Use MillisInternal in the scope and say setup() is not timed by the guard

Some work has no shorter form: bringing up a radio, the first connect of
a network stack, a key generation whose cost is the algorithm. Wrapping
it in this scope moves the loop pass start forward by its duration, so
the blocking warning keeps reporting everything else in the pass and the
component's threshold does not ratchet over it. The comment says what it
is for and that it must never hide code that could be made faster.

esphome/components/api/api_connection.cpp

@@ -2255,12 +2255,7 @@ bool APIConnection::send_message_(uint32_t payload_size, uint16_t message_type,
   // Capacity reserved above, cannot fail
   (void) shared_buf.resize(write_start + payload_size);
   ProtoWriteBuffer buffer{&shared_buf, write_start};
-  uint8_t *end = encode_fn(msg, buffer PROTO_ENCODE_DEBUG_INIT(&shared_buf));
-#ifdef ESPHOME_DEBUG_API
-  assert(end == shared_buf.data() + shared_buf.size());
-#else
-  (void) end;
-#endif
+  encode_fn(msg, buffer PROTO_ENCODE_DEBUG_INIT(&shared_buf));
   return this->send_buffer(ProtoWriteBuffer{&shared_buf}, message_type);
 }
 // encode_to_buffer is defined inline in api_connection.h (ESPHOME_ALWAYS_INLINE)

esphome/components/api/api_connection_buffer.h

@@ -46,13 +46,7 @@ inline uint16_t ESPHOME_ALWAYS_INLINE APIConnection::encode_to_buffer(uint32_t c
     return 0;
   }
   ProtoWriteBuffer buffer{&shared_buf, shared_buf.size() - calculated_size};
-  uint8_t *end = encode_fn(msg, buffer PROTO_ENCODE_DEBUG_INIT(&shared_buf));
-#ifdef ESPHOME_DEBUG_API
-  // A body that writes fewer bytes than calculate_size() promised would ship stale buffer bytes
-  assert(end == shared_buf.data() + shared_buf.size());
-#else
-  (void) end;
-#endif
+  encode_fn(msg, buffer PROTO_ENCODE_DEBUG_INIT(&shared_buf));
 
   return total_calculated_size;
 }

esphome/components/api/proto.h

@@ -287,31 +287,19 @@ class ProtoWriteBuffer {
   uint8_t *pos_;
 };
 
-// A four byte unaligned store is a memcpy call on ESP-IDF (-fno-builtin-memcpy) and on ARM cores without
-// unaligned access (Cortex-M0+, ARM9), so those targets share one outlined byte store helper per fixed32
-// field. Elsewhere the write inlines to a single store, or on ESP8266 to a few stores that measured
-// faster than a call, so it stays inline.
-#if defined(USE_ESP32) || (defined(__arm__) && !defined(__ARM_FEATURE_UNALIGNED))
-#define PROTO_OUTLINE_FOR_SIZE __attribute__((noinline))
-#define PROTO_FIXED32_BYTE_STORES true
-#else
-#define PROTO_OUTLINE_FOR_SIZE inline
-#define PROTO_FIXED32_BYTE_STORES false
-#endif
-
 // Varint encoding thresholds — used by both proto_encode_* free functions and ProtoSize.
 constexpr uint32_t VARINT_MAX_1_BYTE = 1 << 7;   // 128
 constexpr uint32_t VARINT_MAX_2_BYTE = 1 << 14;  // 16384
 
-/// Static encode helpers for the generated encode bodies. Each takes the write cursor by value and
-/// returns it advanced, so outlined calls at -Os chain through the return register instead of a
-/// stack slot. Helpers without a _force suffix skip fields holding the proto3 default.
+/// Static encode helpers for generated encode() functions.
+/// Generated code hoists buffer.pos_ into a local uint8_t *__restrict__ pos,
+/// then calls these methods which take pos by reference. No struct, no overhead.
+/// For sub-messages, pos is synced back to buffer before the call and reloaded after.
 class ProtoEncode {
  public:
   /// Write a multi-byte varint directly through a pos pointer.
   template<typename T>
-  [[nodiscard]] static inline uint8_t *encode_varint_raw_loop(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                              T value) {
+  static inline void encode_varint_raw_loop(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, T value) {
     do {
       PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
       *pos++ = static_cast<uint8_t>(value | 0x80);
@@ -319,49 +307,48 @@ class ProtoEncode {
     } while (value > 0x7F);
     PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
     *pos++ = static_cast<uint8_t>(value);
-    return pos;
   }
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  encode_varint_raw(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint32_t value) {
+  static inline void ESPHOME_ALWAYS_INLINE encode_varint_raw(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                             uint32_t value) {
     if (value < VARINT_MAX_1_BYTE) [[likely]] {
       PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
       *pos++ = static_cast<uint8_t>(value);
-      return pos;
+      return;
     }
-    return encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, value);
+    encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, value);
   }
   /// Encode a varint that is expected to be 1-2 bytes (e.g. zigzag RSSI, small lengths).
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  encode_varint_raw_short(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint32_t value) {
+  static inline void ESPHOME_ALWAYS_INLINE encode_varint_raw_short(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                                   uint32_t value) {
     if (value < VARINT_MAX_1_BYTE) [[likely]] {
       PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
       *pos++ = static_cast<uint8_t>(value);
-      return pos;
+      return;
     }
     if (value < VARINT_MAX_2_BYTE) [[likely]] {
       PROTO_ENCODE_CHECK_BOUNDS(pos, 2);
       *pos++ = static_cast<uint8_t>(value | 0x80);
       *pos++ = static_cast<uint8_t>(value >> 7);
-      return pos;
+      return;
     }
-    return encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, value);
+    encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, value);
   }
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  encode_varint_raw_64(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint64_t value) {
+  static inline void ESPHOME_ALWAYS_INLINE encode_varint_raw_64(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                                uint64_t value) {
     if (value < VARINT_MAX_1_BYTE) [[likely]] {
       PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
       *pos++ = static_cast<uint8_t>(value);
-      return pos;
+      return;
     }
-    return encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, value);
+    encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, value);
   }
   /// Encode a 48-bit MAC address (stored in a uint64) as varint.
   /// Real MAC addresses occupy the full 48 bits (OUI in upper 24), so the
   /// fast path -- any non-zero bit in the top 6 of 48 -- emits exactly 7 bytes
   /// with no per-byte branch. Falls back to the general loop otherwise.
   /// Caller must guarantee value fits in 48 bits (checked in debug builds).
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  encode_varint_raw_48bit(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint64_t value) {
+  static inline void ESPHOME_ALWAYS_INLINE encode_varint_raw_48bit(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                                   uint64_t value) {
 #ifdef ESPHOME_DEBUG_API
     assert(value < (1ULL << (MAC_ADDRESS_SIZE * 8)) && "encode_varint_raw_48bit: value exceeds 48 bits");
 #endif
@@ -376,39 +363,38 @@ class ProtoEncode {
       pos[4] = static_cast<uint8_t>((value >> 28) | 0x80);
       pos[5] = static_cast<uint8_t>((value >> 35) | 0x80);
       pos[6] = static_cast<uint8_t>(value >> 42);
-      return pos + 7;
+      pos += 7;
+      return;
     }
-    return encode_varint_raw_64(pos PROTO_ENCODE_DEBUG_ARG, value);
+    encode_varint_raw_64(pos PROTO_ENCODE_DEBUG_ARG, value);
   }
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  encode_field_raw(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id, uint32_t type) {
-    return encode_varint_raw(pos PROTO_ENCODE_DEBUG_ARG, (field_id << 3) | type);
+  static inline void ESPHOME_ALWAYS_INLINE encode_field_raw(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                            uint32_t field_id, uint32_t type) {
+    encode_varint_raw(pos PROTO_ENCODE_DEBUG_ARG, (field_id << 3) | type);
   }
   /// Write a single precomputed tag byte. Tag must be < 128.
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  write_raw_byte(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint8_t b) {
+  static inline void ESPHOME_ALWAYS_INLINE write_raw_byte(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                          uint8_t b) {
     PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
     *pos++ = b;
-    return pos;
   }
   /// Reserve one byte for later backpatch (e.g., sub-message length).
   /// Advances pos past the reserved byte without writing a value.
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  reserve_byte(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM) {
+  static inline void ESPHOME_ALWAYS_INLINE reserve_byte(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM) {
     PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
-    return pos + 1;
+    pos++;
   }
   /// Write raw bytes to the buffer (no tag, no length prefix).
-  [[nodiscard]] static inline uint8_t *ESPHOME_ALWAYS_INLINE
-  encode_raw(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, const void *data, size_t len) {
+  static inline void ESPHOME_ALWAYS_INLINE encode_raw(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                      const void *data, size_t len) {
     PROTO_ENCODE_CHECK_BOUNDS(pos, len);
     std::memcpy(pos, data, len);
-    return pos + len;
+    pos += len;
   }
   /// Encode tag + 1-byte length + raw string data. For strings with max_data_length < 128.
   /// Tag must be a single-byte varint (< 128). Always encodes (no zero check).
-  [[nodiscard]] static inline uint8_t *encode_short_string_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                                 uint8_t tag, const StringRef &ref) {
+  static inline void encode_short_string_force(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint8_t tag,
+                                               const StringRef &ref) {
 #ifdef ESPHOME_DEBUG_API
     assert(ref.size() < 128 && "encode_short_string_force: string exceeds max_data_length < 128");
 #endif
@@ -416,191 +402,137 @@ class ProtoEncode {
     pos[0] = tag;
     pos[1] = static_cast<uint8_t>(ref.size());
     std::memcpy(pos + 2, ref.c_str(), ref.size());
-    return pos + 2 + ref.size();
+    pos += 2 + ref.size();
   }
-  /// Write a precomputed tag byte + 32-bit value. Outlined on embedded: one copy beats inline stores per field.
-  [[nodiscard]] static PROTO_OUTLINE_FOR_SIZE uint8_t *write_tag_and_fixed32(
-      uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint8_t tag, uint32_t value) {
+  /// Write a precomputed tag byte + 32-bit value in one operation.
+  static inline void ESPHOME_ALWAYS_INLINE write_tag_and_fixed32(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                                 uint8_t tag, uint32_t value) {
     PROTO_ENCODE_CHECK_BOUNDS(pos, 5);
     pos[0] = tag;
-    write_fixed32_le(pos + 1, value);
-    return pos + 5;
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    std::memcpy(pos + 1, &value, 4);
+#else
+    pos[1] = static_cast<uint8_t>(value & 0xFF);
+    pos[2] = static_cast<uint8_t>((value >> 8) & 0xFF);
+    pos[3] = static_cast<uint8_t>((value >> 16) & 0xFF);
+    pos[4] = static_cast<uint8_t>((value >> 24) & 0xFF);
+#endif
+    pos += 5;
   }
-  [[nodiscard]] static inline uint8_t *encode_string_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                           uint32_t field_id, const char *string, size_t len) {
-    pos = encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 2);  // type 2: Length-delimited string
+  static inline void encode_string(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                   const char *string, size_t len, bool force = false) {
+    if (len == 0 && !force)
+      return;
+    encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 2);  // type 2: Length-delimited string
     // NOLINTNEXTLINE(readability-inconsistent-ifelse-braces) -- false positive on [[likely]] attribute
     if (len < VARINT_MAX_1_BYTE) [[likely]] {
       PROTO_ENCODE_CHECK_BOUNDS(pos, 1 + len);
       *pos++ = static_cast<uint8_t>(len);
     } else {
-      pos = encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, len);
+      encode_varint_raw_loop(pos PROTO_ENCODE_DEBUG_ARG, len);
       PROTO_ENCODE_CHECK_BOUNDS(pos, len);
     }
     std::memcpy(pos, string, len);
-    return pos + len;
+    pos += len;
   }
-  [[nodiscard]] static inline uint8_t *encode_string(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                     uint32_t field_id, const char *string, size_t len) {
-    if (len == 0)
-      return pos;
-    return encode_string_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, string, len);
+  static inline void encode_string(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                   const std::string &value, bool force = false) {
+    encode_string(pos PROTO_ENCODE_DEBUG_ARG, field_id, value.data(), value.size(), force);
   }
-  [[nodiscard]] static inline uint8_t *encode_string_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                           uint32_t field_id, const std::string &value) {
-    return encode_string_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, value.data(), value.size());
+  static inline void encode_string(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                   const StringRef &ref, bool force = false) {
+    encode_string(pos PROTO_ENCODE_DEBUG_ARG, field_id, ref.c_str(), ref.size(), force);
   }
-  [[nodiscard]] static inline uint8_t *encode_string(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                     uint32_t field_id, const StringRef &ref) {
-    return encode_string(pos PROTO_ENCODE_DEBUG_ARG, field_id, ref.c_str(), ref.size());
+  static inline void encode_bytes(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                  const uint8_t *data, size_t len, bool force = false) {
+    encode_string(pos PROTO_ENCODE_DEBUG_ARG, field_id, reinterpret_cast<const char *>(data), len, force);
   }
-  [[nodiscard]] static inline uint8_t *encode_string_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                           uint32_t field_id, const StringRef &ref) {
-    return encode_string_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, ref.c_str(), ref.size());
+  static inline void encode_uint32(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                   uint32_t value, bool force = false) {
+    if (value == 0 && !force)
+      return;
+    encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 0);
+    encode_varint_raw(pos PROTO_ENCODE_DEBUG_ARG, value);
   }
-  [[nodiscard]] static inline uint8_t *encode_bytes(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                    uint32_t field_id, const uint8_t *data, size_t len) {
-    return encode_string(pos PROTO_ENCODE_DEBUG_ARG, field_id, reinterpret_cast<const char *>(data), len);
+  static inline void encode_uint64(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                   uint64_t value, bool force = false) {
+    if (value == 0 && !force)
+      return;
+    encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 0);
+    encode_varint_raw_64(pos PROTO_ENCODE_DEBUG_ARG, value);
   }
-  [[nodiscard]] static inline uint8_t *encode_bytes_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                          uint32_t field_id, const uint8_t *data, size_t len) {
-    return encode_string_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, reinterpret_cast<const char *>(data), len);
-  }
-  [[nodiscard]] static inline uint8_t *encode_uint32_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                           uint32_t field_id, uint32_t value) {
-    pos = encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 0);
-    return encode_varint_raw(pos PROTO_ENCODE_DEBUG_ARG, value);
-  }
-  [[nodiscard]] static inline uint8_t *encode_uint32(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                     uint32_t field_id, uint32_t value) {
-    if (value == 0)
-      return pos;
-    return encode_uint32_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, value);
-  }
-  [[nodiscard]] static inline uint8_t *encode_uint64_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                           uint32_t field_id, uint64_t value) {
-    pos = encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 0);
-    return encode_varint_raw_64(pos PROTO_ENCODE_DEBUG_ARG, value);
-  }
-  [[nodiscard]] static inline uint8_t *encode_uint64(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                     uint32_t field_id, uint64_t value) {
-    if (value == 0)
-      return pos;
-    return encode_uint64_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, value);
-  }
-  [[nodiscard]] static inline uint8_t *encode_bool_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                         uint32_t field_id, bool value) {
-    pos = encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 0);
+  static inline void encode_bool(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id, bool value,
+                                 bool force = false) {
+    if (!value && !force)
+      return;
+    encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 0);
     PROTO_ENCODE_CHECK_BOUNDS(pos, 1);
     *pos++ = value ? 0x01 : 0x00;
-    return pos;
   }
-  [[nodiscard]] static inline uint8_t *encode_bool(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                   uint32_t field_id, bool value) {
-    if (!value)
-      return pos;
-    return encode_bool_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, value);
-  }
-  /// Tag + fixed32 for multi-byte tags; single-byte tags use write_tag_and_fixed32.
-  [[nodiscard]] static PROTO_OUTLINE_FOR_SIZE uint8_t *encode_fixed32_force(
-      uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id, uint32_t value) {
-    pos = encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 5);
+  static inline void encode_fixed32(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                    uint32_t value, bool force = false) {
+    if (value == 0 && !force)
+      return;
+    encode_field_raw(pos PROTO_ENCODE_DEBUG_ARG, field_id, 5);
     PROTO_ENCODE_CHECK_BOUNDS(pos, 4);
-    write_fixed32_le(pos, value);
-    return pos + 4;
-  }
-  [[nodiscard]] static inline uint8_t *encode_fixed32(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                      uint32_t field_id, uint32_t value) {
-    if (value == 0)
-      return pos;
-    return encode_fixed32_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, value);
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    std::memcpy(pos, &value, 4);
+    pos += 4;
+#else
+    *pos++ = (value >> 0) & 0xFF;
+    *pos++ = (value >> 8) & 0xFF;
+    *pos++ = (value >> 16) & 0xFF;
+    *pos++ = (value >> 24) & 0xFF;
+#endif
   }
   // NOTE: Wire type 1 (64-bit fixed: double, fixed64, sfixed64) is intentionally
   // not supported to reduce overhead on embedded systems. All ESPHome devices are
   // 32-bit microcontrollers where 64-bit operations are expensive. If 64-bit support
   // is needed in the future, the necessary encoding/decoding functions must be added.
-  [[nodiscard]] static inline uint8_t *encode_float(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                    uint32_t field_id, float value) {
-    return encode_fixed32(pos PROTO_ENCODE_DEBUG_ARG, field_id, float_to_raw(value));
+  static inline void encode_float(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id, float value,
+                                  bool force = false) {
+    uint32_t raw = float_to_raw(value);
+    if (raw == 0 && !force)
+      return;
+    encode_fixed32(pos PROTO_ENCODE_DEBUG_ARG, field_id, raw);
   }
-  [[nodiscard]] static inline uint8_t *encode_float_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                          uint32_t field_id, float value) {
-    return encode_fixed32_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, float_to_raw(value));
-  }
-  [[nodiscard]] static inline uint8_t *encode_int32_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                          uint32_t field_id, int32_t value) {
+  static inline void encode_int32(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id, int32_t value,
+                                  bool force = false) {
     if (value < 0) {
       // negative int32 is always 10 byte long
-      return encode_uint64_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, static_cast<uint64_t>(value));
+      encode_uint64(pos PROTO_ENCODE_DEBUG_ARG, field_id, static_cast<uint64_t>(value), force);
+      return;
     }
-    return encode_uint32_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, static_cast<uint32_t>(value));
+    encode_uint32(pos PROTO_ENCODE_DEBUG_ARG, field_id, static_cast<uint32_t>(value), force);
   }
-  [[nodiscard]] static inline uint8_t *encode_int32(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                    uint32_t field_id, int32_t value) {
-    if (value == 0)
-      return pos;
-    return encode_int32_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, value);
+  static inline void encode_int64(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id, int64_t value,
+                                  bool force = false) {
+    encode_uint64(pos PROTO_ENCODE_DEBUG_ARG, field_id, static_cast<uint64_t>(value), force);
   }
-  [[nodiscard]] static inline uint8_t *encode_int64(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                    uint32_t field_id, int64_t value) {
-    return encode_uint64(pos PROTO_ENCODE_DEBUG_ARG, field_id, static_cast<uint64_t>(value));
+  static inline void encode_sint32(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                   int32_t value, bool force = false) {
+    encode_uint32(pos PROTO_ENCODE_DEBUG_ARG, field_id, encode_zigzag32(value), force);
   }
-  [[nodiscard]] static inline uint8_t *encode_int64_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                          uint32_t field_id, int64_t value) {
-    return encode_uint64_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, static_cast<uint64_t>(value));
+  static inline void encode_sint64(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, uint32_t field_id,
+                                   int64_t value, bool force = false) {
+    encode_uint64(pos PROTO_ENCODE_DEBUG_ARG, field_id, encode_zigzag64(value), force);
   }
-  [[nodiscard]] static inline uint8_t *encode_sint32(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                     uint32_t field_id, int32_t value) {
-    return encode_uint32(pos PROTO_ENCODE_DEBUG_ARG, field_id, encode_zigzag32(value));
-  }
-  [[nodiscard]] static inline uint8_t *encode_sint32_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                           uint32_t field_id, int32_t value) {
-    return encode_uint32_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, encode_zigzag32(value));
-  }
-  [[nodiscard]] static inline uint8_t *encode_sint64(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                     uint32_t field_id, int64_t value) {
-    return encode_uint64(pos PROTO_ENCODE_DEBUG_ARG, field_id, encode_zigzag64(value));
-  }
-  [[nodiscard]] static inline uint8_t *encode_sint64_force(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                           uint32_t field_id, int64_t value) {
-    return encode_uint64_force(pos PROTO_ENCODE_DEBUG_ARG, field_id, encode_zigzag64(value));
-  }
-  /// Sub-message encoding: sync pos to buffer, delegate, read the cursor back.
+  /// Sub-message encoding: sync pos to buffer, delegate, get pos from return value.
   template<typename T>
-  [[nodiscard]] static inline uint8_t *encode_sub_message(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                          ProtoWriteBuffer &buffer, uint32_t field_id, const T &value) {
+  static inline void encode_sub_message(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM, ProtoWriteBuffer &buffer,
+                                        uint32_t field_id, const T &value) {
     buffer.set_pos(pos);
     buffer.encode_sub_message(field_id, value);
-    return buffer.get_pos();
+    pos = buffer.get_pos();
   }
   template<typename T>
-  [[nodiscard]] static inline uint8_t *encode_optional_sub_message(uint8_t *__restrict__ pos PROTO_ENCODE_DEBUG_PARAM,
-                                                                   ProtoWriteBuffer &buffer, uint32_t field_id,
-                                                                   const T &value) {
+  static inline void encode_optional_sub_message(uint8_t *__restrict__ &pos PROTO_ENCODE_DEBUG_PARAM,
+                                                 ProtoWriteBuffer &buffer, uint32_t field_id, const T &value) {
     buffer.set_pos(pos);
     buffer.encode_optional_sub_message(field_id, value);
-    return buffer.get_pos();
-  }
-
- private:
-  /// Unaligned little endian store of four bytes: byte stores where the outlined helper lives (ESP-IDF, ARM
-  /// without unaligned access), otherwise a memcpy the compiler folds into one store. Callers bounds check
-  /// and advance the cursor themselves.
-  static inline void ESPHOME_ALWAYS_INLINE write_fixed32_le(uint8_t *__restrict__ pos, uint32_t value) {
-    if constexpr (PROTO_FIXED32_BYTE_STORES) {
-      // Spelled out so the outlined helper does not itself become a memcpy call
-      pos[0] = static_cast<uint8_t>(value);
-      pos[1] = static_cast<uint8_t>(value >> 8);
-      pos[2] = static_cast<uint8_t>(value >> 16);
-      pos[3] = static_cast<uint8_t>(value >> 24);
-    } else {
-      const uint32_t le = convert_little_endian(value);
-      __builtin_memcpy(pos, &le, 4);
-    }
+    pos = buffer.get_pos();
   }
 };
-#undef PROTO_OUTLINE_FOR_SIZE
-#undef PROTO_FIXED32_BYTE_STORES
 
 #ifdef HAS_PROTO_MESSAGE_DUMP
 /**

esphome/core/application.h

@@ -389,6 +389,7 @@ class Application {
   friend Component;
   friend class Scheduler;
   friend class LoopBlockingGuard;
+  friend class UnavoidableBlockingScope;
 #ifdef USE_RUNTIME_STATS
   friend class runtime_stats::RuntimeStatsCollector;
 #endif
@@ -631,6 +632,55 @@ class LoopBlockingGuard {
   static void __attribute__((noinline, cold)) warn_blocking(uint32_t blocking_time);
 };
 
+/// Leaves a stretch of the current loop pass out of the blocking warning.
+///
+/// Only for work done from a loop pass that cannot be made shorter and
+/// cannot be split across passes: turning on a radio, the first Wi-Fi
+/// connect, a key generation whose cost is the algorithm itself. The warning
+/// then keeps reporting everything else in the pass, and the component's
+/// threshold does not ratchet up over the one step nothing can be done about.
+///
+/// Never use it to paper over a problem that can be solved. A slow driver
+/// call, a loop that could be a state machine, a computation that could be
+/// cached or deferred, a blocking read that could be polled: those are what
+/// the warning exists to find, and wrapping them in this scope hides the
+/// bug instead of fixing it. If in doubt, leave the warning in.
+///
+/// Only work timed by a LoopBlockingGuard is affected, that is a component's
+/// loop() or a scheduler callback; setup() is not timed by the guard, so the
+/// scope has no effect on the warning there. Main loop task only. The watchdog is not fed inside the
+/// scope, so the work must finish within the watchdog timeout, or be paired
+/// with a watchdog::WatchdogManager that raises the timeout for the same
+/// stretch. Scopes may nest; the outermost one decides how much of the pass
+/// is left out.
+/// App.get_loop_component_start_time() reads later in the same pass return
+/// the moved start, so elapsed time across the scope needs millis().
+///
+///   void MyComponent::loop() {
+///     if (this->needs_key_) {
+///       UnavoidableBlockingScope scope;
+///       this->generate_key_();
+///     }
+///   }
+class UnavoidableBlockingScope {
+ public:
+  UnavoidableBlockingScope() : started_(MillisInternal::get()), pass_start_(App.get_loop_component_start_time()) {}
+  ~UnavoidableBlockingScope() {
+    // Move the pass start seen at entry forward by the time spent here, so an
+    // outer scope overrides an inner one instead of adding to it; never past
+    // now, which would underflow the guard's subtraction
+    const uint32_t now = MillisInternal::get();
+    const uint32_t moved = this->pass_start_ + (now - this->started_);
+    App.set_loop_component_start_time_(static_cast<int32_t>(now - moved) < 0 ? now : moved);
+  }
+  UnavoidableBlockingScope(const UnavoidableBlockingScope &) = delete;
+  UnavoidableBlockingScope &operator=(const UnavoidableBlockingScope &) = delete;
+
+ private:
+  uint32_t started_;
+  uint32_t pass_start_;
+};
+
 // Phase A: drain wake notifications and run the scheduler. Invoked on every
 // Application::loop() tick regardless of whether a component phase runs, so
 // scheduler items fire at their requested cadence even when the caller has

esphome/core/millis_internal.h

@@ -51,6 +51,7 @@ class MillisInternal {
   }
   friend class Application;
   friend class LoopBlockingGuard;
+  friend class UnavoidableBlockingScope;
 };
 
 }  // namespace esphome

script/api_protobuf/api_protobuf.py

@@ -131,12 +131,6 @@ def force_str(force: bool) -> str:
     return str(force).lower()
 
 
-def _encode_call(func: str, *args: str, force: bool = False) -> str:
-    """Emit one ProtoEncode call; every helper takes the cursor and returns it advanced."""
-    suffix = "_force" if force else ""
-    return f"pos = ProtoEncode::{func}{suffix}({', '.join(('pos', *args))});"
-
-
 class TypeInfo(ABC):
     """Base class for all type information."""
 
@@ -270,16 +264,14 @@ class TypeInfo(ABC):
     # write_raw_byte(tag) + raw encode instead of the full encode_* method,
     # eliminating the zero-check branch and encode_field_raw indirection.
     # {value} is replaced with the actual field expression.
-    RAW_ENCODE_MAP: dict[str, tuple[str, str]] = {
-        "encode_uint32": ("encode_varint_raw", "{value}"),
-        "encode_uint64": ("encode_varint_raw_64", "{value}"),
-        "encode_sint32": ("encode_varint_raw_short", "encode_zigzag32({value})"),
-        "encode_sint64": ("encode_varint_raw_64", "encode_zigzag64({value})"),
-        "encode_int64": ("encode_varint_raw_64", "static_cast<uint64_t>({value})"),
-        "encode_bool": ("write_raw_byte", "{value} ? 0x01 : 0x00"),
+    RAW_ENCODE_MAP: dict[str, str] = {
+        "encode_uint32": "ProtoEncode::encode_varint_raw(pos, {value});",
+        "encode_uint64": "ProtoEncode::encode_varint_raw_64(pos, {value});",
+        "encode_sint32": "ProtoEncode::encode_varint_raw_short(pos, encode_zigzag32({value}));",
+        "encode_sint64": "ProtoEncode::encode_varint_raw_64(pos, encode_zigzag64({value}));",
+        "encode_int64": "ProtoEncode::encode_varint_raw_64(pos, static_cast<uint64_t>({value}));",
+        "encode_bool": "ProtoEncode::write_raw_byte(pos, {value} ? 0x01 : 0x00);",
     }
-    # Fixed32 value expression for the shared tag+fixed32 writer; None for other wire types
-    fixed32_value_template: str | None = None
 
     def _encode_with_precomputed_tag(self, value_expr: str) -> str | None:
         """Try to emit a precomputed-tag encode for a field.
@@ -296,17 +288,12 @@ class TypeInfo(ABC):
             return None
         max_val = self.max_value
         # Only use RAW_ENCODE_MAP for forced fields or fields with max_value
-        raw = None
+        raw_expr = None
         if self.force or max_val is not None:
-            raw = self.RAW_ENCODE_MAP.get(self.encode_func)
-        if raw is None:
+            raw_expr = self.RAW_ENCODE_MAP.get(self.encode_func)
+        if raw_expr is None:
             return None
-        func, arg = raw
-        body = (
-            _encode_call("write_raw_byte", str(tag))
-            + "\n"
-            + _encode_call(func, arg.format(value=value_expr))
-        )
+        body = f"ProtoEncode::write_raw_byte(pos, {tag});\n{raw_expr.format(value=value_expr)}"
         if self.force:
             return body
         # Non-forced with max_value: inline zero-check + raw encode
@@ -327,43 +314,23 @@ class TypeInfo(ABC):
             return None
         # When max_len < 128, length varint is always 1 byte
         len_encode = (
-            _encode_call("write_raw_byte", f"static_cast<uint8_t>({len_expr})")
+            f"ProtoEncode::write_raw_byte(pos, static_cast<uint8_t>({len_expr}));"
             if max_len is not None and max_len < 128
-            else _encode_call("encode_varint_raw", len_expr)
+            else f"ProtoEncode::encode_varint_raw(pos, {len_expr});"
         )
-        return "\n".join(
-            (
-                _encode_call("write_raw_byte", str(tag)),
-                len_encode,
-                _encode_call("encode_raw", data_expr, len_expr),
-            )
-        )
-
-    def _encode_fixed32_with_precomputed_tag(self, value_expr: str) -> str | None:
-        """Single-byte tag fixed32 write, or None for multi-byte tags."""
-        tag = self.calculate_tag()
-        if tag >= 128:
-            return None
-        if self.force:
-            return _encode_call("write_tag_and_fixed32", str(tag), value_expr)
         return (
-            f"if (uint32_t raw = {value_expr}; raw != 0) [[likely]] {{\n"
-            f"  {_encode_call('write_tag_and_fixed32', str(tag), 'raw')}\n"
-            "}"
+            f"ProtoEncode::write_raw_byte(pos, {tag});\n"
+            f"{len_encode}\n"
+            f"ProtoEncode::encode_raw(pos, {data_expr}, {len_expr});"
         )
 
     @property
     def encode_content(self) -> str:
-        value = f"this->{self.field_name}"
-        if result := self._encode_with_precomputed_tag(value):
+        if result := self._encode_with_precomputed_tag(f"this->{self.field_name}"):
             return result
-        if self.fixed32_value_template is not None and (
-            result := self._encode_fixed32_with_precomputed_tag(
-                self.fixed32_value_template.format(value=value)
-            )
-        ):
-            return result
-        return _encode_call(self.encode_func, str(self.number), value, force=self.force)
+        if self.force:
+            return f"ProtoEncode::{self.encode_func}(pos, {self.number}, this->{self.field_name}, true);"
+        return f"ProtoEncode::{self.encode_func}(pos, {self.number}, this->{self.field_name});"
 
     encode_func = None
 
@@ -668,8 +635,6 @@ class FloatType(FixedSizeTypeMixin, TypeInfo):
     encode_func = "encode_float"
     wire_type = WireType.FIXED32  # Uses wire type 5
 
-    fixed32_value_template = "float_to_raw({value})"
-
     def dump(self, name: str) -> str:
         o = f'snprintf(buffer, sizeof(buffer), "%g", {name});\n'
         o += "out.append(buffer);"
@@ -732,11 +697,11 @@ class UInt64Type(VarintTypeMixin, TypeInfo):
         return self._get_simple_size_calculation(name, force, "uint64")
 
     @property
-    def RAW_ENCODE_MAP(self) -> dict[str, tuple[str, str]]:  # noqa: N802
+    def RAW_ENCODE_MAP(self) -> dict[str, str]:  # noqa: N802
         if self.mac_address:
             return {
                 **TypeInfo.RAW_ENCODE_MAP,
-                "encode_uint64": ("encode_varint_raw_48bit", "{value}"),
+                "encode_uint64": "ProtoEncode::encode_varint_raw_48bit(pos, {value});",
             }
         return TypeInfo.RAW_ENCODE_MAP
 
@@ -804,7 +769,15 @@ class Fixed32Type(FixedSizeTypeMixin, TypeInfo):
         o += "out.append(buffer);"
         return o
 
-    fixed32_value_template = "{value}"
+    @property
+    def encode_content(self) -> str:
+        tag = self.calculate_tag()
+        if self.force and tag < 128:
+            # Emit combined tag+value write: precomputed tag + direct memcpy
+            return f"ProtoEncode::write_tag_and_fixed32(pos, {tag}, this->{self.field_name});"
+        if self.force:
+            return f"ProtoEncode::{self.encode_func}(pos, {self.number}, this->{self.field_name}, true);"
+        return f"ProtoEncode::{self.encode_func}(pos, {self.number}, this->{self.field_name});"
 
     def get_size_calculation(self, name: str, force: bool = False) -> str:
         field_id_size = self.calculate_field_id_size()
@@ -878,12 +851,9 @@ class StringType(TypeInfo):
             f"this->{self.field_name}_ref_.size()",
         ):
             return result
-        return _encode_call(
-            "encode_string",
-            str(self.number),
-            f"this->{self.field_name}_ref_",
-            force=self.force,
-        )
+        if self.force:
+            return f"ProtoEncode::encode_string(pos, {self.number}, this->{self.field_name}_ref_, true);"
+        return f"ProtoEncode::encode_string(pos, {self.number}, this->{self.field_name}_ref_);"
 
     def dump(self, name):
         # If name is 'it', this is a repeated field element - always use string
@@ -981,9 +951,7 @@ class MessageType(TypeInfo):
     @property
     def encode_content(self) -> str:
         # Sub-message encoding needs buffer for backpatch/sync
-        return _encode_call(
-            self.encode_func, "buffer", str(self.number), f"this->{self.field_name}"
-        )
+        return f"ProtoEncode::{self.encode_func}(pos, buffer, {self.number}, this->{self.field_name});"
 
     @property
     def decode_length(self) -> str:
@@ -1090,13 +1058,9 @@ class BytesType(TypeInfo):
             f"this->{self.field_name}_ptr_", f"this->{self.field_name}_len_"
         ):
             return result
-        return _encode_call(
-            "encode_bytes",
-            str(self.number),
-            f"this->{self.field_name}_ptr_",
-            f"this->{self.field_name}_len_",
-            force=self.force,
-        )
+        if self.force:
+            return f"ProtoEncode::encode_bytes(pos, {self.number}, this->{self.field_name}_ptr_, this->{self.field_name}_len_, true);"
+        return f"ProtoEncode::encode_bytes(pos, {self.number}, this->{self.field_name}_ptr_, this->{self.field_name}_len_);"
 
     def dump(self, name: str) -> str:
         ptr_dump = f"format_hex_pretty(this->{self.field_name}_ptr_, this->{self.field_name}_len_)"
@@ -1206,13 +1170,9 @@ class PointerToBytesBufferType(PointerToBufferTypeBase):
             f"this->{self.field_name}", f"this->{self.field_name}_len"
         ):
             return result
-        return _encode_call(
-            "encode_bytes",
-            str(self.number),
-            f"this->{self.field_name}",
-            f"this->{self.field_name}_len",
-            force=self.force,
-        )
+        if self.force:
+            return f"ProtoEncode::encode_bytes(pos, {self.number}, this->{self.field_name}, this->{self.field_name}_len, true);"
+        return f"ProtoEncode::encode_bytes(pos, {self.number}, this->{self.field_name}, this->{self.field_name}_len);"
 
     @property
     def decode_length_content(self) -> str | None:
@@ -1264,19 +1224,16 @@ class PointerToStringBufferType(PointerToBufferTypeBase):
         if max_len is not None and max_len < 128 and self.force:
             tag = self.calculate_tag()
             if tag < 128:
-                return _encode_call(
-                    "encode_short_string_force", str(tag), f"this->{self.field_name}"
-                )
+                return f"ProtoEncode::encode_short_string_force(pos, {tag}, this->{self.field_name});"
         if result := self._encode_bytes_with_precomputed_tag(
             f"this->{self.field_name}.c_str()",
             f"this->{self.field_name}.size()",
         ):
             return result
-        return _encode_call(
-            "encode_string",
-            str(self.number),
-            f"this->{self.field_name}",
-            force=self.force,
+        if self.force:
+            return f"ProtoEncode::encode_string(pos, {self.number}, this->{self.field_name}, true);"
+        return (
+            f"ProtoEncode::encode_string(pos, {self.number}, this->{self.field_name});"
         )
 
     @property
@@ -1464,13 +1421,9 @@ class FixedArrayBytesType(TypeInfo):
             f"this->{self.field_name}", f"this->{self.field_name}_len", max_len=max_len
         ):
             return result
-        return _encode_call(
-            "encode_bytes",
-            str(self.number),
-            f"this->{self.field_name}",
-            f"this->{self.field_name}_len",
-            force=self.force,
-        )
+        if self.force:
+            return f"ProtoEncode::encode_bytes(pos, {self.number}, this->{self.field_name}, this->{self.field_name}_len, true);"
+        return f"ProtoEncode::encode_bytes(pos, {self.number}, this->{self.field_name}, this->{self.field_name}_len);"
 
     def dump(self, name: str) -> str:
         return f"out.append(format_hex_pretty({name}, {name}_len));"
@@ -1567,9 +1520,9 @@ class EnumType(VarintTypeMixin, TypeInfo):
     @property
     def encode_content(self) -> str:
         value_expr = f"static_cast<uint32_t>(this->{self.field_name})"
-        return _encode_call(
-            self.encode_func, str(self.number), value_expr, force=self.force
-        )
+        if self.force:
+            return f"ProtoEncode::{self.encode_func}(pos, {self.number}, {value_expr}, true);"
+        return f"ProtoEncode::{self.encode_func}(pos, {self.number}, {value_expr});"
 
     def dump(self, name: str) -> str:
         return f"out.append_p(proto_enum_to_string<{self.cpp_type}>({name}));"
@@ -1748,9 +1701,9 @@ def _generate_inline_encode_block(
 
     lines = []
     lines.append(f"auto &sub_msg = {element};")
-    lines.append(_encode_call("write_raw_byte", str(tag)))
+    lines.append(f"ProtoEncode::write_raw_byte(pos, {tag});")
     lines.append("uint8_t *len_pos = pos;")
-    lines.append(_encode_call("reserve_byte"))
+    lines.append("ProtoEncode::reserve_byte(pos);")
 
     # Generate inline field encoding for each sub-message field
     for field in sub_desc.field:
@@ -1822,22 +1775,17 @@ class FixedArrayRepeatedType(TypeInfo):
     def _encode_element(self, element: str) -> str:
         """Helper to generate encode statement for a single element."""
         if isinstance(self._ti, EnumType):
-            return _encode_call(
-                self._ti.encode_func,
-                str(self.number),
-                f"static_cast<uint32_t>({element})",
-                force=True,
-            )
+            return f"ProtoEncode::{self._ti.encode_func}(pos, {self.number}, static_cast<uint32_t>({element}), true);"
         # Repeated message elements use encode_sub_message (force=true is default)
         if isinstance(self._ti, MessageType):
             if _is_inline_encode(self._ti.cpp_type):
                 return _generate_inline_encode_block(
                     self.number, self._ti.cpp_type, element
                 )
-            return _encode_call(
-                "encode_sub_message", "buffer", str(self.number), element
-            )
-        return _encode_call(self._ti.encode_func, str(self.number), element, force=True)
+            return f"ProtoEncode::encode_sub_message(pos, buffer, {self.number}, {element});"
+        return (
+            f"ProtoEncode::{self._ti.encode_func}(pos, {self.number}, {element}, true);"
+        )
 
     @property
     def cpp_type(self) -> str:
@@ -2189,18 +2137,13 @@ class RepeatedTypeInfo(TypeInfo):
     def _encode_element_call(self, element: str) -> str:
         """Helper to generate encode call for a single element."""
         if isinstance(self._ti, EnumType):
-            return _encode_call(
-                self._ti.encode_func,
-                str(self.number),
-                f"static_cast<uint32_t>({element})",
-                force=True,
-            )
+            return f"ProtoEncode::{self._ti.encode_func}(pos, {self.number}, static_cast<uint32_t>({element}), true);"
         # Repeated message elements use encode_sub_message (force=true is default)
         if isinstance(self._ti, MessageType):
-            return _encode_call(
-                "encode_sub_message", "buffer", str(self.number), element
-            )
-        return _encode_call(self._ti.encode_func, str(self.number), element, force=True)
+            return f"ProtoEncode::encode_sub_message(pos, buffer, {self.number}, {element});"
+        return (
+            f"ProtoEncode::{self._ti.encode_func}(pos, {self.number}, {element}, true);"
+        )
 
     @property
     def encode_content(self) -> str:
@@ -2209,7 +2152,7 @@ class RepeatedTypeInfo(TypeInfo):
             # Special handling for const char* elements (when container_no_template contains "const char")
             if "const char" in self._container_no_template:
                 o = f"for (const char *it : *this->{self.field_name}) {{\n"
-                o += f"  {_encode_call(self._ti.encode_func, str(self.number), 'it', 'strlen(it)', force=True)}\n"
+                o += f"  ProtoEncode::{self._ti.encode_func}(pos, {self.number}, it, strlen(it), true);\n"
             else:
                 o = f"for (const auto &it : *this->{self.field_name}) {{\n"
                 o += f"  {self._encode_element_call('it')}\n"

tests/components/api/test_proto_mac_varint.cpp

@@ -59,7 +59,7 @@ static void verify_mac(uint64_t mac, size_t expected_bytes) {
 #ifdef ESPHOME_DEBUG_API
   uint8_t *proto_debug_end_ = api_buf.data() + api_buf.size();
 #endif
-  pos = ProtoEncode::encode_varint_raw_48bit(pos PROTO_ENCODE_DEBUG_ARG, mac);
+  ProtoEncode::encode_varint_raw_48bit(pos PROTO_ENCODE_DEBUG_ARG, mac);
   size_t new_len = pos - api_buf.data();
 
   EXPECT_EQ(new_len, expected_bytes) << "mac=0x" << std::hex << mac << std::dec;

tests/components/core/test_blocking_scope.cpp

@@ -0,0 +1,103 @@
+#include <gtest/gtest.h>
+
+#include "esphome/core/application.h"
+#include "esphome/core/hal.h"
+
+namespace esphome {
+
+// The scope must push the pass start forward by the time it covers and by
+// nothing else, so the blocking guard sees only the work outside it
+TEST(UnavoidableBlockingScope, ExcludesItsDurationFromThePass) {
+  const uint32_t pass_start = millis();
+  LoopBlockingGuard guard(nullptr, nullptr, pass_start);
+  ASSERT_EQ(App.get_loop_component_start_time(), pass_start);
+
+  const uint32_t before = millis();
+  {
+    UnavoidableBlockingScope scope;
+    delay(30);
+  }
+  const uint32_t excused = millis() - before;
+
+  const uint32_t moved = App.get_loop_component_start_time() - pass_start;
+  EXPECT_GE(moved, 30u);
+  EXPECT_LE(moved, excused);
+}
+
+TEST(UnavoidableBlockingScope, ZeroLengthScopeLeavesTheStartAlone) {
+  const uint32_t pass_start = millis();
+  LoopBlockingGuard guard(nullptr, nullptr, pass_start);
+  const uint32_t before = millis();
+  { UnavoidableBlockingScope scope; }
+  EXPECT_LE(App.get_loop_component_start_time() - pass_start, millis() - before);
+}
+
+// Nested scopes leave out the outer span exactly once, and never move the
+// start past now
+TEST(UnavoidableBlockingScope, NestedScopesExcuseTheOuterSpanOnce) {
+  const uint32_t pass_start = millis();
+  LoopBlockingGuard guard(nullptr, nullptr, pass_start);
+  const uint32_t before = millis();
+  {
+    UnavoidableBlockingScope outer;
+    {
+      UnavoidableBlockingScope inner;
+      delay(30);
+    }
+    delay(5);
+  }
+  const uint32_t excused = millis() - before;
+  const uint32_t moved = App.get_loop_component_start_time() - pass_start;
+  EXPECT_GE(moved, 35u);
+  EXPECT_LE(moved, excused);
+  EXPECT_GE(static_cast<int32_t>(millis() - App.get_loop_component_start_time()), 0);
+}
+
+namespace {
+// Static: the guard publishes the component to App and nothing clears it.
+// One instance per test, since a ratcheted threshold is permanent
+class DummyComponent : public Component {};
+DummyComponent &blocking_test_component(size_t index) {
+  static DummyComponent components[2];
+  return components[index];
+}
+}  // namespace
+
+// The excused stretch must neither warn nor ratchet the component's threshold
+TEST(UnavoidableBlockingScope, ExcusedStretchDoesNotRatchetTheThreshold) {
+  DummyComponent &component = blocking_test_component(0);
+  uint32_t threshold_before = 0;
+  component.should_warn_of_blocking(0, threshold_before);
+  {
+    LoopBlockingGuard guard(&component, nullptr, millis());
+    {
+      UnavoidableBlockingScope scope;
+      delay(WARN_IF_BLOCKING_OVER_CS * 10U + 20);
+    }
+    guard.finish();
+  }
+  uint32_t threshold_after = 0;
+  component.should_warn_of_blocking(0, threshold_after);
+  EXPECT_EQ(threshold_after, threshold_before);
+}
+
+// Work outside the scope is still measured and still ratchets
+TEST(UnavoidableBlockingScope, WorkOutsideTheScopeStillRatchetsTheThreshold) {
+  DummyComponent &component = blocking_test_component(1);
+  uint32_t threshold_before = 0;
+  component.should_warn_of_blocking(0, threshold_before);
+  {
+    LoopBlockingGuard guard(&component, nullptr, millis());
+    {
+      UnavoidableBlockingScope scope;
+      delay(20);
+    }
+    delay(WARN_IF_BLOCKING_OVER_CS * 10U + 20);
+    guard.finish();
+  }
+  uint32_t threshold_after = 0;
+  component.should_warn_of_blocking(0, threshold_after);
+  EXPECT_GT(threshold_after, threshold_before);
+}
+
+}  // namespace esphome

tests/integration/fixtures/api_encode_boundaries.yaml

@@ -1,58 +0,0 @@
-esphome:
-  name: api-encode-boundaries-test
-  # Top-level area fills DeviceInfoResponse.suggested_area (field 16, a two-byte tag)
-  area:
-    id: kitchen_area
-    name: Kitchen
-  on_boot:
-    - sensor.template.publish:
-        id: zero_then_value
-        state: 0.0
-
-host:
-api:
-logger:
-  level: DEBUG
-
-sensor:
-  - platform: template
-    name: "Zero Then Value"
-    id: zero_then_value
-    # Negative int32 takes the ten byte varint path
-    accuracy_decimals: -2
-    update_interval: never
-
-text_sensor:
-  - platform: template
-    name: "Long Text"
-    id: long_text
-    update_interval: never
-
-number:
-  - platform: template
-    name: "Negative Number"
-    optimistic: true
-    min_value: -1000
-    max_value: 1000
-    step: 0.5
-    initial_value: -123.5
-
-select:
-  - platform: template
-    name: "Long Option Select"
-    optimistic: true
-    options:
-      - short
-      - "option-with-a-name-long-enough-that-its-length-prefix-needs-two-varint-bytes-when-the-list-entities-response-is-encoded-xxxxxxxxxx"
-    initial_option: short
-
-button:
-  - platform: template
-    name: "Publish Values"
-    on_press:
-      - sensor.template.publish:
-          id: zero_then_value
-          state: 12.5
-      - text_sensor.template.publish:
-          id: long_text
-          state: !lambda return std::string(200, 'y');

tests/integration/state_utils.py

@@ -57,46 +57,6 @@ async def wait_for_state(
     return await asyncio.wait_for(future, timeout=timeout)
 
 
-class StateWaiter:
-    """Route one state subscription to any number of predicate waits."""
-
-    def __init__(self) -> None:
-        self._waiters: list[
-            tuple[Callable[[EntityState], bool], asyncio.Future[EntityState]]
-        ] = []
-
-    def on_state(self, state: EntityState) -> None:
-        for predicate, future in self._waiters:
-            if future.done():
-                continue
-            try:
-                matched = predicate(state)
-            except Exception as exc:  # noqa: BLE001  the wait re-raises it, the callback must not die
-                future.set_exception(exc)
-                continue
-            if matched:
-                future.set_result(state)
-
-    async def expect(
-        self,
-        predicate: Callable[[EntityState], bool],
-        timeout: float = 5.0,
-        label: str | None = None,
-    ) -> EntityState:
-        """Wait for the next state matching ``predicate``; states seen before this call do not count."""
-        entry = (predicate, asyncio.get_running_loop().create_future())
-        self._waiters.append(entry)
-        try:
-            async with asyncio.timeout(timeout):
-                return await entry[1]
-        except TimeoutError:
-            raise TimeoutError(
-                f"no state matched {label or predicate} within {timeout}s"
-            ) from None
-        finally:
-            self._waiters.remove(entry)
-
-
 def find_entity[T: EntityInfo](
     entities: list[EntityInfo],
     object_id_substring: str,

tests/integration/test_api_encode_boundaries.py

@@ -1,78 +0,0 @@
-"""Encode paths at their branch boundaries: zero skipped float, fixed32 state, negative int32,
-length prefixes of two varint bytes and two byte field tags."""
-
-from __future__ import annotations
-
-import asyncio
-
-from aioesphomeapi import (
-    NumberState,
-    SelectInfo,
-    SensorInfo,
-    SensorState,
-    TextSensorState,
-)
-import pytest
-
-from .state_utils import InitialStateHelper, StateWaiter, require_entity
-from .types import APIClientConnectedFactory, RunCompiledFunction
-
-LONG_OPTION = (
-    "option-with-a-name-long-enough-that-its-length-prefix-needs-two-varint-bytes-"
-    "when-the-list-entities-response-is-encoded-xxxxxxxxxx"
-)
-
-
-@pytest.mark.asyncio
-async def test_api_encode_boundaries(
-    yaml_config: str,
-    run_compiled: RunCompiledFunction,
-    api_client_connected: APIClientConnectedFactory,
-) -> None:
-    async with run_compiled(yaml_config), api_client_connected() as client:
-        device_info, (entities, _) = await asyncio.gather(
-            client.device_info(), client.list_entities_services()
-        )
-        assert device_info.suggested_area == "Kitchen"
-
-        sensor = require_entity(entities, "zero_then_value", SensorInfo)
-        assert sensor.accuracy_decimals == -2
-        select = require_entity(entities, "long_option_select", SelectInfo)
-        assert len(LONG_OPTION) >= 128
-        assert select.options == ["short", LONG_OPTION]
-        text = require_entity(entities, "long_text")
-        number = require_entity(entities, "negative_number")
-        button = require_entity(entities, "publish_values")
-
-        initial = InitialStateHelper(entities)
-        waiter = StateWaiter()
-        client.subscribe_states(initial.on_state_wrapper(waiter.on_state))
-        await initial.wait_for_initial_states()
-
-        # A float of exactly zero is skipped on the wire and must still read as 0.0, not missing
-        first = initial.initial_states[sensor.key]
-        assert isinstance(first, SensorState)
-        assert first.state == 0.0 and not first.missing_state
-        first_number = initial.initial_states[number.key]
-        assert isinstance(first_number, NumberState)
-        assert first_number.state == -123.5
-
-        client.button_command(button.key)
-        await asyncio.gather(
-            waiter.expect(
-                lambda s: (
-                    isinstance(s, SensorState)
-                    and s.key == sensor.key
-                    and s.state == 12.5
-                ),
-                label="sensor 12.5",
-            ),
-            waiter.expect(
-                lambda s: (
-                    isinstance(s, TextSensorState)
-                    and s.key == text.key
-                    and s.state == "y" * 200
-                ),
-                label="text 200 x y",
-            ),
-        )

tests/unit_tests/components/api/test_api_proto.py

@@ -380,13 +380,3 @@ def test_api_version_minor_is_at_least_15() -> None:
         "clients to see api_version >= 1.15 in HelloResponse before they will "
         "ever request it."
     )
-
-
-def test_generated_encode_calls_keep_the_cursor() -> None:
-    """No generated ProtoEncode call may drop the returned cursor."""
-    dropped = [
-        line
-        for line in CPP_TEXT.splitlines()
-        if "ProtoEncode::" in line and "pos = ProtoEncode::" not in line
-    ]
-    assert not dropped, dropped[:5]

tests/unit_tests/components/api/test_api_protobuf_generator.py

@@ -15,11 +15,9 @@ import pytest
 
 sys.path.insert(0, str(Path(__file__).parents[4] / "script" / "api_protobuf"))
 
-import aioesphomeapi.api_options_pb2 as pb  # noqa: E402
 from api_protobuf import (  # noqa: E402
     MAX_MESSAGE_ID,
     _make_ifdef_line,
-    create_field_type_info,
     get_varint64_ifdef,
     validate_message_id,
 )
@@ -45,14 +43,7 @@ UINT64 = descriptor_pb2.FieldDescriptorProto.TYPE_UINT64
 INT64 = descriptor_pb2.FieldDescriptorProto.TYPE_INT64
 SINT64 = descriptor_pb2.FieldDescriptorProto.TYPE_SINT64
 UINT32 = descriptor_pb2.FieldDescriptorProto.TYPE_UINT32
-INT32 = descriptor_pb2.FieldDescriptorProto.TYPE_INT32
-SINT32 = descriptor_pb2.FieldDescriptorProto.TYPE_SINT32
 FIXED64 = descriptor_pb2.FieldDescriptorProto.TYPE_FIXED64
-FIXED32 = descriptor_pb2.FieldDescriptorProto.TYPE_FIXED32
-FLOAT = descriptor_pb2.FieldDescriptorProto.TYPE_FLOAT
-BOOL = descriptor_pb2.FieldDescriptorProto.TYPE_BOOL
-STRING = descriptor_pb2.FieldDescriptorProto.TYPE_STRING
-BYTES = descriptor_pb2.FieldDescriptorProto.TYPE_BYTES
 
 
 def test_no_varint64_fields() -> None:
@@ -116,69 +107,3 @@ def test_message_id_at_maximum_is_accepted() -> None:
 def test_message_id_above_maximum_is_rejected() -> None:
     with pytest.raises(ValueError, match="exceeds the plaintext"):
         validate_message_id(MAX_MESSAGE_ID + 1, "TooBigMessage")
-
-
-def _field(
-    field_type: int, number: int = 1, *, force: bool = False, repeated: bool = False
-) -> descriptor_pb2.FieldDescriptorProto:
-    field = descriptor_pb2.FieldDescriptorProto(
-        name="value", number=number, type=field_type
-    )
-    if repeated:
-        field.label = descriptor_pb2.FieldDescriptorProto.LABEL_REPEATED
-    if force:
-        field.options.Extensions[pb.force] = True
-    return field
-
-
-def _encode_field(
-    field_type: int, number: int = 1, force: bool = False, repeated: bool = False
-) -> str:
-    """Return the encode statement the generator emits for one encode-only field."""
-    field = _field(field_type, number, force=force, repeated=repeated)
-    return create_field_type_info(
-        field, needs_decode=False, needs_encode=True
-    ).encode_content
-
-
-SCALAR_TYPES = [
-    BOOL,
-    UINT32,
-    INT32,
-    UINT64,
-    INT64,
-    SINT32,
-    FLOAT,
-    FIXED32,
-    STRING,
-    BYTES,
-]
-
-
-@pytest.mark.parametrize("field_type", SCALAR_TYPES)
-def test_forced_fields_use_the_force_overload_or_raw_writes(field_type: int) -> None:
-    content = _encode_field(field_type, force=True)
-    assert (
-        "_force(" in content
-        or "write_raw_byte(" in content
-        or "write_tag_and_fixed32(" in content
-    ), content
-
-
-@pytest.mark.parametrize("field_type", [FLOAT, FIXED32])
-def test_single_byte_tag_fixed32_shares_the_outlined_writer(field_type: int) -> None:
-    unconditional = _encode_field(field_type, force=True)
-    assert unconditional.count("write_tag_and_fixed32(pos, 13,") == 1, unconditional
-    guarded = _encode_field(field_type, force=False)
-    assert guarded.startswith("if ("), guarded
-    assert "[[likely]]" in guarded
-    assert "write_tag_and_fixed32(pos, 13," in guarded
-
-
-@pytest.mark.parametrize("field_type", [FLOAT, FIXED32])
-def test_multi_byte_tag_fixed32_falls_back_to_the_generic_helper(
-    field_type: int,
-) -> None:
-    content = _encode_field(field_type, number=16)
-    assert "write_tag_and_fixed32" not in content, content
-    assert content.startswith("pos = ProtoEncode::encode_"), content