lua metaprogram improvmeents

scripts/duffle.lua

@@ -1338,7 +1338,6 @@ M.GTE_CR_ALIAS_GROUPS = {
 -- first must be written before second.
 M.GTE_PACKED_SLOT_RELATIONS = {
 	{ slot = 2, first = "gte_cr_RT13", second = "gte_cr_RT22" },
-	{ slot = 4, first = "gte_cr_RT22", second = "gte_cr_RT33" },
 }
 
 -- Operand-class table for the COP2->GPR load-delay check.
@@ -2924,18 +2923,17 @@ end
 -------------------------------------------------------------------------------
 -- find_atom_proc_decl_for — backward walk for MipsAtom_Proc_ name extraction.
 --
--- After the `sym` arg was dropped from MipsAtom_Proc_, the atom name is
--- derived from the preceding `MipsAtom* X_proc(args)` function declaration.
+-- The atom name is the preceding `MipsAtom* ident(args)` function ident.
 -- This function walks backward from `before_pos` to find it.
 --
--- Returns (raw_name, args_inner) or (nil, nil).
---   raw_name   — e.g. "normalize_v3s4" (the _proc suffix is stripped)
---   args_inner — e.g. "AtomArena_R aa, U4 r_scratch, ..."
+-- Returns (raw_name, args_inner, func_ident, after_paren) or (nil, nil).
+--   raw_name    — the function ident as written
+--   args_inner  — e.g. "AtomArena_R aa, U4 r_scratch, ..."
+--   after_paren — source position after the function `)`
 --
 -- The walk finds the LAST "MipsAtom*" before before_pos, then skips
 -- whitespace + qualifiers (internal, I_, FI_, comments) until it finds an
--- ident followed by "(". That ident is the function name (with _proc suffix);
--- the suffix is stripped to get raw_name. The parens contents are the args.
+-- ident followed by "(". That ident is the name. The parens contents are the args.
 -------------------------------------------------------------------------------
 function M.find_atom_proc_decl_for(source, before_pos, mips_atom_ptr_len)
 	local search_pos = 1
@@ -2980,9 +2978,9 @@ function M.find_atom_proc_decl_for(source, before_pos, mips_atom_ptr_len)
 		-- check if the next non-ws char after ident is "("
 		local next_pos = M.skip_ws_and_cmt(source, ident_end)
 		if source:sub(next_pos, next_pos) == "(" then
-			local inner = M.read_parens(source, next_pos)
+			local inner, after_paren = M.read_parens(source, next_pos)
 			if inner then
-				return ident, inner, ident
+				return ident, inner, ident, after_paren
 			end
 		end
 		-- ident not followed by "(" — it's a qualifier; skip it

scripts/passes/components.lua

@@ -7,7 +7,7 @@
 --- then resolves the function-args string from the preceding `FI_ Slice_MipsCode ac_X(...)` declaration via a backward walk.
 ---
 --- `MipsAtom_Proc_(X, ab, { body })` declarations (kind="atom_proc") are ATOMS, not components, and are deliberately excluded —
---- atoms get emitted via `tb_emit(tb, code_<name>)` linker symbols, not inlined as `mac_*` macros.
+--- the ELF symbol is the C ident. Raw `MipsCode code_*` is leftover, not the atom rule.
 ---
 --- Emits one `gen/macs.h` per *immediate source directory* with `#define mac_X(sig) \` macros plus `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
 --- All sources inside the same directory contribute to the same file (per-directory aggregation).
@@ -187,7 +187,7 @@ local function project_components(source, scan)
 		-- Only `MipsAtomComp_(ac_X)` (kind="comp_bare") and `MipsAtomComp_Proc_(ac_X, ...)` (kind="comp_proc")
 		-- are COMPONENTS — they get inlined via `mac_<name>` aliases inside atom bodies.
 		-- `MipsAtom_Proc_` (kind="atom_proc") is an ATOM (ends with `mac_yield()`); it gets emitted via
-		-- `tb_emit(tb, code_<name>)` (linker symbol), NOT inlined as a macro. Including `atom_proc` here
+		-- `tb_emit` of the C ident, NOT inlined as a macro. Including `atom_proc` here
 		-- would incorrectly emit `mac_<name>` aliases for atoms, polluting `gen/macs.h`.
 		-- See `docs/duffle_dsl_primer.md` §"mac_* aliases" for the contract.
 		if a.kind == "comp_bare" or a.kind == "comp_proc" then
@@ -393,6 +393,10 @@ end
 --- @param cache        table<string, integer>
 --- @return integer
 local function gp0_contrib_rec(name, comp_by_name, cache)
+	if name:match("^insert_ot_tag") then
+		cache[name] = 0
+		return 0
+	end
 	if cache[name] ~= nil then return cache[name] end
 	cache[name] = -1
 	local cc = comp_by_name[name]
@@ -408,8 +412,15 @@ local function gp0_contrib_rec(name, comp_by_name, cache)
 					-- Nested `mac_X(...)` call: recurse.
 					local nested = ident:sub(MAC_PREFIX_LEN + 1)
 					n = n + gp0_contrib_rec(nested, comp_by_name, cache)
+				elseif ident == "gte_sw" then
+					n = n + 1
 				elseif ident == "store_word" or ident == "store_half" or ident == "store_byte" then
-					if trimmed:find("R_PrimCursor", 1, true) then
+					if trimmed:find("R_PrimCursor", 1, true)
+						or trimmed:find("O_(Poly_", 1, true)
+						or trimmed:find("r_prim_cursor", 1, true)
+						or trimmed:find("r_primitive_cursor", 1, true)
+						or trimmed:find("r_base", 1, true)
+					then
 						n = n + 1
 					end
 				end

scripts/passes/offsets.lua

@@ -1,7 +1,8 @@
 --- passes/offsets.lua — Branch-offset generator.
 ---
 --- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
---- for `MipsAtom_(name)` and `MipsCode code_<name>` declarations, computes the word offset
+--- for `MipsAtom_(name)` and leftover `MipsCode code_*` declarations, computes the word offset
+--- (ELF symbol is the C ident; raw `code_*` is leftover, not the atom rule)
 --- from each `atom_offset(F, T)` marker to its target `atom_label(T)` declaration, and emits
 --- `gen/offsets.h` with one `#define _atom_offset_F_T = N` per branch.
 ---

scripts/passes/report.lua

@@ -419,6 +419,29 @@ local function render_section_annotations(add, view)
 	add("")
 end
 
+local function render_section_component_annotations(add, view)
+	local rows = {}
+	for _, src in ipairs(view.sources) do
+		for _, info in ipairs((src.scan and src.scan.component_atom_infos) or {}) do
+			rows[#rows + 1] = {
+				source = source_basename(src.path),
+				line   = info.info_line or 0,
+				name   = info.atom_name or "?",
+				reads  = (#(info.reads or {}) > 0 and table.concat(info.reads, ",")) or "—",
+				writes = (#(info.writes or {}) > 0 and table.concat(info.writes, ",")) or "—",
+			}
+		end
+	end
+	if #rows == 0 then add("_(none)_"); add(""); return end
+	add("| source | line | name | reads | writes |")
+	add("|--------|------|------|-------|--------|")
+	for _, r in ipairs(rows) do
+		add(string.format("| %s | %d | %s | %s | %s |",
+			r.source, r.line, r.name, r.reads, r.writes))
+	end
+	add("")
+end
+
 local function render_section_binds(add, view)
 	local wrote = false
 	for _, src in ipairs(view.sources) do
@@ -610,21 +633,123 @@ local function render_section_relations(add, view)
 	if not wrote then add("_(none)_"); add("") end
 end
 
+local HIDDEN_UNLESS_WRITTEN = {
+	R_AT = true, R_TapePtr = true, R_AtomJmp = true,
+}
+
+local PHYSICAL_GPR = {
+	R_T0 = true, R_T1 = true, R_T2 = true, R_T3 = true,
+	R_T4 = true, R_T5 = true, R_T6 = true, R_T7 = true,
+	R_V0 = true, R_V1 = true,
+}
+
+local function encoder_wrote_key(atom, key)
+	for _, ev in ipairs((atom.paths and atom.paths.word_events) or {}) do
+		for _, dest in pairs(ev.gpr_keys or {}) do
+			if dest == key then return true end
+		end
+	end
+	return false
+end
+
+local function written_name_for(key, atom)
+	local slot = key:match("^reguse:.+:(.+)$")
+	if slot then
+		local param = atom.reg_use_param_name
+		if param and param ~= "" then return param .. "." .. slot end
+		return slot
+	end
+	return key
+end
+
+local function aliases_for_key(key, atom, view)
+	local slot = key:match("^reguse:.+:(.+)$")
+	if not slot then return "—" end
+	local schema_name = atom.reg_use_schema_name
+	local schema = view.corpus and view.corpus.reg_use_schemas and view.corpus.reg_use_schemas[schema_name]
+	if not schema then return "—" end
+	for _, s in ipairs(schema.slots or {}) do
+		if s.name == slot then
+			local names = {}
+			for _, alias in ipairs(s.aliases or {}) do
+				if alias ~= slot then names[#names + 1] = alias end
+			end
+			if #names == 0 then
+				if s.aliases and #s.aliases > 0 then return table.concat(s.aliases, ", ") end
+				return "—"
+			end
+			return table.concat(names, ", ")
+		end
+	end
+	return "—"
+end
+
+local function physical_for_key(key, atom, view)
+	if PHYSICAL_GPR[key] then return key end
+	local corpus = view.corpus or {}
+	local alias = (corpus.register_alias_registry or {})[key]
+	if type(alias) == "table" then
+		local phys = alias.physical or alias.gpr or alias.code_name
+		if type(phys) == "string" and PHYSICAL_GPR[phys] then return phys end
+		if type(alias.name) == "string" and PHYSICAL_GPR[alias.name] then return alias.name end
+	elseif type(alias) == "string" and PHYSICAL_GPR[alias] then
+		return alias
+	end
+	local atom_map = (corpus.atom_auto_regs or {})[atom.name]
+	if type(atom_map) == "table" then
+		local slot = key:match("^reguse:.+:(.+)$") or key
+		local bound = atom_map[slot] or atom_map["R_" .. slot]
+		if type(bound) == "string" and PHYSICAL_GPR[bound] then return bound end
+	end
+	return "—"
+end
+
+local function last_relation_for(key, atom)
+	local last = nil
+	for _, rel in ipairs((atom.paths and atom.paths.relations) or {}) do
+		local dest = rel.destination or rel.producer_destination
+		if dest == key then last = rel end
+	end
+	if not last then return "—" end
+	local sem = last.semantic or "?"
+	local a = last.producer_word
+	local b = last.consumer_word
+	if a and b then return string.format("%s w%s→%s", sem, tostring(a), tostring(b)) end
+	return sem
+end
+
 local function render_section_forward(add, view)
 	local wrote = false
 	for _, a in ipairs(view.decls) do
 		local gpr = a.paths and a.paths.forward_state and a.paths.forward_state.gpr_values
 		local keys = {}
 		for k in pairs(gpr or {}) do
-			if k ~= "R_0" then keys[#keys + 1] = k end
+			if k == "R_0" then
+				-- hidden
+			elseif HIDDEN_UNLESS_WRITTEN[k] and not encoder_wrote_key(a, k) then
+				-- hidden
+			else
+				keys[#keys + 1] = k
+			end
 		end
 		if #keys > 0 then
 			wrote = true
 			add("### " .. a.name)
+			add("| written | aliases | physical | lattice | last relation |")
+			add("|---|---|---|---|---|")
 			table.sort(keys)
 			for _, k in ipairs(keys) do
 				local slot = gpr[k]
-				add(string.format("- `%s` %s", k, (slot and slot.kind) or "unknown"))
+				local lattice = "—"
+				if slot and slot.kind == "constant" then
+					lattice = tostring(slot.value)
+				end
+				add(string.format("| `%s` | %s | %s | %s | %s |",
+					written_name_for(k, a),
+					aliases_for_key(k, a, view),
+					physical_for_key(k, a, view),
+					lattice,
+					last_relation_for(k, a)))
 			end
 			add("")
 		end
@@ -637,6 +762,7 @@ local SECTION_RENDERERS = {
 	{ header = "## Components",            render = render_section_components },
 	{ header = "## RegUse schemas",        render = render_section_reguse },
 	{ header = "## Annotations",           render = render_section_annotations },
+	{ header = "## Component annotations", render = render_section_component_annotations },
 	{ header = "## Binds_* structs",       render = render_section_binds },
 	{ header = "## Phases / views / ctx",  render = render_section_phases },
 	{ header = "## Register aliases",      render = render_section_aliases },
@@ -644,7 +770,7 @@ local SECTION_RENDERERS = {
 	{ header = "## Collisions",            render = render_section_collisions },
 	{ header = "## Findings",              render = render_section_findings },
 	{ header = "## Relations",             render = render_section_relations },
-	{ header = "## Forward GPR",           render = render_section_forward },
+	{ header = "## GPR model",             render = render_section_forward },
 }
 
 --- Render the consolidated per-module markdown (`build/<module>.atom_meta_report.md`).

scripts/passes/scan_source.lua

@@ -416,31 +416,51 @@ local function walk_body_fields(body, build_field)
 	return fields
 end
 
--- Parse the `<type> <field>;` declarations from a Struct_ body.
+-- Parse the `<type> <field>[, <field>...];` declarations from a Struct_ body.
+-- After the type and `*` chain, keep reading `, ident` until `;`.
+-- Same type, same pointer depth for every name on that list.
 -- Returns the raw fields array with `{name, type_name, pointer_depth}` only (NO offset / byte_size).
 -- The propagation pass `resolve_struct_field_sizes` walks each struct's fields AFTER type resolution and populates offset + byte_size in place.
--- Returns (fields). The aggregate byte_count is computed in the propagation pass (it depends on whether every field's type resolved).
 local function parse_struct_body_fields(body)
-	return walk_body_fields(body, function(type_name, type_end, after_type)
-		-- Parse the trailing `*` chain to derive pointer_depth.
-		local depth, cursor = 0, after_type
-		while cursor <= #body and body:sub(cursor, cursor) == "*" do
-			depth  = depth  + 1
-			cursor = cursor + 1
-			cursor = duffle.skip_ws_and_cmt(body, cursor)
+	local fields = {}
+	local body_pos = 1
+	local body_len = #body
+	while body_pos <= body_len do
+		body_pos = duffle.skip_ws_and_cmt(body, body_pos)
+		if body_pos > body_len then break end
+		local type_name, type_end = duffle.read_ident(body, body_pos)
+		if not type_name then
+			body_pos = body_pos + 1
+		else
+			local depth, cursor = 0, duffle.skip_ws_and_cmt(body, type_end)
+			while cursor <= body_len and body:sub(cursor, cursor) == "*" do
+				depth = depth + 1
+				cursor = duffle.skip_ws_and_cmt(body, cursor + 1)
+			end
+			while cursor <= body_len do
+				local field_ident, field_end = duffle.read_ident(body, cursor)
+				if not field_ident then break end
+				fields[#fields + 1] = {
+					name          = field_ident,
+					type_name     = type_name,
+					pointer_depth = depth,
+					offset        = nil,
+					byte_size     = nil,
+				}
+				cursor = duffle.skip_ws_and_cmt(body, field_end)
+				if body:sub(cursor, cursor) == "," then
+					cursor = duffle.skip_ws_and_cmt(body, cursor + 1)
+				else
+					break
+				end
+			end
+			if cursor <= body_len and body:sub(cursor, cursor) == ";" then
+				cursor = cursor + 1
+			end
+			body_pos = cursor
 		end
-		-- Read the field ident immediately after the type chain.
-		local field_ident, field_end = duffle.read_ident(body, cursor)
-		if not field_ident then return nil, type_end + 1 end
-		return {
-			name          = field_ident,
-			type_name     = type_name,
-			pointer_depth = depth,
-			-- offset + byte_size filled by resolve_struct_field_sizes
-			offset        = nil,
-			byte_size     = nil,
-		}, field_end
-	end)
+	end
+	return fields
 end
 
 -- Parse the `Enum_(<underlying>, <name>) { <body> }` body for entries.
@@ -1299,33 +1319,6 @@ local function parse_atom_info_after_decl(source, after_paren, raw_name, line_of
 	return info_after
 end
 
---- Parse `atom_info(...)` immediately before `before_pos` (the MipsAtom_Proc_ token).
-local function parse_atom_info_before(source, before_pos, raw_name, line_of, out, dest)
-	local i = before_pos - 1
-	while i >= 1 and source:sub(i, i):match("%s") do i = i - 1 end
-	if source:sub(i, i) ~= ")" then return end
-	local depth = 0
-	local j = i
-	while j >= 1 do
-		local c = source:sub(j, j)
-		if c == ")" then
-			depth = depth + 1
-		elseif c == "(" then
-			depth = depth - 1
-			if depth == 0 then break end
-		end
-		j = j - 1
-	end
-	if j < 1 then return end
-	local k = j - 1
-	while k >= 1 and source:sub(k, k):match("%s") do k = k - 1 end
-	local ident_end = k
-	while k >= 1 and source:sub(k, k):match("[%w_]") do k = k - 1 end
-	local ident_start = k + 1
-	if source:sub(ident_start, ident_end) ~= "atom_info" then return end
-	parse_atom_info_after_decl(source, ident_start, raw_name, line_of, out, dest)
-end
-
 --- Parse: `MipsAtom_(<name>) [atom_info(<binds>, <reads>, <writes>)] { <body> }`
 --- @param source    string
 --- @param pos       integer
@@ -1466,10 +1459,15 @@ local function parse_mips_atom_proc(source, pos, ident_end, line_of, out)
 
 	-- The atom name is the preceding function ident as written
 	-- (`internal MipsAtom* X(...)`). The first macro arg is the arena.
-	local raw_name, args_inner, func_ident = duffle.find_atom_proc_decl_for(source, open_paren, MIPS_ATOM_PTR_LEN)
+	local raw_name, args_inner, func_ident, after_func_paren =
+		duffle.find_atom_proc_decl_for(source, open_paren, MIPS_ATOM_PTR_LEN)
 	if not raw_name then raw_name = "?" end
 	local name = strip_ac_prefix(raw_name)
-	parse_atom_info_before(source, pos, name, line_of, out, out.atom_infos)
+	if after_func_paren then
+		parse_atom_info_after_decl(source, after_func_paren, name, line_of, out, out.atom_infos)
+	else
+		parse_atom_info_after_decl(source, pos, name, line_of, out, out.atom_infos)
+	end
 	local reg_use_schema_name = nil
 	local reg_use_param_name  = nil
 	if args_inner then
@@ -1617,18 +1615,18 @@ local function register_typedef_alias(underlying, name, pos, line_of, out)
 	}
 end
 
--- Layout of Reg_<type> only. Never the C data struct (V3_S4 has a pad field).
-local REG_ALLOC_FIELDS = {
-	Reg_V3_S4 = { "x", "y", "z" },
-	Reg_P3_S4 = { "x", "y", "z" },
-	Reg_V3_S2 = { "x", "y", "z" },
-}
-
 local parse_reg_use_schema_body
 
 local function fields_for_reg_type(type_name, type_registry)
 	local reg_name = "Reg_" .. type_name
 	local entry = type_registry and type_registry[reg_name]
+	if entry and entry.fields and #entry.fields > 0 then
+		local names = {}
+		for _, field in ipairs(entry.fields) do
+			if field.name then names[#names + 1] = field.name end
+		end
+		if #names > 0 then return names end
+	end
 	if entry and entry.body and parse_reg_use_schema_body then
 		local schema = parse_reg_use_schema_body(entry.body, type_registry)
 		if schema and schema.slots then
@@ -1639,10 +1637,13 @@ local function fields_for_reg_type(type_name, type_registry)
 			if #names > 0 then return names end
 		end
 	end
-	return REG_ALLOC_FIELDS[reg_name]
+	return nil
 end
 
-parse_reg_use_schema_body = function(body, type_registry)
+parse_reg_use_schema_body = function(body, type_registry, opts)
+	opts = opts or {}
+	local require_types = opts.require_types == true
+	local pending = false
 	local slots = {}
 	local alias_to_slot = {}
 	local slot_names = {}
@@ -1789,10 +1790,14 @@ parse_reg_use_schema_body = function(body, type_registry)
 					errors[#errors + 1] = { kind = "reguse_malformed" }
 					return nil, errors
 				end
-				typed_fields = fields_for_reg_type(duffle.trim(type_inner), type_registry)
+				local type_ident = duffle.trim(type_inner)
+				typed_fields = fields_for_reg_type(type_ident, type_registry)
 				if not typed_fields then
-					errors[#errors + 1] = { kind = "reguse_malformed" }
-					return nil, errors
+					if require_types then
+						errors[#errors + 1] = { kind = "reguse_unknown_reg_type", type_name = type_ident }
+					else
+						pending = true
+					end
 				end
 				after = duffle.skip_ws_and_cmt(body, after_paren)
 			end
@@ -1808,11 +1813,13 @@ parse_reg_use_schema_body = function(body, type_registry)
 				return nil, errors
 			end
 			for _, n in ipairs(names) do
-				if typed_fields then
-					for _, field in ipairs(typed_fields) do
-						local path = n .. "." .. field
-						if not add_alias(path, path) then return nil, errors end
-						if not add_slot(path, { path }, readonly) then return nil, errors end
+				if first == "Reg_" then
+					if typed_fields then
+						for _, field in ipairs(typed_fields) do
+							local path = n .. "." .. field
+							if not add_alias(path, path) then return nil, errors end
+							if not add_slot(path, { path }, readonly) then return nil, errors end
+						end
 					end
 				else
 					if not add_alias(n, n) then return nil, errors end
@@ -1827,10 +1834,15 @@ parse_reg_use_schema_body = function(body, type_registry)
 		::continue::
 	end
 	if #slots == 0 then
-		errors[#errors + 1] = { kind = "reguse_malformed" }
+		if pending and not require_types then
+			return { slots = slots, alias_to_slot = alias_to_slot, pending = true }, errors
+		end
+		if #errors == 0 then
+			errors[#errors + 1] = { kind = "reguse_malformed" }
+		end
 		return nil, errors
 	end
-	return { slots = slots, alias_to_slot = alias_to_slot }, errors
+	return { slots = slots, alias_to_slot = alias_to_slot, pending = pending }, errors
 end
 
 --- Parse: `typedef` declarations.
@@ -2231,6 +2243,137 @@ local DECL_PARSERS = {
 -- Only the bare `atom_dbg_skip` marker reaches `parse_dbg_skip_marker`.
 -- Unknown identifiers follow the same unrelated-token path as every other unsupported source token.
 
+local TAPE_SKIP_MACROS = {
+	MipsAtom_ = true,
+	MipsAtom_Proc_ = true,
+	MipsAtomComp_ = true,
+	MipsAtomComp_Proc_ = true,
+	MipsAtomComp_ProcMap_ = true,
+	Struct_ = true,
+	Enum_ = true,
+}
+
+local function collect_addrs_assigns(text)
+	local addrs = {}
+	local pos = 1
+	local n = #text
+	while pos <= n do
+		pos = duffle.skip_ws_and_cmt(text, pos)
+		if pos > n then break end
+		local ident, ident_end = duffle.read_ident(text, pos)
+		if ident == "addrs" then
+			local after = duffle.skip_ws_and_cmt(text, ident_end)
+			if text:sub(after, after) == "[" then
+				local inner, after_br = duffle.read_brackets(text, after)
+				local idx = inner and tonumber(duffle.trim(inner))
+				after_br = duffle.skip_ws_and_cmt(text, after_br or after)
+				if idx and text:sub(after_br, after_br) == "=" then
+					local rhs = duffle.skip_ws_and_cmt(text, after_br + 1)
+					local rhs_ident = duffle.read_ident(text, rhs)
+					if rhs_ident then addrs[idx] = rhs_ident end
+					pos = rhs
+				else
+					pos = after_br or (after + 1)
+				end
+			else
+				pos = ident_end
+			end
+		elseif ident then
+			pos = ident_end
+		else
+			pos = pos + 1
+		end
+	end
+	return addrs
+end
+
+local function collect_tb_emits(body, addrs)
+	local names = {}
+	local pos = 1
+	local n = #body
+	while pos <= n do
+		pos = duffle.skip_ws_and_cmt(body, pos)
+		if pos > n then break end
+		local ident, ident_end = duffle.read_ident(body, pos)
+		if ident == "tb_emit_" or ident == "tb_emit" then
+			local after = duffle.skip_ws_and_cmt(body, ident_end)
+			if body:sub(after, after) == "(" then
+				local inner, after_p = duffle.read_parens(body, after)
+				local name
+				if ident == "tb_emit_" then
+					name = duffle.trim(inner or ""):match("^([%w_]+)")
+				else
+					local args = duffle.split_top_level_commas(inner or "")
+					local last = duffle.trim(args[#args] or "")
+					local idx = last:match("^addrs%s*%[%s*(%d+)%s*%]$")
+					if idx then
+						name = addrs[tonumber(idx)]
+					else
+						name = last:match("([%w_]+)$")
+					end
+				end
+				if name then names[#names + 1] = name end
+				pos = after_p or (after + 1)
+			else
+				pos = ident_end
+			end
+		elseif ident then
+			pos = ident_end
+		else
+			pos = pos + 1
+		end
+	end
+	return names
+end
+
+-- Linear appearance order of tb_emit / tb_emit_ in each C function body.
+-- Commented-out emits are skipped by skip_ws_and_cmt. No C if/loop CFG.
+local function scan_tape_chains(source)
+	local addrs = collect_addrs_assigns(source)
+	local chains = {}
+	local pos = 1
+	local n = #source
+	while pos <= n do
+		pos = duffle.skip_ws_and_cmt(source, pos)
+		if pos > n then break end
+		local ident, ident_end = duffle.read_ident(source, pos)
+		if ident and TAPE_SKIP_MACROS[ident] then
+			local after = duffle.skip_ws_and_cmt(source, ident_end)
+			if source:sub(after, after) == "(" then
+				local _, after_p = duffle.read_parens(source, after)
+				after = duffle.skip_ws_and_cmt(source, after_p or after)
+			end
+			if source:sub(after, after) == "{" then
+				local _, after_b = duffle.read_braces(source, after)
+				pos = after_b or (after + 1)
+			else
+				pos = after
+			end
+		elseif ident then
+			local after = duffle.skip_ws_and_cmt(source, ident_end)
+			if source:sub(after, after) == "(" then
+				local _, after_p = duffle.read_parens(source, after)
+				after = duffle.skip_ws_and_cmt(source, after_p or after)
+				if source:sub(after, after) == "{" then
+					local body, after_b = duffle.read_braces(source, after)
+					local names = collect_tb_emits(body or "", addrs)
+					if #names > 0 then
+						chains[#chains + 1] = names
+					end
+					pos = after_b or (after + 1)
+				else
+					pos = after
+				end
+			else
+				pos = ident_end
+			end
+		else
+			pos = pos + 1
+		end
+	end
+	return chains
+end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- The single source walker
 -- ════════════════════════════════════════════════════════════════════════════
@@ -2343,6 +2486,7 @@ local function scan_source(source, source_file, code_macros, code_macro_bodies)
 	-- Runs AFTER the source walk so all typedef / Struct_ / Enum_ declarations have been parsed into `out.type_name_registry`.
 	-- Mutates each entry's `byte_size` field in place; fields with pointer_depth > 0 already carry byte_size = 4 from parse time and are unaffected.
 	propagate_type_sizes(out)
+	out.tape_chains = scan_tape_chains(source)
 
 	return out
 end
@@ -2511,6 +2655,7 @@ local function merge_corpus_registries(corpus)
 	corpus.collisions              = corpus.collisions              or {}
 	corpus.reg_use_schemas         = corpus.reg_use_schemas         or {}
 	corpus.reg_use_errors          = corpus.reg_use_errors          or {}
+	corpus.tape_chains             = corpus.tape_chains             or {}
 
 	-- Replace the existing corpus collections with empty tables so a re-run on the same corpus produces identical state (deterministic merge).
 	-- This is safe because M.run is the only writer to these tables within a single orchestrator invocation.
@@ -2518,6 +2663,7 @@ local function merge_corpus_registries(corpus)
 		"register_alias_registry", "type_name_registry", "binds_by_name",
 		"atoms_by_name", "atom_views", "atom_ctxs", "atom_phases",
 		"atom_infos", "component_atom_infos", "collisions", "reg_use_schemas", "reg_use_errors",
+		"tape_chains",
 	}) do
 		corpus[key] = {}
 	end
@@ -2622,6 +2768,53 @@ local function merge_corpus_registries(corpus)
 			for _, err in ipairs(scan.reg_use_errors or {}) do
 				corpus.reg_use_errors[#corpus.reg_use_errors + 1] = err
 			end
+			for _, chain in ipairs(scan.tape_chains or {}) do
+				corpus.tape_chains[#corpus.tape_chains + 1] = chain
+			end
+		end
+	end
+end
+
+local SCHEMA_BODY_ERROR = {
+	reguse_malformed          = true,
+	reguse_unknown_reg_type   = true,
+	reguse_duplicate_alias    = true,
+	reguse_duplicate_slot     = true,
+	reguse_const_reg_spelling = true,
+	reguse_mixed_const        = true,
+}
+
+-- Re-parse every RegUse_* body against the merged type_name_registry.
+-- Scan-time expansion still runs when Reg_T is in the same source.
+-- Missing Reg_T after merge is reguse_unknown_reg_type, not a fallback table.
+local function resolve_reg_use_schemas(corpus)
+	local kept = {}
+	for _, err in ipairs(corpus.reg_use_errors or {}) do
+		if not SCHEMA_BODY_ERROR[err.kind] then
+			kept[#kept + 1] = err
+		end
+	end
+	corpus.reg_use_errors = kept
+
+	for name, type_entry in pairs(corpus.type_name_registry or {}) do
+		if name:sub(1, 7) == "RegUse_" and type_entry.body then
+			local fresh, errs = parse_reg_use_schema_body(
+				type_entry.body, corpus.type_name_registry, { require_types = true })
+			if fresh then
+				fresh.name = name
+				local old = corpus.reg_use_schemas[name]
+				fresh.source_file = (old and old.source_file) or type_entry.source_file
+				fresh.source_line = (old and old.source_line) or type_entry.source_line
+				corpus.reg_use_schemas[name] = fresh
+			else
+				corpus.reg_use_schemas[name] = nil
+			end
+			for _, err in ipairs(errs or {}) do
+				err.schema_name = name
+				err.source_file = type_entry.source_file
+				err.source_line = type_entry.source_line
+				corpus.reg_use_errors[#corpus.reg_use_errors + 1] = err
+			end
 		end
 	end
 end
@@ -2710,6 +2903,7 @@ function M.run(ctx)
 
 	-- Merge per-source scans into the corpus registries (see merge_corpus_registries for first-wins + collision discipline).
 	merge_corpus_registries(corpus)
+	resolve_reg_use_schemas(corpus)
 
 	-- code_macros and code_macro_bodies are function-local; the GC reclaims them on M.run return.
 	return { outputs = {}, errors = {}, warnings = {} }

scripts/passes/static_analysis.lua

@@ -340,7 +340,9 @@ local function classify_tokens(tokens)
 		local shape = ident:match("^mac_format_([%w_]+)_color$")
 		if shape then mac_format_shape = shape end
 		if ident:match("^mac_gte_store_[%w_]+$")  then is_gte_store = true end
-		if ident:match("^mac_insert_ot_tag_[%w_]+$") then is_ot_tag    = true end
+		if ident == "mac_insert_ot_tag" or ident:match("^mac_insert_ot_tag_[%w_]+$") then
+			is_ot_tag = true
+		end
 
 		-- O_(<arg1>, <arg2>) / S_(<arg>) captures (used by check_abi_handoff).
 		-- Cheap pattern match — anchored, fails fast on non-matching tokens.
@@ -1782,17 +1784,59 @@ local function check_yield_load_tail_pairing(atom, _pipe_ctx, findings)
 	end
 
 	-- ── Rule 2: `mac_yield_tail()` is valid if every path that reaches it already ran `mac_yield_load()`.
+	-- MIPS delay slot always runs. Successors skip the BD token for control flow,
+	-- but the yield walk still counts that token as executed.
 	local function load_covers_tail(tail_idx)
-		for i = 1, tail_idx - 1 do
-			if tc[i].ident == "mac_yield_load" then
-				local prev_i = skip_delay(i - 1, -1)
-				local prev   = prev_i and tc[prev_i] or nil
-				if prev and (prev.is_branch or prev.is_atom_label) then
-					return true
-				end
+		local labels = {}
+		for i = 1, n do
+			if tc[i].is_atom_label and tc[i].label_name then
+				labels[tc[i].label_name] = i
 			end
 		end
-		return false
+		local function is_load(idx)
+			return tc[idx] and tc[idx].ident == "mac_yield_load"
+		end
+		local reached_without = false
+		local reached_any = false
+		local path_n = 0
+		local MAX_PATHS = 64
+		local function dfs(idx, saw_load, visited)
+			if path_n >= MAX_PATHS then return end
+			if visited[idx] then return end
+			local vis = {}
+			for k, v in pairs(visited) do vis[k] = v end
+			vis[idx] = true
+			local saw = saw_load or is_load(idx)
+			if tc[idx].is_branch and idx + 1 <= n then
+				saw = saw or is_load(idx + 1)
+			end
+			if idx == tail_idx then
+				path_n = path_n + 1
+				reached_any = true
+				if not saw then reached_without = true end
+				return
+			end
+			if tc[idx].is_yield or tc[idx].is_terminal_jump then
+				return
+			end
+			if tc[idx].is_branch then
+				if not tc[idx].is_unconditional_jump and idx + 2 <= n then
+					dfs(idx + 2, saw, vis)
+				end
+				local label = tc[idx].branch_label
+				if label and labels[label] then
+					local dest = labels[label] + 1
+					if dest <= n then dfs(dest, saw, vis) end
+				end
+				return
+			end
+			if idx + 1 <= n then
+				dfs(idx + 1, saw, vis)
+			end
+		end
+		dfs(1, false, {})
+		if not reached_any then return true end
+		return not reached_without
 	end
 
 	for tok_idx = 1, n do
@@ -1951,6 +1995,7 @@ local function check_gpu_portstore_shape(atom, pipe_ctx, findings)
 	local contrib        = 0
 	local saw_format     = false
 	local saw_prim_write = false
+	local saw_tag        = false
 
 	-- Reads from tc_entry fields pre-computed by classify_tokens (R3 lift).
 	-- Eliminates 4 per-token string matches (mac_format_X_color + mac_gte_store_<shape> + mac_insert_ot_tag_<shape> + R_PrimCursor)
@@ -1981,30 +2026,55 @@ local function check_gpu_portstore_shape(atom, pipe_ctx, findings)
 			local comp    = pipe_ctx.components_by_name[bare]
 			local n       = comp and comp.gp0_contrib
 			if    n then contrib = contrib + n end
+			-- insert_ot_tag writes the packet tag. Count it once when the atom
+			-- has no raw O_(Poly_*, tag) store.
+			if not saw_tag then
+				contrib = contrib + 1
+				saw_tag = true
+			end
 		end
 		if tc_entry.writes_r_prim_cursor then
 			saw_prim_write = true
 		end
+		-- A raw store to O_(Poly_*, tag) is the packet tag word, counted once.
+		if tc_entry.o_arg1 and tc_entry.o_arg1:match("^Poly_") and tc_entry.o_arg2 == "tag" then
+			if tc_entry.is_store_word or tc_entry.ident == "gte_sw" then
+				if not saw_tag then
+					contrib = contrib + 1
+					saw_tag = true
+				end
+			end
+		end
 	end
 
-	-- Token-name gp0_contrib is 0 when bodies use gte_sw. Count expanded prim-buffer stores.
+	-- Token-name gp0_contrib is 0 when bodies use uncounted stores.
+	-- Once gte_sw is taught, prefer that sum. Do not also count every PrimCursor store.
 	if contrib == 0 then
+		local seen_field = {}
 		for _, ev in ipairs(atom.paths.word_events or {}) do
 			local enc = ev.encoder or ""
 			if enc == "store_word" or enc == "store_half" or enc == "store_byte" or enc == "gte_sw" then
 				local text = (ev.call_text or "") .. " " .. (ev.root_call_text or "")
-				local hit = text:find("R_PrimCursor", 1, true)
-				if not hit then
-					for _, arg in ipairs(ev.args or {}) do
-						if tostring(arg):find("R_PrimCursor", 1, true) then
-							hit = true
-							break
+				if text:find("insert_ot_tag", 1, true) then
+					-- OT list mutation, not a packet word.
+				else
+					local field = text:match("O_%(([^%)]+)%)") or text
+					if not seen_field[field] then
+						local hit = text:find("R_PrimCursor", 1, true)
+						if not hit then
+							for _, arg in ipairs(ev.args or {}) do
+								if tostring(arg):find("R_PrimCursor", 1, true) then
+									hit = true
+									break
+								end
+							end
+						end
+						if hit then
+							seen_field[field] = true
+							contrib = contrib + 1
 						end
 					end
 				end
-				if hit then
-					contrib = contrib + 1
-				end
 			end
 		end
 	end
@@ -2429,8 +2499,14 @@ local function check_binds_no_substruct_deref(_src, pipe_ctx, findings)
 				local body_line  = a.line + (line_in_body[tokens[ti].rel] or 0)
 
 				local type_entry = resolve_type_with_fields(type_name, type_registry, 1)
-				-- PSYQ opaques such as DisplayEnv have no fields table. Do not invent the layout.
-				local skip_opaque = type_name == "DisplayEnv" and (not type_entry or not type_entry.fields)
+				local no_fields = not type_entry or not type_entry.fields or #type_entry.fields == 0
+				local raw_entry = type_registry[type_name]
+				local is_typedef_to_struct = raw_entry
+					and raw_entry.kind == "typedef"
+					and raw_entry.underlying_type
+					and type_registry[raw_entry.underlying_type]
+					and type_registry[raw_entry.underlying_type].fields
+				local skip_opaque = no_fields and not is_typedef_to_struct
 				if skip_opaque then
 					-- leave this token
 				elseif not type_entry or not type_entry.fields then
@@ -2681,36 +2757,45 @@ local function check_gte_cr_TR_naming(atom, _pipe_ctx, findings)
 	end
 end
 
-local function check_gte_cr_alias_writes_xatom(_src, pipe_ctx, findings)
-	local slot_state = {}
-	for _, atom in ipairs(pipe_ctx.atoms or {}) do
-		atom.paths = atom.paths or {}
-		atom.paths.forward_state = atom.paths.forward_state or {}
-		local outgoing = {}
-		for slot, prev in pairs(slot_state) do
-			outgoing[slot] = prev
-		end
-		for _, w in ipairs(ctrl_writes_in_atom(atom)) do
-			local group = find_alias_pair_for(w.alias, duffle)
-			if group then
-				local slot = group[1]
-				local prev = slot_state[slot]
-				if prev and prev.alias ~= w.alias and prev.atom ~= atom.name then
-					findings[#findings + 1] = {
-						atom  = atom.name or "",
-						line  = w.line,
-						check = "gte_cr_alias_writes_xatom",
-						kind  = "warning",
-						msg   = string.format(
-							"atom '%s' writes %s to C2[%d]; atom '%s' already wrote %s"
-							, atom.name or "", w.alias, slot, prev.atom, prev.alias),
-					}
+local function check_gte_cr_alias_writes_xatom(src, pipe_ctx, findings)
+	-- Walk tape chains once (first source only). Atoms in no chain stay per-atom.
+	local first = pipe_ctx.source_order and pipe_ctx.source_order[1]
+	if first and src ~= first then return end
+	local atoms_by_name = pipe_ctx.atoms_by_name or {}
+	for _, chain in ipairs(pipe_ctx.tape_chains or {}) do
+		local slot_state = {}
+		for _, name in ipairs(chain) do
+			local atom = atoms_by_name[name]
+			if atom then
+				atom.paths = atom.paths or {}
+				atom.paths.forward_state = atom.paths.forward_state or {}
+				local outgoing = {}
+				for slot, prev in pairs(slot_state) do
+					outgoing[slot] = prev
 				end
-				slot_state[slot] = { alias = w.alias, atom = atom.name, line = w.line }
-				outgoing[slot] = slot_state[slot]
+				for _, w in ipairs(ctrl_writes_in_atom(atom)) do
+					local group = find_alias_pair_for(w.alias, duffle)
+					if group then
+						local slot = group[1]
+						local prev = slot_state[slot]
+						if prev and prev.alias ~= w.alias and prev.atom ~= atom.name then
+							findings[#findings + 1] = {
+								atom  = atom.name or "",
+								line  = w.line,
+								check = "gte_cr_alias_writes_xatom",
+								kind  = "warning",
+								msg   = string.format(
+									"atom '%s' writes %s to C2[%d]; atom '%s' already wrote %s"
+									, atom.name or "", w.alias, slot, prev.atom, prev.alias),
+							}
+						end
+						slot_state[slot] = { alias = w.alias, atom = atom.name, line = w.line }
+						outgoing[slot] = slot_state[slot]
+					end
+				end
+				atom.paths.forward_state.ctrl_writes_by_slot = outgoing
 			end
 		end
-		atom.paths.forward_state.ctrl_writes_by_slot = outgoing
 	end
 end
 
@@ -2738,74 +2823,103 @@ local function check_gte_packed_writes(atom, _pipe_ctx, findings)
 end
 
 local function check_ctc2_chain_source_preservation(atom, _pipe_ctx, findings)
-	local live = {}
-	local function mark_live(src, alias)
-		if src and alias and alias:match("^gte_cr_RT") then
-			live[src] = true
-		end
+	-- Fire only when a load sits before a later RT ctc2 that still names that GPR.
+	-- A load after the last RT ctc2 and before the command is a legal reload.
+	local events = (atom.paths and atom.paths.word_events) or {}
+	local function event_src(ev)
+		if ev.gpr_keys and ev.gpr_keys[1] then return ev.gpr_keys[1] end
+		local src = ev.args and ev.args[1]
+		if type(src) == "string" then src = src:match("^[%w_.]+") end
+		return src
 	end
-	for _, ev in ipairs((atom.paths and atom.paths.word_events) or {}) do
-		local enc = ev.encoder or ""
-		if enc == "gte_mv_to_ctrl_r" then
-			local src = ev.args and ev.args[1]
-			if type(src) == "string" then src = src:match("[%w_]+") end
-			local alias = ev.args and ev.args[2]
-			if type(alias) ~= "string" or not alias:match("^gte_cr_") then
-				alias = ctrl_alias_from_text(ev.call_text)
-			end
-			mark_live(src, alias)
-		elseif enc == "load_word" then
-			local dest = ev.args and ev.args[1]
-			if type(dest) == "string" then dest = dest:match("[%w_]+") end
-			if dest and live[dest] == true then
-				live[dest] = "clobbered"
-			end
-		elseif enc:match("^gte_cmdw_") then
-			for gpr, state in pairs(live) do
-				if state == "clobbered" then
-					findings[#findings + 1] = {
-						atom  = atom.name or "",
-						line  = ev.line or atom.line,
-						check = "ctc2_chain_source_preservation",
-						kind  = "warning",
-						msg   = string.format(
-							"atom '%s' reloads %s after ctc2 into RT and before %s"
-							, atom.name or "", gpr, enc),
-					}
-				end
-			end
-			live = {}
+	local function event_alias(ev)
+		local alias = ev.args and ev.args[2]
+		if type(alias) ~= "string" or not alias:match("^gte_cr_") then
+			alias = ctrl_alias_from_text(ev.call_text)
 		end
+		return alias
 	end
-	if not next((atom.paths and atom.paths.word_events) or {}) then
-		local pending = {}
-		for _, t in ipairs((atom.paths and atom.paths.tokens) or {}) do
-			local tok = t.tok or ""
-			local ident = tok:match("^([%w_]+)") or ""
-			if ident == "gte_mv_to_ctrl_r" then
-				mark_live(tok:match("%(%s*([%w_]+)"), ctrl_alias_from_text(tok))
-			elseif ident == "load_word" then
-				local dest = tok:match("%(%s*([%w_]+)")
-				if dest and live[dest] == true then live[dest] = "clobbered" end
-			elseif ident:match("^gte_cmdw_") then
-				for gpr, state in pairs(live) do
-					if state == "clobbered" then
-						pending[#pending + 1] = { gpr = gpr, enc = ident }
+	if #events > 0 then
+		for i, ev in ipairs(events) do
+			local enc = ev.encoder or ""
+			if enc == "load_word" then
+				local dest = event_src(ev)
+				if dest then
+					local earlier = false
+					for j = i - 1, 1, -1 do
+						local prev = events[j]
+						local prev_enc = prev.encoder or ""
+						if prev_enc:match("^gte_cmdw_") then break end
+						if prev_enc == "gte_mv_to_ctrl_r" then
+							local prev_src = event_src(prev)
+							local prev_alias = event_alias(prev)
+							if prev_src == dest and prev_alias and prev_alias:match("^gte_cr_RT") then
+								earlier = true
+								break
+							end
+						end
+					end
+					if earlier then
+						for j = i + 1, #events do
+							local later = events[j]
+							local later_enc = later.encoder or ""
+							if later_enc:match("^gte_cmdw_") then
+								break
+							end
+							if later_enc == "gte_mv_to_ctrl_r" then
+								local later_src = event_src(later)
+								local later_alias = event_alias(later)
+								if later_src == dest and later_alias and later_alias:match("^gte_cr_RT") then
+									findings[#findings + 1] = {
+										atom  = atom.name or "",
+										line  = ev.line or atom.line,
+										check = "ctc2_chain_source_preservation",
+										kind  = "warning",
+										msg   = string.format(
+											"atom '%s' reloads %s before a later ctc2 that still names it"
+											, atom.name or "", dest),
+									}
+									break
+								end
+							end
+						end
 					end
 				end
-				live = {}
 			end
 		end
-		for _, p in ipairs(pending) do
-			findings[#findings + 1] = {
-				atom  = atom.name or "",
-				line  = atom.line,
-				check = "ctc2_chain_source_preservation",
-				kind  = "warning",
-				msg   = string.format(
-					"atom '%s' reloads %s after ctc2 into RT and before %s"
-					, atom.name or "", p.gpr, p.enc),
-			}
+		return
+	end
+	local tokens = (atom.paths and atom.paths.tokens) or {}
+	for i, t in ipairs(tokens) do
+		local tok = t.tok or ""
+		local ident = tok:match("^([%w_]+)") or ""
+		if ident == "load_word" then
+			local dest = tok:match("%(%s*([%w_]+)")
+			if dest then
+				for j = i + 1, #tokens do
+					local later = tokens[j].tok or ""
+					local later_ident = later:match("^([%w_]+)") or ""
+					if later_ident:match("^gte_cmdw_") then
+						break
+					end
+					if later_ident == "gte_mv_to_ctrl_r" then
+						local later_src = later:match("%(%s*([%w_]+)")
+						local later_alias = ctrl_alias_from_text(later)
+						if later_src == dest and later_alias and later_alias:match("^gte_cr_RT") then
+							findings[#findings + 1] = {
+								atom  = atom.name or "",
+								line  = atom.line,
+								check = "ctc2_chain_source_preservation",
+								kind  = "warning",
+								msg   = string.format(
+									"atom '%s' reloads %s before a later ctc2 that still names it"
+									, atom.name or "", dest),
+							}
+							break
+						end
+					end
+				end
+			end
 		end
 	end
 end
@@ -2893,7 +3007,6 @@ local function check_immediate_field_width(atom, pipe_ctx, findings)
 					end
 				end
 			end
-			::continue_token::
 		end
 	end
 end
@@ -2901,6 +3014,167 @@ end
 	end
 end
 
+local SCRATCH_GPRS = {
+	R_T0 = true, R_T1 = true, R_T2 = true, R_T3 = true,
+	R_AT = true, R_V0 = true, R_V1 = true,
+}
+
+local function token_arg_list(tok)
+	local inner = (tok or ""):match("%b()")
+	if not inner then return {} end
+	return duffle.split_top_level_commas(inner:sub(2, -2))
+end
+
+local function arg_as_gpr(arg)
+	arg = duffle.trim(arg or "")
+	return arg:match("^R_[%w_]+$")
+end
+
+local function collect_gpr_traffic(tokens)
+	local reads, writes = {}, {}
+	for _, t in ipairs(tokens or {}) do
+		local tok = t.tok or t
+		local ident = (tok or ""):match("^([%w_]+)") or ""
+		if ident:sub(1, 4) ~= "mac_"
+			and not (duffle.DELAY_MARKERS and duffle.DELAY_MARKERS[ident])
+			and ident ~= "nop" and ident ~= "atom_label" and ident ~= "atom_offset"
+		then
+			local args = token_arg_list(tok)
+			local fx = (duffle.INSTRUCTION_GPR_EFFECTS or {})[ident]
+			if fx then
+				for _, pos in ipairs(fx.reads or {}) do
+					local g = arg_as_gpr(args[pos])
+					if g then reads[g] = true end
+				end
+				for _, pos in ipairs(fx.writes or {}) do
+					local g = arg_as_gpr(args[pos])
+					if g then writes[g] = true end
+				end
+			else
+				for _, arg in ipairs(args) do
+					local g = arg_as_gpr(arg)
+					if g then
+						reads[g] = true
+						writes[g] = true
+					end
+				end
+			end
+		end
+	end
+	return reads, writes
+end
+
+local function gpr_set_from_list(list)
+	local s = {}
+	for _, name in ipairs(list or {}) do
+		if type(name) == "string" then s[name] = true end
+	end
+	return s
+end
+
+local function gpr_set_eq(a, b)
+	for k in pairs(a) do if not b[k] then return false end end
+	for k in pairs(b) do if not a[k] then return false end end
+	return true
+end
+
+local function gpr_set_keys(s)
+	local keys = {}
+	for k in pairs(s) do keys[#keys + 1] = k end
+	table.sort(keys)
+	return keys
+end
+
+local function check_atom_calls_inferred_traffic(atom, pipe_ctx, findings)
+	if atom.kind ~= "atom" and atom.kind ~= "atom_proc" then return end
+	if is_runtime_helper(atom) then return end
+	local info = pipe_ctx.info_by_atom and pipe_ctx.info_by_atom[atom.name]
+	if not info then return end
+	if #(info.reads or {}) == 0 and #(info.writes or {}) == 0 then return end
+
+	local tokens = (atom.paths and atom.paths.tokens) or atom.body_tokens
+	local reads, writes = collect_gpr_traffic(tokens)
+	for _, t in ipairs(tokens or {}) do
+		local ident = ((t.tok or t) or ""):match("^([%w_]+)") or ""
+		if ident:sub(1, 4) == "mac_" then
+			local bare = ident:sub(5)
+			local idx = pipe_ctx.component_body_index and pipe_ctx.component_body_index[bare]
+			local comp = (pipe_ctx.components_by_name or {})[bare]
+				or (pipe_ctx.atoms_by_name or {})[bare]
+			local body_toks = (idx and idx.body_tokens)
+				or (comp and (comp.body_tokens or (comp.paths and comp.paths.tokens)))
+			local cr, cw = collect_gpr_traffic(body_toks)
+			for k in pairs(cr) do reads[k] = true end
+			for k in pairs(cw) do writes[k] = true end
+		end
+	end
+
+	local decl_r = gpr_set_from_list(info.reads)
+	local decl_w = gpr_set_from_list(info.writes)
+	local function keep_inferred(inferred, declared)
+		local out = {}
+		for k in pairs(inferred) do
+			if k == "R_0" then
+				if declared[k] then out[k] = true end
+			elseif SCRATCH_GPRS[k] then
+				if declared[k] then out[k] = true end
+			else
+				out[k] = true
+			end
+		end
+		return out
+	end
+	reads = keep_inferred(reads, decl_r)
+	writes = keep_inferred(writes, decl_w)
+	if not gpr_set_eq(decl_r, reads) or not gpr_set_eq(decl_w, writes) then
+		findings[#findings + 1] = {
+			atom  = atom.name,
+			line  = info.info_line or atom.line,
+			check = "atom_calls_inferred_traffic",
+			kind  = "warning",
+			msg   = string.format(
+				"atom '%s' declared [%s]/[%s] != inferred [%s]/[%s]",
+				atom.name,
+				table.concat(gpr_set_keys(decl_r), ","),
+				table.concat(gpr_set_keys(decl_w), ","),
+				table.concat(gpr_set_keys(reads), ","),
+				table.concat(gpr_set_keys(writes), ",")),
+		}
+	end
+end
+
+local function check_component_self_consistency(src, pipe_ctx, findings)
+	local first = pipe_ctx.source_order and pipe_ctx.source_order[1]
+	if first and src ~= first then return end
+	local infos = pipe_ctx.component_atom_infos or {}
+	local atoms_by_name = pipe_ctx.atoms_by_name or {}
+	for _, ai in ipairs(infos) do
+		local name = ai.atom_name or ai.name
+		local atom = name and atoms_by_name[name]
+		if atom and not atom.debug_skip then
+			local tokens = (atom.paths and atom.paths.tokens) or atom.body_tokens
+			local reads, writes = collect_gpr_traffic(tokens)
+			local decl_r = gpr_set_from_list(ai.reads)
+			local decl_w = gpr_set_from_list(ai.writes)
+			if not gpr_set_eq(decl_r, reads) or not gpr_set_eq(decl_w, writes) then
+				findings[#findings + 1] = {
+					atom  = name,
+					line  = ai.info_line or atom.line,
+					check = "component_self_consistency",
+					kind  = "warning",
+					msg   = string.format(
+						"component '%s' atom_reads/atom_writes [%s]/[%s] != body [%s]/[%s]",
+						name,
+						table.concat(gpr_set_keys(decl_r), ","),
+						table.concat(gpr_set_keys(decl_w), ","),
+						table.concat(gpr_set_keys(reads), ","),
+						table.concat(gpr_set_keys(writes), ",")),
+				}
+			end
+		end
+	end
+end
+
 -- CHECK_RULES — data-driven check dispatch (Muratori: data over control flow)
 -- ════════════════════════════════════════════════════════════════════════════
 
@@ -2937,6 +3211,8 @@ local CHECK_RULES = {
 	{ name = "enum_alias_membership",           per_source = check_enum_alias_membership           },
 	{ name = "atom_type_consistency",           per_source = check_atom_type_consistency           },
 	{ name = "binds_no_substruct_deref",        per_source = check_binds_no_substruct_deref        },
+	{ name = "component_self_consistency",      per_source = check_component_self_consistency      },
+	{ name = "atom_calls_inferred_traffic",     per_atom   = check_atom_calls_inferred_traffic     },
 }
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -2973,6 +3249,11 @@ local function build_corpus_pipe_ctx(ctx)
 		-- `MipsAtomComp_` body by `passes/components.lua::compute_components_metadata`.
 		-- Keyed by bare name (e.g. `format_f3_color`, `gte_store_f3`); the `mac_` prefix at call sites is stripped before lookup.
 		components_by_name      = corpus.components              or {},
+		atoms_by_name           = corpus.atoms_by_name           or {},
+		tape_chains             = corpus.tape_chains             or {},
+		source_order            = corpus.source_order            or {},
+		component_atom_infos    = corpus.component_atom_infos    or {},
+		atom_infos              = corpus.atom_infos              or {},
 		-- Corpus-wide ordered list of atom_info records (source-order + duplicates).
 		atom_infos_list         = corpus.atom_infos              or {},
 		-- Corpus-wide collisions (recorded by scan_source.merge_corpus_registries).
@@ -3031,6 +3312,11 @@ local function validate(ctx, src, corpus_pipe_ctx)
 		type_name_registry      = corpus_pipe_ctx.type_name_registry,
 		-- Per-component metadata (cycle_cost + gp0_contrib) auto-derived from the original `MipsAtomComp_` body by `passes/components.lua::compute_components_metadata`.
 		components_by_name      = corpus_pipe_ctx.components_by_name,
+		atoms_by_name           = corpus_pipe_ctx.atoms_by_name,
+		tape_chains             = corpus_pipe_ctx.tape_chains,
+		source_order            = corpus_pipe_ctx.source_order,
+		component_atom_infos    = corpus_pipe_ctx.component_atom_infos,
+		atom_infos_all          = corpus_pipe_ctx.atom_infos,
 	}
 	-- Shared cross-source component-body index is owned by the corpus (`corpus.component_body_index`, populated by `passes/components.lua`).
 	-- Per-atom checks consume the corpus-owned index directly.