diff --git a/code/duffle/gen/duffle.macs.h b/code/duffle/gen/duffle.macs.h
index 8d24ce4..b6a1716 100644
--- a/code/duffle/gen/duffle.macs.h
+++ b/code/duffle/gen/duffle.macs.h
@@ -2,7 +2,7 @@
 #pragma once
 #endif
 // Auto-generated by tape_atom_annotation_pass.lua — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
+// Source: C:/projects/Pikuma/ps1/code/duffle/lottes_tape.h
 // Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
 // + auto word-counts (so tape_atom.metadata.h stays manual-only
 //   for encoding macros).
@@ -85,7 +85,7 @@ WORD_COUNT(mac_pack_color_word, 3)
 
 #define mac_format_f3_color(r, g, b) \
 	mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
-WORD_COUNT(mac_format_f3_color, 1)
+WORD_COUNT(mac_format_f3_color, 3)
 
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
  * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
@@ -100,7 +100,7 @@ WORD_COUNT(mac_gte_store_f3_post_rtpt, 3)
 ,	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1) \
 ,	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2) \
 ,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
-WORD_COUNT(mac_format_g4_color, 4)
+WORD_COUNT(mac_format_g4_color, 12)
 
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
  * G4 triangle portion to p0/p1/p2.
diff --git a/code/duffle/lottes_tape.h b/code/duffle/lottes_tape.h
index ce0616e..e8aaa1e 100644
--- a/code/duffle/lottes_tape.h
+++ b/code/duffle/lottes_tape.h
@@ -13,7 +13,9 @@
 typedef U4 const MipsCode;
 typedef Slice_(MipsCode);
 typedef Slice_MipsCode MipsAtom;
+
 #define MipsAtom_(sym)               MipsCode tmpl(code,sym) [] align_(4) =
+
 // Bare form: file-scope declaration with hardcoded body.
 // Used for components with no args (e.g., ac_load_tri_indices) or
 // identifier-args (hardcoded register names).
@@ -21,6 +23,7 @@ typedef Slice_MipsCode MipsAtom;
 // expands to:
 //   MipsCode ac_X[] align_(4) = { body };
 #define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
+
 // Function form: function-body block that returns a MipsAtom slice.
 // Used for components with value-args (e.g., ac_format_f3_color).
 //   FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
diff --git a/code/gte_hello/gen/gte_hello.offsets.h b/code/gte_hello/gen/gte_hello.offsets.h
index 7415184..cba333b 100644
--- a/code/gte_hello/gen/gte_hello.offsets.h
+++ b/code/gte_hello/gen/gte_hello.offsets.h
@@ -1,13 +1,13 @@
 // Auto-generated by tape_atom_offset_gen.meta.lua — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\gte_hello\hello_gte_tape.c
+// Source: C:/projects/Pikuma/ps1/code/gte_hello/hello_gte_tape.c
 #pragma once
 
 #pragma region hello_gte_tape
 
 
-// --- atom: cube_g4_face (42 words) ---
+// --- atom: cube_g4_face (50 words) ---
 
-#define _atom_offset_cull_cube_g4_face_exit           26
+#define _atom_offset_cull_cube_g4_face_exit           34
 #define _atom_offset_bounds_chk_cube_g4_face_exit     2
 
 enum {
diff --git a/scripts/duffle.lua b/scripts/duffle.lua
index 5799c10..ac91c7d 100644
--- a/scripts/duffle.lua
+++ b/scripts/duffle.lua
@@ -336,9 +336,9 @@ function M.split_top_level_commas(body)
     local token_start = 1
     while i <= #body do
         local c = body:byte(i)
-        if c == 40 then local _, a = M.read_parens(body, i); i = a    -- '('
-        elseif c == 123 then local _, a = M.read_braces(body, i); i = a  -- '{'
-        elseif c == 91 then local _, a = M.read_brackets(body, i); i = a  -- '['
+        if c == 40 then local _, a = M.read_parens(body, i); i = a
+        elseif c == 123 then local _, a = M.read_braces(body, i); i = a
+        elseif c == 91 then local _, a = M.read_brackets(body, i); i = a
         elseif c == 44 then  -- ','
             tokens[#tokens + 1] = body:sub(token_start, i - 1)
             i = i + 1
@@ -347,7 +347,8 @@ function M.split_top_level_commas(body)
             local nx = M.skip_str_or_cmt(body, i)
             -- Don't advance token_start on skip: a trailing `/* ... */` after
             -- a macro call is part of the current entry's chunk, not the
-            -- start of the next one.
+            -- start of the next one. This keeps comments merged into the
+            -- next entry's chunk (so they don't inflate the word count).
             if nx > i then i = nx else i = i + 1 end
         end
     end
diff --git a/scripts/tape_atom.offset_gen.meta.lua b/scripts/tape_atom.offset_gen.meta.lua
index eb18aa9..4d458a1 100644
--- a/scripts/tape_atom.offset_gen.meta.lua
+++ b/scripts/tape_atom.offset_gen.meta.lua
@@ -208,6 +208,7 @@ local function scan_atom_body(body, word_counts)
     local pos      = 0
     local labels   = {}
     local branches = {}
+    local _dbg_count = 0
     for _, tok in ipairs(split_top_level_commas(body)) do
         local k    = 1
         local tlen = #tok
@@ -426,33 +427,60 @@ end
 -- Main
 -- ============================================================
 
+-- Scan a directory recursively for files matching a pattern.
+-- No regex (per the no_regex constraint) — uses plain byte matching.
+local function scan_dir(dir, suffix)
+	local results = {}
+	local p = io.popen('dir /b /s "' .. dir .. '\\' .. suffix .. '" 2>nul')
+	if not p then return results end
+	for raw_line in p:lines() do
+		-- dir /b outputs paths with backslashes; normalize to forward.
+		local path = raw_line:gsub("\\", "/")
+		results[#results + 1] = path
+	end
+	p:close()
+	return results
+end
+
 local function main(args)
 	if #args < 2 then
 		print("Usage: luajit gen_atom_offsets.lua <metadata.h> <source1> [source2 ...]")
 		os.exit(1)
 	end
-	local word_counts = load_word_counts(args[1])
-	for i = 2, #args do
-		local source_path  = args[i]
-		-- Also load the auto-generated <dir>.macs.h for this source's
-		-- directory (if it exists). The metaprogram emits
-		-- WORD_COUNT(mac_X, N) entries there for the MipsAtomComp_
-		-- declarations in the source. Merging ensures the offset
-		-- computation has the right word counts (no "unknown macro"
-		-- warnings for the auto-generated components).
-		local source_dir   = dirname(source_path)
-		-- dirname() keeps the trailing separator; basename_no_ext()
-		-- on a path with a trailing separator returns "". Strip it.
-		if #source_dir > 0 and (source_dir:sub(-1) == "/" or source_dir:sub(-1) == "\\") then
-			source_dir = source_dir:sub(1, -2)
-		end
-		local macs_path    = source_dir .. "/gen/" .. basename_no_ext(source_dir) .. ".macs.h"
-		local ok, mc        = pcall(load_word_counts, macs_path)
-		if ok then
+	-- Find the project root (parent of the metadata file's dir).
+	local meta_dir = dirname(args[1])
+	if #meta_dir > 0 and (meta_dir:sub(-1) == "/" or meta_dir:sub(-1) == "\\") then
+		meta_dir = meta_dir:sub(1, -2)
+	end
+	local project_root = dirname(meta_dir)
+	-- Scan the project for all *.macs.h files and load their WORD_COUNT
+	-- entries as the source of truth. The metaprogram's recursive
+	-- counting (which accounts for macro-in-macro expansion) gives
+	-- the correct values; the metadata file's manually-maintained
+	-- entries are NOT loaded here (they may be wrong for recursive cases).
+	local word_counts = {}
+	local macs_files = scan_dir(project_root, "*.macs.h")
+	for _, macs_path in ipairs(macs_files) do
+		-- Convert forward slashes to backslashes for Windows file open.
+		local win_path = macs_path:gsub("/", "\\")
+		local ok, mc = pcall(load_word_counts, win_path)
+		if ok and type(mc) == "table" then
 			for name, count in pairs(mc) do
 				word_counts[name] = count
 			end
 		end
+	end
+	-- Also load the metadata's encoding-macro entries (which are NOT
+	-- in the macs.h, since the macs.h is component-only). These are
+	-- the regular 1-word-per-macro entries like load_word, add_ui, etc.
+	local meta_counts = load_word_counts(args[1])
+	for name, count in pairs(meta_counts) do
+		if not word_counts[name] then
+			word_counts[name] = count
+		end
+	end
+	for i = 2, #args do
+		local source_path  = args[i]
 		process_source(source_path, word_counts)
 	end
 end
diff --git a/scripts/tape_atom_annotation_pass.lua b/scripts/tape_atom_annotation_pass.lua
index f411b52..b63e913 100644
--- a/scripts/tape_atom_annotation_pass.lua
+++ b/scripts/tape_atom_annotation_pass.lua
@@ -892,6 +892,62 @@ local function convert_line_comments_to_block(s)
 	return result
 end
 
+-- Compute the word count of a component body, accounting for
+-- macro expansion. Each comma-separated entry in the body is a
+-- "slot" that contributes its own word count. For most entries
+-- (regular MIPS instructions) the count is 1. For `mac_Y(...)`
+-- calls, the count is the word count of mac_Y (recursive lookup).
+--
+-- The component list passed in is the full set of components in the
+-- source, so we can resolve any `mac_X` call to its definition.
+local function compute_component_word_count(c, components)
+	-- Build a lookup table: mac_X_name -> component
+	local comp_by_name = {}
+	for _, cc in ipairs(components) do
+		comp_by_name[cc.name] = cc
+	end
+	-- Cache to avoid infinite recursion (if components cycle).
+	local cache = {}
+	local function rec(name)
+		if cache[name] then return cache[name] end
+		cache[name] = 0  -- mark as in-progress
+		local cc = comp_by_name[name]
+		local n
+		if cc then
+			local body_tokens = {}
+			for _, t in ipairs(duffle.split_top_level_commas(cc.body)) do
+				local trimmed = duffle.trim(t)
+				if trimmed ~= "" then body_tokens[#body_tokens + 1] = trimmed end
+			end
+			n = 0
+			for _, t in ipairs(body_tokens) do
+				-- Read the first identifier from the token.
+				local ident, after = duffle.read_ident(t, 1)
+				-- Components are stored without the `mac_` prefix
+				-- (e.g. "format_f3_color"). The token has `mac_format_f3_color(...)`,
+				-- so strip the `mac_` prefix to look up the component.
+				local comp_name = ident
+				if comp_name and comp_name:sub(1, 4) == "mac_" then
+					comp_name = comp_name:sub(5)
+				end
+				if comp_name and comp_by_name[comp_name] then
+					-- It's a `mac_X(...)` call. Recurse.
+					n = n + rec(comp_name)
+				else
+					-- Regular instruction (or some other token): 1 word.
+					n = n + 1
+				end
+			end
+		else
+			-- Not a known component: assume 1 word (regular instruction).
+			n = 1
+		end
+		cache[name] = n
+		return n
+	end
+	return rec(c.name)
+end
+
 local function emit_component_macros_h(source_path, components)
 		if #components == 0 then return end
 
@@ -956,19 +1012,36 @@ local function emit_component_macros_h(source_path, components)
 						sig = "..."
 				end
 
+        -- Compute the word count of the component body, accounting for
+        -- macro expansion. Each comma-separated entry in the body is one
+        -- "instruction slot", but a `mac_Y(...)` call expands to Y's
+        -- word count. The offset_gen and the metadata's WORD_COUNT table
+        -- both use this resolved count.
+        --
+        -- We compute it once per component (not per token) and emit
+        -- the value in the WORD_COUNT entry. The lookup table `counts_by_name`
+        -- is built from the components list (which find_component_atoms
+        -- already populated) and falls back to 1 for non-mac tokens.
+        local counts_by_name = {}
+        for _, cc in ipairs(components) do
+            local cc_count = compute_component_word_count(cc, components)
+            counts_by_name["mac_" .. cc.name] = cc_count
+        end
+        local n = compute_component_word_count(c, components)
+
         if n > 0 then
             -- Convert `//` line comments to `/* */` block comments in each
             -- token. C macros use `\` line-continuations; if a `//` comment
             -- appears before a `\`, the rest of the line (including the
             -- continuation) is consumed by the `//`, breaking the macro.
             -- Converting to `/* */` preserves the macro structure.
-            for j = 1, n do
+            for j = 1, #tokens do
                 tokens[j] = convert_line_comments_to_block(tokens[j])
             end
             -- Emit the mac_<X>(<sig>) macro
             lines[#lines + 1] = "#define mac_" .. c.name .. "(" .. sig .. ") \\"
             lines[#lines + 1] = "\t" .. tokens[1] .. " \\"
-            for j = 2, n do
+            for j = 2, #tokens do
                 lines[#lines + 1] = ",\t" .. tokens[j] .. " \\"
             end
 						-- Strip the trailing line-continuation on the last body line.