diff --git a/runtime/pb/wire.lua b/runtime/pb/wire.lua
index 9482ba76201a18749d88c6479777acb04ffec593..6262045a766e3063d3e08a1b790d7b20326b4fc3 100644
--- a/runtime/pb/wire.lua
+++ b/runtime/pb/wire.lua
@@ -97,7 +97,22 @@     return encode_varint(field_id * 8 + wire_type)
 end
 M.encode_tag = encode_tag
 
+-- The 1-byte varint fast path is duplicated at every hot decode call
+-- site (decode_tag, decode_len, decode_int32 / int64 / uint32 / uint64 /
+-- sint32 / sint64 / bool, skip_field VARINT branch) instead of being
+-- factored into a helper. Reason: LuaJIT 2.1 inlines a small called
+-- function into the caller's trace, so when the parent guard exits to
+-- a side trace for the multi-byte case, the side trace has to return
+-- from the inlined frame — and LuaJIT can't stitch that return back
+-- to the parent, dropping to interpreter dispatch. Inlining the fast
+-- path literally keeps the side trace inside the caller's own frame,
+-- where stitching works.
 local function decode_tag(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then
+        return bit.rshift(b, 3), bit.band(b, 7), pos + 1
+    end
     local v, npos = decode_varint(buf, pos)
     v = tonumber(v)
     return bit.rshift(v, 3), bit.band(v, 7), npos
@@ -242,8 +257,16 @@ M.encode_len = encode_len
 
 -- decode_len(buf, pos) -> string, new_pos
 local function decode_len(buf, pos)
-    local len, npos = decode_varint(buf, pos)
-    len = tonumber(len)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    local len, npos
+    if b < 0x80 then
+        len = b; npos = pos + 1
+    else
+        local v
+        v, npos = decode_varint(buf, pos)
+        len = tonumber(v)
+    end
     if npos + len - 1 > #buf then error("truncated LEN payload", 0) end
     return buf:sub(npos, npos + len - 1), npos + len
 end
@@ -277,33 +300,58 @@ M.encode_bytes    = encode_len
 -- (encode_fixed32, encode_fixed64, encode_float, encode_double already on M)
 
 -- Decoders --------------------------------------------------------------------
+-- The 1-byte fast path is inlined at every varint-based scalar decoder
+-- (see comment above decode_tag). Each decoder reads the first byte,
+-- handles the common 0..127 case in straight-line code, and falls
+-- through to decode_varint only for multi-byte values.
 local function decode_int32(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then return b, pos + 1 end  -- 0..127 fits int32 directly
     local u, np = decode_varint(buf, pos)
     return tonumber(INT64(u)), np  -- truncated to int32 range via int64 sign-extension
 end
 local function decode_int64(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then return INT64(b), pos + 1 end
     local u, np = decode_varint(buf, pos)
     return INT64(u), np
 end
 local function decode_uint32(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then return b, pos + 1 end
     local u, np = decode_varint(buf, pos)
     local n = tonumber(UINT64(u))
     if n < 0 then n = n + 0x100000000 end
     return n, np
 end
 local function decode_uint64(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then return UINT64(b), pos + 1 end
     local u, np = decode_varint(buf, pos)
     return UINT64(u), np
 end
 local function decode_sint32(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then return zigzag_decode32(b), pos + 1 end
     local u, np = decode_varint(buf, pos)
     return zigzag_decode32(tonumber(u)), np
 end
 local function decode_sint64(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then return zigzag_decode64(b), pos + 1 end
     local u, np = decode_varint(buf, pos)
     return zigzag_decode64(u), np
 end
 local function decode_bool(buf, pos)
+    local b = buf:byte(pos)
+    if b == nil then error("truncated varint at offset " .. pos, 0) end
+    if b < 0x80 then return b ~= 0, pos + 1 end
     local u, np = decode_varint(buf, pos)
     return u ~= UINT64_ZERO, np
 end
@@ -363,11 +411,17 @@ -- skip_field(buf, pos, wire_type) -> new_pos
 -- ---------------------------------------------------------------------------
 local function skip_field(buf, pos, wire_type)
     if wire_type == M.WIRE_VARINT then
+        local b = buf:byte(pos)
+        if b == nil then error("truncated varint at offset " .. pos, 0) end
+        if b < 0x80 then return pos + 1 end
         local _, npos = decode_varint(buf, pos)
         return npos
     elseif wire_type == M.WIRE_I64 then
         return pos + 8
     elseif wire_type == M.WIRE_LEN then
+        local b = buf:byte(pos)
+        if b == nil then error("truncated varint at offset " .. pos, 0) end
+        if b < 0x80 then return pos + 1 + b end
         local len, npos = decode_varint(buf, pos)
         return npos + tonumber(len)
     elseif wire_type == M.WIRE_I32 then