = bodyOnly and "body" or "full"
  local c = cache[map.id]
  if c and c[slot] ~= nil then return c[slot] or nil end
  local key = jobKey(map.id, slot)
  local job = jobIndex[key]
  if not job then
    job = { id = map.id, map = map, slot = slot, masks = masks,
            urgent = urgent or false, gen = gen[map.id] or 0 }
    jobIndex[key] = job
    jobs[#jobs + 1] = job
  elseif urgent then
    job.urgent = true
  end
  return nil
end

function ChunkMesher.pending()
  return #jobs
end

-- Advance queued builds inside a per-frame time budget. Urgent jobs (the
-- current map) come first and get the larger slice -- the first voxel
-- frame after a toggle is worth more milliseconds than a neighbour
-- popping in one frame later. `covered` says the world pass is hidden
-- this frame (a warp's fade, a menu): nothing visible can hitch, so the
-- slice opens up and a door fade swallows most of a destination build.
local URGENT_SLICE = 0.012
local IDLE_SLICE = 0.005
local COVERED_SLICE = 0.030

function ChunkMesher.pump(covered)
  if #jobs == 0 then return end
  local pick = jobs[1]
  for _, j in ipairs(jobs) do
    if j.urgent then
      pick = j
      break
    end
  end
  local slice = covered and COVERED_SLICE
                or (pick.urgent and URGENT_SLICE or IDLE_SLICE)
  local deadline = clock() + slice
  while pick do
    if not pick.co then
      pick.co = coroutine.create(runJob)
    end
    Budget.begin(pick.co, deadline - clock())
    local ok, err = coroutine.resume(pick.co, pick)
    Budget.finish()
    if not ok then
      finishJob(pick, false, err)
    elseif coroutine.status(pick.co) == "dead" then
      finishJob(pick, true)
    else
      return   -- slice spent mid-build; resume next frame
    end
    if clock() >= deadline or #jobs == 0 then return end
    pick = jobs[1]
    for _, j in ipairs(jobs) do
      if j.urgent then
        pick = j
        break
      end
    end
  end
end

-- Meshes for `map`, built SYNCHRONOUSLY on first use -- the historical
-- contract, kept for probes and any direct caller. `false` is cached for
-- a map whose mesh could not be built so a headless run does not retry
-- every frame. `masks` (the full variant's neighbour-body rects) is
-- static per map id -- a map's connections never change -- so it caches
-- like everything else.
function ChunkMesher.get(map, bodyOnly, masks)
  local slot = bodyOnly and "body" or "full"
  local c = entry(map.id)
  if c.grass == nil then
    local okG, grass = pcall(buildGrassMesh, map)
    c.grass = (okG and grass) or false
  end
  if c[slot] == nil then
    local ok, mesh = pcall(ChunkMesher.build, map, bodyOnly, masks)
    if not ok then
      print("[warn] voxel mesh build failed for " .. tostring(map.id)
            .. ": " .. tostring(mesh))
    end
    c[slot] = (ok and mesh) or false
    local key = jobKey(map.id, slot)
    local job = jobIndex[key]
    if job then finishJob(job, true) end
  end
  return c[slot] or nil
end

-- The cached mesh, or nil -- never builds. The async path's read side.
function ChunkMesher.peek(map, bodyOnly)
  local c = cache[map.id]
  local mesh = c and c[bodyOnly and "body" or "full"]
  return mesh or nil
end

function ChunkMesher.grass(map)
  local c = cache[map.id]
  return c and c.grass or nil
end

-- Evict everything outside `live` (a set of map ids): far maps' meshes
-- are released -- GPU buffer and LOVE's CPU copy both -- and their
-- Structures analysis dropped. The live set is the current map plus its
-- rendered neighbours, so memory stays bounded by what is on or near the
-- screen instead of growing with every area ever visited.
--
-- The PREVIOUS live set is retained too: warping into a building
-- collapses the set to one small interior, and evicting the town at the
-- door means rebuilding the whole neighbourhood on the way out -- a
-- flat-world flash after every house. One set of history makes the
-- round trip free while staying bounded at two neighbourhoods.
local prevLive = {}

function ChunkMesher.setLive(live)
  for id, c in pairs(cache) do
    if not live[id] and not prevLive[id] then
      releaseEntry(c)
      cache[id] = nil
      gen[id] = (gen[id] or 0) + 1
      Structures.invalidate(id)
    end
  end
  for i = #jobs, 1, -1 do
    local job = jobs[i]
    if not live[job.id] and not prevLive[job.id] then
      jobIndex[jobKey(job.id, job.slot)] = nil
      table.remove(jobs, i)
    end
  end
  prevLive = live
end

-- Drop one map's mesh (Cut swapped a block) or all of them (hot reload).
-- Structures' analysis is derived from the same block layer, so it drops
-- in the same breath; in-flight builds of the map are cancelled through
-- the generation counter.
function ChunkMesher.invalidate(mapId)
  Structures.invalidate(mapId)
  if mapId then
    local c = cache[mapId]
    if c then releaseEntry(c) end
    cache[mapId] = nil
    gen[mapId] = (gen[mapId] or 0) + 1
  else
    for _, c in pairs(cache) do releaseEntry(c) end
    cache = {}
    for id in pairs(gen) do gen[id] = gen[id] + 1 end
  end
  for i = #jobs, 1, -1 do
    local job = jobs[i]
    if mapId == nil or job.id == mapId then
      jobIndex[jobKey(job.id, job.slot)] = nil
      table.remove(jobs, i)
    end
  end
end

Assets.register(function() ChunkMesher.invalidate() end)

return ChunkMesher