fixed.

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).
@@ -60,7 +60,7 @@ WORD_COUNT(mac_load_tri_verts, 18)
 ,	shift_lleft( R_AT, R_PrimCursor, S_(polytag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
 ,	shift_lright(R_AT, R_AT,         S_(polytag_len_bits)) \
 ,	store_word(  R_AT, R_T1,         O_(PolyTag,bf_addr_len)) /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_f3, 10)
+WORD_COUNT(mac_insert_ot_tag_f3, 11)
 
 /* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
  * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
@@ -75,7 +75,7 @@ WORD_COUNT(mac_insert_ot_tag_f3, 10)
 ,	shift_lleft( R_AT, R_PrimCursor, S_(polytag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
 ,	shift_lright(R_AT, R_AT,         S_(polytag_len_bits)) \
 ,	store_word(  R_AT, R_T1,         O_(PolyTag,bf_addr_len)) /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_g4, 10)
+WORD_COUNT(mac_insert_ot_tag_g4, 11)
 
 #define mac_pack_color_word(off, code, r, g, b) \
 	load_upper_i(R_AT, (code) << 8  | (b)) \

code/gte_hello/gen/gte_hello.offsets.h

@@ -1,24 +1,24 @@
 // 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 (50 words) ---
+// --- atom: cube_g4_face (87 words) ---
 
-#define _atom_offset_cull_cube_g4_face_exit           34
+#define _atom_offset_cull_cube_g4_face_exit           48
-#define _atom_offset_bounds_chk_cube_g4_face_exit     2
+#define _atom_offset_bounds_chk_cube_g4_face_exit     12
 
 enum {
     atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
     atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
 };
 
-// --- atom: floor_f3_face (47 words) ---
+// --- atom: floor_f3_face (66 words) ---
 
-#define _atom_offset_culling_floor_f3_face_exit       14
+#define _atom_offset_culling_floor_f3_face_exit       29
-#define _atom_offset_bounds_chk_floor_f3_face_exit    3
+#define _atom_offset_bounds_chk_floor_f3_face_exit    13
 
 enum {
     atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,

code/gte_hello/hello_gte_tape.c

@@ -9,10 +9,8 @@
 #endif
 
 #pragma region MACs (Mips Atom components)
-/* The macros mac_format_f3_color and mac_gte_store_f3 moved to
+
- * lottes_tape.h during the Phase 3 gp.h overhaul. Both are now RGB-form
+
- * (mac_format_f3_color takes _r, _g, _b byte values rather than raw
- * 16-bit half-words). */
 
 #pragma endregion MACs
 

scripts/duffle.lua

@@ -330,30 +330,85 @@ end
 
 -- Split a brace-body into top-level comma-separated tokens. Honors nested
 -- parens/braces/brackets and skips strings/comments.
+--
+-- FIX (2026-07-09): split at top-level NEWLINES and SEMICOLONS too, AND
+-- emit a token break after a top-level comment/string. Previous behavior
+-- glued the macro call after a comment into the same token, so
+-- `word_count_of_token` only saw the leading ident (often nil after
+-- stripping the comment), undercounting the body. See Phase 1 of the
+-- branch-offset regression investigation. Pure-comment / pure-string
+-- chunks (which now appear between real statements) are filtered out so
+-- they contribute 0 words instead of 1.
 function M.split_top_level_commas(body)
     local tokens = {}
     local i = 1
+    local len = #body
     local token_start = 1
-    while i <= #body do
+
+    -- True iff `chunk` contains any non-whitespace, non-comment, non-string
+    -- content (i.e., real token material). Walks through ws + comments
+    -- individually so a chunk like "  /* trailing */ shift_lleft(...)"
+    -- is correctly classified as having real content (the macro call).
+    local function has_real_content(chunk)
+        local k = 1
+        local klen = #chunk
+        while k <= klen do
+            if M.is_space(chunk:sub(k, k)) then
+                k = k + 1
+            else
+                local nx = M.skip_str_or_cmt(chunk, k)
+                if nx > k then
+                    k = nx  -- skipped a comment or string
+                else
+                    return true  -- found real content
+                end
+            end
+        end
+        return false
+    end
+
+    local function emit(end_pos)
+        if end_pos >= token_start then
+            local chunk = body:sub(token_start, end_pos)
+            if M.trim(chunk) ~= "" and has_real_content(chunk) then
+                tokens[#tokens + 1] = chunk
+            end
+            token_start = end_pos + 1
+        end
+    end
+
+    while i <= len do
         local c = body:byte(i)
-        if c == 40 then local _, a = M.read_parens(body, i); i = a
+        if c == 40 then                              -- '('
-        elseif c == 123 then local _, a = M.read_braces(body, i); i = a
+            local _, a = M.read_parens(body, i); i = a
-        elseif c == 91 then local _, a = M.read_brackets(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)
+            emit(i - 1)
+            i = i + 1
+            token_start = i
+        elseif c == 59 then                           -- ';'
+            emit(i - 1)
+            i = i + 1
+            token_start = i
+        elseif c == 10 then                           -- '\n'
+            emit(i - 1)
             i = i + 1
             token_start = i
         else
             local nx = M.skip_str_or_cmt(body, i)
-            -- Don't advance token_start on skip: a trailing `/* ... */` after
+            if nx > i then
-            -- a macro call is part of the current entry's chunk, not the
+                -- Skipped a comment or string at top level: emit token break.
-            -- start of the next one. This keeps comments merged into the
+                i = nx
-            -- next entry's chunk (so they don't inflate the word count).
+                emit(i - 1)
-            if nx > i then i = nx else i = i + 1 end
+            else
+                i = i + 1
             end
         end
-    local last = body:sub(token_start)
+    end
-    if M.trim(last) ~= "" then tokens[#tokens + 1] = last end
+    emit(len)
     return tokens
 end
 

scripts/tape_atom_annotation_pass.lua

@@ -897,10 +897,18 @@ end
 -- "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).
+-- For encoding macros with a known multi-word count (e.g.
+-- `mask_upper` = 2), the count is taken from `word_counts`.
 --
 -- 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)
+--
+-- FIX (2026-07-09): also consult `word_counts` (encoding macros from
+-- tape_atom.metadata.h) for non-component tokens. Previously `mask_upper`
+-- inside `mac_insert_ot_tag_*` was assumed to be 1 word instead of 2,
+-- so any atom using `mac_insert_ot_tag_*` had its branch offsets off by
+-- 1 word. See Phase 4b of the branch-offset regression investigation.
+local function compute_component_word_count(c, components, word_counts)
 	-- Build a lookup table: mac_X_name -> component
 	local comp_by_name = {}
 	for _, cc in ipairs(components) do
@@ -933,8 +941,15 @@ local function compute_component_word_count(c, components)
 				if comp_name and comp_by_name[comp_name] then
 					-- It's a `mac_X(...)` call. Recurse.
 					n = n + rec(comp_name)
+				elseif comp_name and word_counts and word_counts[comp_name] then
+					-- Encoding macro or pseudo-instruction (e.g. mask_upper = 2,
+					-- nop2 = 2). Trust the metadata — tape_atom.metadata.h is the
+					-- single source of truth for word counts.
+					n = n + word_counts[comp_name]
 				else
-					-- Regular instruction (or some other token): 1 word.
+					-- Unrecognized token. Fall back to 1 word. tape_atom.metadata.h
+					-- should declare a WORD_COUNT entry for every macro used in
+					-- a component body — add the missing entry if you see drift.
 					n = n + 1
 				end
 			end
@@ -948,7 +963,7 @@ local function compute_component_word_count(c, components)
 	return rec(c.name)
 end
 
-local function emit_component_macros_h(source_path, components)
+local function emit_component_macros_h(source_path, components, word_counts)
 		if #components == 0 then return end
 
 		local dir          = duffle.dirname(source_path)
@@ -1024,10 +1039,10 @@ local function emit_component_macros_h(source_path, components)
         -- 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)
+            local cc_count = compute_component_word_count(cc, components, word_counts)
             counts_by_name["mac_" .. cc.name] = cc_count
         end
-        local n = compute_component_word_count(c, components)
+        local n = compute_component_word_count(c, components, word_counts)
 
         if n > 0 then
             -- Convert `//` line comments to `/* */` block comments in each
@@ -1593,7 +1608,7 @@ local function main(args)
 				local source_text = duffle.read_file(source_path)
 				local components  = find_component_atoms(source_text)
 				if #components > 0 then
-						emit_component_macros_h(source_path, components)
+						emit_component_macros_h(source_path, components, word_counts)
 				end
 		end