diff --git a/.clang-tidy.hash b/.clang-tidy.hash
index cd61d9ec48..60c3776aa8 100644
--- a/.clang-tidy.hash
+++ b/.clang-tidy.hash
@@ -1 +1 @@
-d48687d988ae2a94a9973226df773478a7db1d52133545f07aa05e34fc678dcf
+10c432ae818f9ed7fd4a0176a04467b1f2634363f5ec985045a6d72747f60b90
diff --git a/Doxyfile b/Doxyfile
index 599109b43f..fb85e8028c 100644
--- a/Doxyfile
+++ b/Doxyfile
@@ -48,7 +48,7 @@ PROJECT_NAME           = ESPHome
 # could be handy for archiving the generated documentation or if some version
 # control system is used.
 
-PROJECT_NUMBER         = 2026.4.0b2
+PROJECT_NUMBER         = 2026.4.0b3
 
 # Using the PROJECT_BRIEF tag one can provide an optional one line description
 # for a project that appears at the top of each page and should give viewer a
diff --git a/esphome/__main__.py b/esphome/__main__.py
index 25b404ae45..7879cdad0c 100644
--- a/esphome/__main__.py
+++ b/esphome/__main__.py
@@ -750,8 +750,15 @@ def upload_using_esptool(
             platformio_api.FlashImage(
                 path=idedata.firmware_bin_path, offset=firmware_offset
             ),
-            *idedata.extra_flash_images,
         ]
+        for image in idedata.extra_flash_images:
+            if not image.path.is_file():
+                _LOGGER.warning(
+                    "Skipping missing flash image declared by platform: %s",
+                    image.path,
+                )
+                continue
+            flash_images.append(image)
 
     mcu = "esp8266"
     if CORE.is_esp32:
diff --git a/esphome/components/adc/sensor.py b/esphome/components/adc/sensor.py
index bab2762f00..09e09f0dc1 100644
--- a/esphome/components/adc/sensor.py
+++ b/esphome/components/adc/sensor.py
@@ -2,7 +2,11 @@ import logging
 
 import esphome.codegen as cg
 from esphome.components import sensor, voltage_sampler
-from esphome.components.esp32 import get_esp32_variant, include_builtin_idf_component
+from esphome.components.esp32 import (
+    get_esp32_variant,
+    include_builtin_idf_component,
+    require_adc_oneshot_iram,
+)
 from esphome.components.nrf52.const import AIN_TO_GPIO, EXTRA_ADC
 from esphome.components.zephyr import (
     zephyr_add_overlay,
@@ -24,6 +28,7 @@ from esphome.const import (
     PlatformFramework,
 )
 from esphome.core import CORE
+from esphome.types import ConfigType
 
 from . import (
     ATTENUATION_MODES,
@@ -65,6 +70,13 @@ def validate_config(config):
     return config
 
 
+def _require_adc_iram(config: ConfigType) -> ConfigType:
+    """Register ADC oneshot IRAM requirement during config validation."""
+    if CORE.is_esp32:
+        require_adc_oneshot_iram()
+    return config
+
+
 ADCSensor = adc_ns.class_(
     "ADCSensor", sensor.Sensor, cg.PollingComponent, voltage_sampler.VoltageSampler
 )
@@ -95,6 +107,7 @@ CONFIG_SCHEMA = cv.All(
     )
     .extend(cv.polling_component_schema("60s")),
     validate_config,
+    _require_adc_iram,
 )
 
 CONF_ADC_CHANNEL_ID = "adc_channel_id"
diff --git a/esphome/components/api/api.proto b/esphome/components/api/api.proto
index e4d0c2d16d..f906cfb8d7 100644
--- a/esphome/components/api/api.proto
+++ b/esphome/components/api/api.proto
@@ -671,6 +671,7 @@ message SensorStateResponse {
   option (source) = SOURCE_SERVER;
   option (ifdef) = "USE_SENSOR";
   option (no_delay) = true;
+  option (speed_optimized) = true;
 
   fixed32 key = 1 [(force) = true];
   float state = 2;
@@ -777,9 +778,10 @@ message SubscribeLogsResponse {
   option (source) = SOURCE_SERVER;
   option (log) = false;
   option (no_delay) = false;
+  option (speed_optimized) = true;
 
-  LogLevel level = 1;
-  bytes message = 3;
+  LogLevel level = 1 [(force) = true];
+  bytes message = 3 [(force) = true];
 }
 
 // ==================== NOISE ENCRYPTION ====================
@@ -1638,6 +1640,7 @@ message BluetoothLERawAdvertisementsResponse {
   option (source) = SOURCE_SERVER;
   option (ifdef) = "USE_BLUETOOTH_PROXY";
   option (no_delay) = true;
+  option (speed_optimized) = true;
 
   repeated BluetoothLERawAdvertisement advertisements = 1 [(fixed_array_with_length_define) = "BLUETOOTH_PROXY_ADVERTISEMENT_BATCH_SIZE"];
 }
diff --git a/esphome/components/api/api_options.proto b/esphome/components/api/api_options.proto
index dacc290e31..d5d0b37e8d 100644
--- a/esphome/components/api/api_options.proto
+++ b/esphome/components/api/api_options.proto
@@ -23,6 +23,7 @@ extend google.protobuf.MessageOptions {
     optional bool no_delay = 1040 [default=false];
     optional string base_class = 1041;
     optional bool inline_encode = 1042 [default=false];
+    optional bool speed_optimized = 1043 [default=false];
 }
 
 extend google.protobuf.FieldOptions {
diff --git a/esphome/components/api/api_pb2.cpp b/esphome/components/api/api_pb2.cpp
index c2d513f0d3..f304c85282 100644
--- a/esphome/components/api/api_pb2.cpp
+++ b/esphome/components/api/api_pb2.cpp
@@ -745,7 +745,9 @@ uint32_t ListEntitiesSensorResponse::calculate_size() const {
 #endif
   return size;
 }
-uint8_t *SensorStateResponse::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {
+__attribute__((optimize("O2")))  // NOLINT(clang-diagnostic-unknown-attributes)
+uint8_t *
+SensorStateResponse::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {
   uint8_t *__restrict__ pos = buffer.get_pos();
   ProtoEncode::write_tag_and_fixed32(pos PROTO_ENCODE_DEBUG_ARG, 13, this->key);
   ProtoEncode::encode_float(pos PROTO_ENCODE_DEBUG_ARG, 2, this->state);
@@ -755,7 +757,9 @@ uint8_t *SensorStateResponse::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG
 #endif
   return pos;
 }
-uint32_t SensorStateResponse::calculate_size() const {
+__attribute__((optimize("O2")))  // NOLINT(clang-diagnostic-unknown-attributes)
+uint32_t
+SensorStateResponse::calculate_size() const {
   uint32_t size = 0;
   size += 5;
   size += ProtoSize::calc_float(1, this->state);
@@ -912,16 +916,22 @@ bool SubscribeLogsRequest::decode_varint(uint32_t field_id, proto_varint_value_t
   }
   return true;
 }
-uint8_t *SubscribeLogsResponse::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {
+__attribute__((optimize("O2")))  // NOLINT(clang-diagnostic-unknown-attributes)
+uint8_t *
+SubscribeLogsResponse::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {
   uint8_t *__restrict__ pos = buffer.get_pos();
-  ProtoEncode::encode_uint32(pos PROTO_ENCODE_DEBUG_ARG, 1, static_cast<uint32_t>(this->level));
-  ProtoEncode::encode_bytes(pos PROTO_ENCODE_DEBUG_ARG, 3, this->message_ptr_, this->message_len_);
+  ProtoEncode::encode_uint32(pos PROTO_ENCODE_DEBUG_ARG, 1, static_cast<uint32_t>(this->level), true);
+  ProtoEncode::write_raw_byte(pos PROTO_ENCODE_DEBUG_ARG, 26);
+  ProtoEncode::encode_varint_raw(pos PROTO_ENCODE_DEBUG_ARG, this->message_len_);
+  ProtoEncode::encode_raw(pos PROTO_ENCODE_DEBUG_ARG, this->message_ptr_, this->message_len_);
   return pos;
 }
-uint32_t SubscribeLogsResponse::calculate_size() const {
+__attribute__((optimize("O2")))  // NOLINT(clang-diagnostic-unknown-attributes)
+uint32_t
+SubscribeLogsResponse::calculate_size() const {
   uint32_t size = 0;
-  size += this->level ? 2 : 0;
-  size += ProtoSize::calc_length(1, this->message_len_);
+  size += 2;
+  size += ProtoSize::calc_length_force(1, this->message_len_);
   return size;
 }
 #ifdef USE_API_NOISE
@@ -2328,7 +2338,9 @@ bool SubscribeBluetoothLEAdvertisementsRequest::decode_varint(uint32_t field_id,
   }
   return true;
 }
-uint8_t *BluetoothLERawAdvertisementsResponse::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {
+__attribute__((optimize("O2")))  // NOLINT(clang-diagnostic-unknown-attributes)
+uint8_t *
+BluetoothLERawAdvertisementsResponse::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {
   uint8_t *__restrict__ pos = buffer.get_pos();
   for (uint16_t i = 0; i < this->advertisements_len; i++) {
     auto &sub_msg = this->advertisements[i];
@@ -2350,7 +2362,9 @@ uint8_t *BluetoothLERawAdvertisementsResponse::encode(ProtoWriteBuffer &buffer P
   }
   return pos;
 }
-uint32_t BluetoothLERawAdvertisementsResponse::calculate_size() const {
+__attribute__((optimize("O2")))  // NOLINT(clang-diagnostic-unknown-attributes)
+uint32_t
+BluetoothLERawAdvertisementsResponse::calculate_size() const {
   uint32_t size = 0;
   for (uint16_t i = 0; i < this->advertisements_len; i++) {
     auto &sub_msg = this->advertisements[i];
diff --git a/esphome/components/atm90e32/atm90e32.h b/esphome/components/atm90e32/atm90e32.h
index c44a11e3ed..95154812cb 100644
--- a/esphome/components/atm90e32/atm90e32.h
+++ b/esphome/components/atm90e32/atm90e32.h
@@ -111,14 +111,14 @@ class ATM90E32Component : public PollingComponent,
 #endif
   float get_reference_voltage(uint8_t phase) {
 #ifdef USE_NUMBER
-    return (phase >= 0 && phase < 3 && ref_voltages_[phase]) ? ref_voltages_[phase]->state : 120.0;  // Default voltage
+    return (phase < 3 && ref_voltages_[phase]) ? ref_voltages_[phase]->state : 120.0;  // Default voltage
 #else
     return 120.0;  // Default voltage
 #endif
   }
   float get_reference_current(uint8_t phase) {
 #ifdef USE_NUMBER
-    return (phase >= 0 && phase < 3 && ref_currents_[phase]) ? ref_currents_[phase]->state : 5.0f;  // Default current
+    return (phase < 3 && ref_currents_[phase]) ? ref_currents_[phase]->state : 5.0f;  // Default current
 #else
     return 5.0f;   // Default current
 #endif
diff --git a/esphome/components/esp32/__init__.py b/esphome/components/esp32/__init__.py
index cd38c82dd8..7b3f9da3da 100644
--- a/esphome/components/esp32/__init__.py
+++ b/esphome/components/esp32/__init__.py
@@ -676,7 +676,7 @@ ARDUINO_FRAMEWORK_VERSION_LOOKUP = {
     "dev": cv.Version(3, 3, 8),
 }
 ARDUINO_PLATFORM_VERSION_LOOKUP = {
-    cv.Version(3, 3, 8): cv.Version(55, 3, 38),
+    cv.Version(3, 3, 8): cv.Version(55, 3, 38, "1"),
     cv.Version(3, 3, 7): cv.Version(55, 3, 37),
     cv.Version(3, 3, 6): cv.Version(55, 3, 36),
     cv.Version(3, 3, 5): cv.Version(55, 3, 35),
@@ -724,7 +724,7 @@ ESP_IDF_PLATFORM_VERSION_LOOKUP = {
     cv.Version(
         6, 0, 0
     ): "https://github.com/pioarduino/platform-espressif32.git#prep_IDF6",
-    cv.Version(5, 5, 4): cv.Version(55, 3, 38),
+    cv.Version(5, 5, 4): cv.Version(55, 3, 38, "1"),
     cv.Version(5, 5, 3, "1"): cv.Version(55, 3, 37),
     cv.Version(5, 5, 3): cv.Version(55, 3, 37),
     cv.Version(5, 5, 2): cv.Version(55, 3, 37),
@@ -744,8 +744,8 @@ ESP_IDF_PLATFORM_VERSION_LOOKUP = {
 # The platform-espressif32 version
 #  - https://github.com/pioarduino/platform-espressif32/releases
 PLATFORM_VERSION_LOOKUP = {
-    "recommended": cv.Version(55, 3, 38),
-    "latest": cv.Version(55, 3, 38),
+    "recommended": cv.Version(55, 3, 38, "1"),
+    "latest": cv.Version(55, 3, 38, "1"),
     "dev": "https://github.com/pioarduino/platform-espressif32.git#develop",
 }
 
@@ -1058,6 +1058,7 @@ CONF_DISABLE_MBEDTLS_PEER_CERT = "disable_mbedtls_peer_cert"
 CONF_DISABLE_MBEDTLS_PKCS7 = "disable_mbedtls_pkcs7"
 CONF_DISABLE_REGI2C_IN_IRAM = "disable_regi2c_in_iram"
 CONF_DISABLE_FATFS = "disable_fatfs"
+CONF_ADC_ONESHOT_IN_IRAM = "adc_oneshot_in_iram"
 
 # VFS requirement tracking
 # Components that need VFS features can call require_vfs_*() functions
@@ -1071,6 +1072,7 @@ KEY_MBEDTLS_PEER_CERT_REQUIRED = "mbedtls_peer_cert_required"
 KEY_MBEDTLS_PKCS7_REQUIRED = "mbedtls_pkcs7_required"
 KEY_FATFS_REQUIRED = "fatfs_required"
 KEY_MBEDTLS_SHA512_REQUIRED = "mbedtls_sha512_required"
+KEY_ADC_ONESHOT_IRAM_REQUIRED = "adc_oneshot_iram_required"
 
 
 def require_vfs_select() -> None:
@@ -1168,6 +1170,17 @@ def require_fatfs() -> None:
     CORE.data[KEY_ESP32][KEY_FATFS_REQUIRED] = True
 
 
+def require_adc_oneshot_iram() -> None:
+    """Mark that ADC oneshot IRAM safety is required by a component.
+
+    Call this from components that use the ADC oneshot driver. When flash cache is
+    disabled (e.g., during NVS writes by WiFi, BLE, Zigbee, or power management),
+    the ADC oneshot read function must be in IRAM to avoid crashes.
+    This sets CONFIG_ADC_ONESHOT_CTRL_FUNC_IN_IRAM.
+    """
+    CORE.data[KEY_ESP32][KEY_ADC_ONESHOT_IRAM_REQUIRED] = True
+
+
 def _parse_idf_component(value: str) -> ConfigType:
     """Parse IDF component shorthand syntax like 'owner/component^version'"""
     # Match operator followed by version-like string (digit or *)
@@ -1268,6 +1281,7 @@ FRAMEWORK_SCHEMA = cv.Schema(
                 cv.Optional(CONF_DISABLE_MBEDTLS_PEER_CERT, default=True): cv.boolean,
                 cv.Optional(CONF_DISABLE_MBEDTLS_PKCS7, default=True): cv.boolean,
                 cv.Optional(CONF_DISABLE_REGI2C_IN_IRAM, default=True): cv.boolean,
+                cv.Optional(CONF_ADC_ONESHOT_IN_IRAM, default=False): cv.boolean,
                 cv.Optional(CONF_DISABLE_FATFS, default=True): cv.boolean,
             }
         ),
@@ -2068,6 +2082,16 @@ async def to_code(config):
     if advanced[CONF_DISABLE_REGI2C_IN_IRAM]:
         add_idf_sdkconfig_option("CONFIG_ESP_REGI2C_CTRL_FUNC_IN_IRAM", False)
 
+    # Place ADC oneshot control functions in IRAM for cache safety
+    # When flash cache is disabled (during NVS writes by WiFi, BLE, Zigbee, Thread,
+    # power management, etc.), ADC reads will crash if these functions are in flash.
+    # Components using ADC call require_adc_oneshot_iram() to force this.
+    if (
+        CORE.data[KEY_ESP32].get(KEY_ADC_ONESHOT_IRAM_REQUIRED, False)
+        or advanced[CONF_ADC_ONESHOT_IN_IRAM]
+    ):
+        add_idf_sdkconfig_option("CONFIG_ADC_ONESHOT_CTRL_FUNC_IN_IRAM", True)
+
     # Disable FATFS support
     # Components that need FATFS (SD card, etc.) can call require_fatfs()
     if CORE.data[KEY_ESP32].get(KEY_FATFS_REQUIRED, False):
diff --git a/esphome/components/globals/__init__.py b/esphome/components/globals/__init__.py
index ec6730a41c..46725fe6dd 100644
--- a/esphome/components/globals/__init__.py
+++ b/esphome/components/globals/__init__.py
@@ -108,8 +108,13 @@ async def globals_set_to_code(config, action_id, template_arg, args):
     full_id, paren = await cg.get_variable_with_full_id(config[CONF_ID])
     template_arg = cg.TemplateArguments(full_id.type, *template_arg)
     var = cg.new_Pvariable(action_id, template_arg, paren)
+    # Use the global's value_type alias as the lambda return type so
+    # TemplatableFn stores a direct function pointer instead of going through
+    # the deprecated converting trampoline when the value expression deduces
+    # to a different type (e.g. int literal assigned to a float global).
+    value_type = cg.RawExpression(f"{full_id.type}::value_type")
     templ = await cg.templatable(
-        config[CONF_VALUE], args, None, to_exp=cg.RawExpression, wrap_constant=True
+        config[CONF_VALUE], args, value_type, to_exp=cg.RawExpression
     )
     cg.add(var.set_value(templ))
     return var
diff --git a/esphome/components/i2s_audio/microphone/__init__.py b/esphome/components/i2s_audio/microphone/__init__.py
index 761cbb7f48..1392d1d4ec 100644
--- a/esphome/components/i2s_audio/microphone/__init__.py
+++ b/esphome/components/i2s_audio/microphone/__init__.py
@@ -36,7 +36,7 @@ I2SAudioMicrophone = i2s_audio_ns.class_(
 )
 
 INTERNAL_ADC_VARIANTS = [esp32.VARIANT_ESP32]
-PDM_VARIANTS = [esp32.VARIANT_ESP32, esp32.VARIANT_ESP32S3]
+PDM_VARIANTS = [esp32.VARIANT_ESP32, esp32.VARIANT_ESP32S3, esp32.VARIANT_ESP32P4]
 
 
 def _validate_esp32_variant(config):
diff --git a/esphome/components/light/addressable_light.cpp b/esphome/components/light/addressable_light.cpp
index 2f6ffc9a38..d2f5913f4b 100644
--- a/esphome/components/light/addressable_light.cpp
+++ b/esphome/components/light/addressable_light.cpp
@@ -58,6 +58,12 @@ void AddressableLightTransformer::start() {
   // our transition will handle brightness, disable brightness in correction.
   this->light_.correction_.set_local_brightness(255);
   this->target_color_ *= to_uint8_scale(end_values.get_brightness() * end_values.get_state());
+
+  // Uniformity scan is deferred to the first apply() call. start() can run before the underlying
+  // LED output's setup() has allocated its frame buffer (e.g. on_boot at priority > HARDWARE
+  // triggering a transition), and reading through ESPColorView would deref a null buffer.
+  this->uniform_start_scanned_ = false;
+  this->uniform_start_is_uniform_ = false;
 }
 
 inline constexpr uint8_t subtract_scaled_difference(uint8_t a, uint8_t b, int32_t scale) {
@@ -97,12 +103,57 @@ optional<LightColorValues> AddressableLightTransformer::apply() {
   // non-linear when applying small deltas.
 
   if (smoothed_progress > this->last_transition_progress_ && this->last_transition_progress_ < 1.f) {
-    int32_t scale = int32_t(256.f * std::max((1.f - smoothed_progress) / (1.f - this->last_transition_progress_), 0.f));
-    for (auto led : this->light_) {
-      led.set_rgbw(subtract_scaled_difference(this->target_color_.red, led.get_red(), scale),
-                   subtract_scaled_difference(this->target_color_.green, led.get_green(), scale),
-                   subtract_scaled_difference(this->target_color_.blue, led.get_blue(), scale),
-                   subtract_scaled_difference(this->target_color_.white, led.get_white(), scale));
+    // Lazy uniformity scan: deferred from start() so the LED output's setup() has run and the
+    // frame buffer is valid. When every LED already has the same color (the common case: plain
+    // turn_on/turn_off on a uniform strip), interpolate math-only against a single start color.
+    // Avoiding the per-step read-back through the 8-bit stored byte prevents gamma round-trip
+    // quantization from stalling the fade at low values (e.g. gamma 2.8 pre-gamma values <27
+    // round to stored 0, freezing progress).
+    if (!this->uniform_start_scanned_) {
+      this->uniform_start_scanned_ = true;
+      if (this->light_.size() > 0) {
+        Color first = this->light_[0].get();
+        bool uniform = true;
+        for (int32_t i = 1; i < this->light_.size(); i++) {
+          if (this->light_[i].get() != first) {
+            uniform = false;
+            break;
+          }
+        }
+        if (uniform) {
+          this->uniform_start_color_ = first;
+          this->uniform_start_is_uniform_ = true;
+        }
+      }
+    }
+    if (this->uniform_start_is_uniform_) {
+      // All LEDs started at the same color: compute the interpolated value once and write it to
+      // every LED. No read-back, so each LED's stored byte advances through every gamma threshold
+      // as smoothed_progress crosses it, instead of stalling at 0 for low pre-gamma values.
+      //
+      // Trade-off: any mid-transition writes to individual LEDs (e.g. from a user lambda) will be
+      // overwritten on the next apply() here. The fallback path below would have respected them
+      // via its read-back. Concurrent per-LED mutation during a transition isn't a pattern we
+      // support, so this is acceptable.
+      // lerp(start, target, progress) via existing helper: target - (target-start)*(1-progress).
+      const Color &start = this->uniform_start_color_;
+      int32_t remaining = int32_t(256.f * (1.f - smoothed_progress));
+      uint8_t r = subtract_scaled_difference(this->target_color_.red, start.red, remaining);
+      uint8_t g = subtract_scaled_difference(this->target_color_.green, start.green, remaining);
+      uint8_t b = subtract_scaled_difference(this->target_color_.blue, start.blue, remaining);
+      uint8_t w = subtract_scaled_difference(this->target_color_.white, start.white, remaining);
+      for (auto led : this->light_) {
+        led.set_rgbw(r, g, b, w);
+      }
+    } else {
+      int32_t scale =
+          int32_t(256.f * std::max((1.f - smoothed_progress) / (1.f - this->last_transition_progress_), 0.f));
+      for (auto led : this->light_) {
+        led.set_rgbw(subtract_scaled_difference(this->target_color_.red, led.get_red(), scale),
+                     subtract_scaled_difference(this->target_color_.green, led.get_green(), scale),
+                     subtract_scaled_difference(this->target_color_.blue, led.get_blue(), scale),
+                     subtract_scaled_difference(this->target_color_.white, led.get_white(), scale));
+      }
     }
     this->last_transition_progress_ = smoothed_progress;
     this->light_.schedule_show();
diff --git a/esphome/components/light/addressable_light.h b/esphome/components/light/addressable_light.h
index 17cdb7d6f6..0202ad380a 100644
--- a/esphome/components/light/addressable_light.h
+++ b/esphome/components/light/addressable_light.h
@@ -115,6 +115,9 @@ class AddressableLightTransformer : public LightTransformer {
   AddressableLight &light_;
   float last_transition_progress_{0.0f};
   Color target_color_{};
+  Color uniform_start_color_{};
+  bool uniform_start_scanned_{false};
+  bool uniform_start_is_uniform_{false};
 };
 
 }  // namespace esphome::light
diff --git a/esphome/components/micro_wake_word/__init__.py b/esphome/components/micro_wake_word/__init__.py
index de95e4961b..5ab1e4bb80 100644
--- a/esphome/components/micro_wake_word/__init__.py
+++ b/esphome/components/micro_wake_word/__init__.py
@@ -452,7 +452,7 @@ async def to_code(config):
 
     esp32.add_idf_component(name="espressif/esp-tflite-micro", ref="1.3.3~1")
     # Pin esp-nn for stable future builds (esp-tflite-micro depends on esp-nn)
-    esp32.add_idf_component(name="espressif/esp-nn", ref="1.2.1")
+    esp32.add_idf_component(name="espressif/esp-nn", ref="1.1.2")
 
     cg.add_build_flag("-DTF_LITE_STATIC_MEMORY")
     cg.add_build_flag("-DTF_LITE_DISABLE_X86_NEON")
diff --git a/esphome/components/midea/air_conditioner.cpp b/esphome/components/midea/air_conditioner.cpp
index 50521cf238..69e0d46d2d 100644
--- a/esphome/components/midea/air_conditioner.cpp
+++ b/esphome/components/midea/air_conditioner.cpp
@@ -28,7 +28,8 @@ void AirConditioner::on_status_change() {
   if (this->base_.getAutoconfStatus() == dudanov::midea::AUTOCONF_OK &&
       this->base_.getCapabilities().supportFrostProtectionPreset() && !this->frost_protection_set_) {
     // Read existing presets (set by codegen), append frost protection, write back
-    const auto &existing = this->get_traits().get_supported_custom_presets();
+    auto traits = this->get_traits();
+    const auto &existing = traits.get_supported_custom_presets();
     bool found = false;
     for (const char *p : existing) {
       if (strcmp(p, Constants::FREEZE_PROTECTION) == 0) {
diff --git a/esphome/components/mipi_spi/mipi_spi.h b/esphome/components/mipi_spi/mipi_spi.h
index 423226b1d7..2242be6c17 100644
--- a/esphome/components/mipi_spi/mipi_spi.h
+++ b/esphome/components/mipi_spi/mipi_spi.h
@@ -234,9 +234,9 @@ class MipiSpi : public display::Display,
   }
 
   void dump_config() override {
-    internal_dump_config(this->model_, this->get_width(), this->get_height(), OFFSET_WIDTH, OFFSET_HEIGHT, MADCTL,
-                         this->invert_colors_, DISPLAYPIXEL * 8, IS_BIG_ENDIAN, this->brightness_, this->cs_,
-                         this->reset_pin_, this->dc_pin_, this->mode_, this->data_rate_, BUS_TYPE,
+    internal_dump_config(this->model_, this->get_width(), this->get_height(), OFFSET_WIDTH, OFFSET_HEIGHT,
+                         (uint8_t) MADCTL, this->invert_colors_, DISPLAYPIXEL * 8, IS_BIG_ENDIAN, this->brightness_,
+                         this->cs_, this->reset_pin_, this->dc_pin_, this->mode_, this->data_rate_, BUS_TYPE,
                          HAS_HARDWARE_ROTATION);
   }
 
@@ -305,7 +305,7 @@ class MipiSpi : public display::Display,
       this->write_command_(BRIGHTNESS, this->brightness_.value());
 
     // calculate new madctl value from base value adjusted for rotation
-    uint8_t madctl = MADCTL;  // lower 8 bits only
+    uint8_t madctl = (uint8_t) MADCTL;  // lower 8 bits only
     constexpr bool use_flips = (MADCTL & MADCTL_FLIP_FLAG) != 0;
     constexpr uint8_t x_mask = use_flips ? MADCTL_XFLIP : MADCTL_MX;
     constexpr uint8_t y_mask = use_flips ? MADCTL_YFLIP : MADCTL_MY;
diff --git a/esphome/components/nextion/nextion.cpp b/esphome/components/nextion/nextion.cpp
index b0e14b5ea3..e42f7ca216 100644
--- a/esphome/components/nextion/nextion.cpp
+++ b/esphome/components/nextion/nextion.cpp
@@ -36,8 +36,9 @@ bool Nextion::send_command_(const std::string &command) {
   }
 
 #ifdef USE_NEXTION_COMMAND_SPACING
-  if (!this->connection_state_.ignore_is_setup_ && !this->command_pacer_.can_send()) {
-    ESP_LOGN(TAG, "Command spacing: delaying command '%s'", command.c_str());
+  const uint32_t now = App.get_loop_component_start_time();
+  if (!this->connection_state_.ignore_is_setup_ && !this->command_pacer_.can_send(now)) {
+    ESP_LOGN(TAG, "Command spacing: delaying '%s'", command.c_str());
     return false;
   }
 #endif  // USE_NEXTION_COMMAND_SPACING
@@ -48,6 +49,16 @@ bool Nextion::send_command_(const std::string &command) {
   const uint8_t to_send[3] = {0xFF, 0xFF, 0xFF};
   this->write_array(to_send, sizeof(to_send));
 
+#ifdef USE_NEXTION_COMMAND_SPACING
+  // Mark sent immediately after writing to UART. The pacer enforces inter-command
+  // spacing from the transmit side. Marking on ACK (0x01) would leave last_command_time_
+  // at zero indefinitely, making can_send() always return true and spacing a no-op.
+  // ignore_is_setup_ commands (setup/init sequence) bypass spacing intentionally.
+  if (!this->connection_state_.ignore_is_setup_) {
+    this->command_pacer_.mark_sent(now);
+  }
+#endif  // USE_NEXTION_COMMAND_SPACING
+
   return true;
 }
 
@@ -253,11 +264,8 @@ bool Nextion::send_command(const char *command) {
   if ((!this->is_setup() && !this->connection_state_.ignore_is_setup_) || this->is_sleeping())
     return false;
 
-  if (this->send_command_(command)) {
-    this->add_no_result_to_queue_("command");
-    return true;
-  }
-  return false;
+  this->add_no_result_to_queue_with_command_("command", command);
+  return true;
 }
 
 bool Nextion::send_command_printf(const char *format, ...) {
@@ -274,11 +282,8 @@ bool Nextion::send_command_printf(const char *format, ...) {
     return false;
   }
 
-  if (this->send_command_(buffer)) {
-    this->add_no_result_to_queue_("command_printf");
-    return true;
-  }
-  return false;
+  this->add_no_result_to_queue_with_command_("command_printf", buffer);
+  return true;
 }
 
 #ifdef NEXTION_PROTOCOL_LOG
@@ -349,25 +354,43 @@ void Nextion::loop() {
   }
 
 #ifdef USE_NEXTION_COMMAND_SPACING
-  // Try to send any pending commands if spacing allows
   this->process_pending_in_queue_();
+#ifdef USE_NEXTION_WAVEFORM
+  if (!this->waveform_queue_.empty()) {
+    this->check_pending_waveform_();
+  }
+#endif  // USE_NEXTION_WAVEFORM
 #endif  // USE_NEXTION_COMMAND_SPACING
 }
 
 #ifdef USE_NEXTION_COMMAND_SPACING
 void Nextion::process_pending_in_queue_() {
-  if (this->nextion_queue_.empty() || !this->command_pacer_.can_send()) {
-    return;
-  }
+#ifdef USE_NEXTION_MAX_COMMANDS_PER_LOOP
+  size_t commands_sent = 0;
+#endif  // USE_NEXTION_MAX_COMMANDS_PER_LOOP
 
-  // Check if first item in queue has a pending command
-  auto *front_item = this->nextion_queue_.front();
-  if (front_item && !front_item->pending_command.empty()) {
-    if (this->send_command_(front_item->pending_command)) {
-      // Command sent successfully, clear the pending command
-      front_item->pending_command.clear();
-      ESP_LOGVV(TAG, "Pending command sent: %s", front_item->component->get_variable_name().c_str());
+  for (auto *item : this->nextion_queue_) {
+    if (item == nullptr || item->pending_command.empty()) {
+      continue;  // Already sent, waiting for ACK — skip, don't stop
     }
+
+#ifdef USE_NEXTION_MAX_COMMANDS_PER_LOOP
+    if (++commands_sent > this->max_commands_per_loop_) {
+      ESP_LOGV(TAG, "Pending cmds: loop limit reached, deferring");
+      break;
+    }
+#endif  // USE_NEXTION_MAX_COMMANDS_PER_LOOP
+
+    const uint32_t now = App.get_loop_component_start_time();
+    if (!this->command_pacer_.can_send(now)) {
+      break;  // Spacing not elapsed, stop for this loop iteration
+    }
+
+    if (!this->send_command_(item->pending_command)) {
+      break;  // Unexpected send failure, stop
+    }
+    item->pending_command.clear();
+    ESP_LOGVV(TAG, "Pending cmd sent: %s", item->component->get_variable_name().c_str());
   }
 }
 #endif  // USE_NEXTION_COMMAND_SPACING
@@ -470,10 +493,6 @@ void Nextion::process_nextion_commands_() {
             this->setup_callback_.call();
           }
         }
-#ifdef USE_NEXTION_COMMAND_SPACING
-        this->command_pacer_.mark_sent();  // Here is where we should mark the command as sent
-        ESP_LOGN(TAG, "Command spacing: marked command sent");
-#endif
         break;
       case 0x02:  // invalid Component ID or name was used
         ESP_LOGW(TAG, "Invalid component ID/name");
@@ -1079,10 +1098,18 @@ void Nextion::add_no_result_to_queue_(const std::string &variable_name) {
 }
 
 /**
- * @brief
+ * @brief Send a command and enqueue it for response tracking.
  *
- * @param variable_name Variable name for the queue
- * @param command
+ * Callers are responsible for checking is_sleeping() before calling this
+ * method. The sleep guard is deliberately absent here because some callers
+ * (e.g. add_no_result_to_queue_with_ignore_sleep_printf_()) are explicitly
+ * sleep-safe and must bypass it.
+ *
+ * If USE_NEXTION_COMMAND_SPACING is enabled and the pacer is not ready,
+ * the command is saved in the queue entry for retry rather than dropped.
+ *
+ * @param variable_name Name of the variable or component associated with the command.
+ * @param command       The raw command string to send.
  */
 void Nextion::add_no_result_to_queue_with_command_(const std::string &variable_name, const std::string &command) {
   if ((!this->is_setup() && !this->connection_state_.ignore_is_setup_) || command.empty())
@@ -1263,9 +1290,22 @@ void Nextion::add_to_get_queue(NextionComponentBase *component) {
 
   std::string command = "get " + component->get_variable_name_to_send();
 
+#ifdef USE_NEXTION_COMMAND_SPACING
+  // Always enqueue first so the response handler is present when the command
+  // is eventually sent. Store the command for retry if spacing blocked it;
+  // process_pending_in_queue_() will transmit it when the pacer allows.
+  nextion_queue->pending_command = command;
+  this->nextion_queue_.push_back(nextion_queue);
+  if (this->send_command_(command)) {
+    nextion_queue->pending_command.clear();
+  }
+#else   // USE_NEXTION_COMMAND_SPACING
   if (this->send_command_(command)) {
     this->nextion_queue_.push_back(nextion_queue);
+  } else {
+    delete nextion_queue;  // NOLINT(cppcoreguidelines-owning-memory)
   }
+#endif  // USE_NEXTION_COMMAND_SPACING
 }
 
 #ifdef USE_NEXTION_WAVEFORM
@@ -1309,10 +1349,10 @@ void Nextion::check_pending_waveform_() {
   char command[24];  // "addt " + uint8 + "," + uint8 + "," + uint8 + null = max 17 chars
   buf_append_printf(command, sizeof(command), 0, "addt %u,%u,%zu", component->get_component_id(),
                     component->get_wave_channel_id(), buffer_to_send);
-  if (!this->send_command_(command)) {
-    delete nb;  // NOLINT(cppcoreguidelines-owning-memory)
-    this->waveform_queue_.pop();
-  }
+  // If spacing or setup state blocks the send, leave the entry at the front
+  // of waveform_queue_ for retry on the next loop iteration via
+  // check_pending_waveform_(). Only pop on a successful send.
+  this->send_command_(command);
 }
 #endif  // USE_NEXTION_WAVEFORM
 
diff --git a/esphome/components/nextion/nextion.h b/esphome/components/nextion/nextion.h
index c84a5cd49c..c62772ac75 100644
--- a/esphome/components/nextion/nextion.h
+++ b/esphome/components/nextion/nextion.h
@@ -55,15 +55,20 @@ class NextionCommandPacer {
   uint8_t get_spacing() const { return spacing_ms_; }
 
   /**
-   * @brief Check if enough time has passed to send next command
-   * @return true if enough time has passed since last command
+   * @brief Check if enough time has passed to send the next command.
+   * @param now Current timestamp in milliseconds (use App.get_loop_component_start_time()
+   *            for consistency with the rest of the queue timing).
+   * @return true if the spacing interval has elapsed since the last command was sent.
    */
-  bool can_send() const { return (millis() - last_command_time_) >= spacing_ms_; }
+  bool can_send(uint32_t now) const { return (now - last_command_time_) >= spacing_ms_; }
 
   /**
-   * @brief Mark a command as sent, updating the timing
+   * @brief Record the transmit timestamp for the most recently sent command.
+   * @param now Current timestamp in milliseconds, as returned by
+   *            App.get_loop_component_start_time(). Must use the same clock
+   *            source as can_send() to avoid unsigned underflow.
    */
-  void mark_sent() { last_command_time_ = millis(); }
+  void mark_sent(uint32_t now) { last_command_time_ = now; }
 
  private:
   uint8_t spacing_ms_;
diff --git a/esphome/components/packages/__init__.py b/esphome/components/packages/__init__.py
index 3a15b5b95a..3f3df75351 100644
--- a/esphome/components/packages/__init__.py
+++ b/esphome/components/packages/__init__.py
@@ -321,12 +321,15 @@ def _walk_packages(
         return config
     packages = config[CONF_PACKAGES]
 
-    if not isinstance(packages, (dict, list)):
-        raise cv.Invalid(
-            f"Packages must be a key to value mapping or list, got {type(packages)} instead"
-        )
-
     with cv.prepend_path(CONF_PACKAGES):
+        if isinstance(packages, yaml_util.IncludeFile):
+            # If the packages key is an IncludeFile, resolve it first before processing.
+            packages, _ = resolve_include(packages, [], context, strict_undefined=False)
+        if not isinstance(packages, (dict, list)):
+            raise cv.Invalid(
+                f"Packages must be a key to value mapping or list, got {type(packages)} instead"
+            )
+
         if not isinstance(packages, dict):
             _walk_package_list(packages, callback, context)
         elif (result := _walk_package_dict(packages, callback, context)) is not None:
diff --git a/esphome/components/tcs34725/tcs34725.cpp b/esphome/components/tcs34725/tcs34725.cpp
index 4fe87de0ca..1098d8de5f 100644
--- a/esphome/components/tcs34725/tcs34725.cpp
+++ b/esphome/components/tcs34725/tcs34725.cpp
@@ -315,7 +315,7 @@ void TCS34725Component::set_integration_time(TCS34725IntegrationTime integration
     my_integration_time_regval = integration_time;
     this->integration_time_auto_ = false;
   }
-  this->integration_time_ = (256.f - my_integration_time_regval) * 2.4f;
+  this->integration_time_ = (256.f - (float) my_integration_time_regval) * 2.4f;
   ESP_LOGI(TAG, "TCS34725I Integration time set to: %.1fms", this->integration_time_);
 }
 void TCS34725Component::set_gain(TCS34725Gain gain) {
diff --git a/esphome/components/web_server/ota/ota_web_server.cpp b/esphome/components/web_server/ota/ota_web_server.cpp
index 95b166901a..9812714ec0 100644
--- a/esphome/components/web_server/ota/ota_web_server.cpp
+++ b/esphome/components/web_server/ota/ota_web_server.cpp
@@ -114,7 +114,25 @@ void OTARequestHandler::handleUpload(AsyncWebServerRequest *request, const Platf
                                      uint8_t *data, size_t len, bool final) {
   ota::OTAResponseTypes error_code = ota::OTA_RESPONSE_OK;
 
-  if (index == 0 && !this->ota_backend_) {
+  // First byte of a new upload: index==0 with actual data. (web_server_idf
+  // fires a separate start-marker call with data==nullptr/len==0 before the
+  // first real chunk; gate on len>0 so we only trigger once per upload.)
+  if (index == 0 && len > 0) {
+    // If a previous upload was interrupted (e.g. client closed the tab, TCP
+    // reset) the backend from that session may still be open. Tear it down
+    // so flash state doesn't get concatenated with the new image (which can
+    // produce a technically-valid-sized but corrupted firmware that bricks
+    // the device once it reboots).
+    if (this->ota_backend_) {
+      ESP_LOGW(TAG, "New OTA upload received while previous session was still open; aborting previous session");
+      this->ota_backend_->abort();
+#ifdef USE_OTA_STATE_LISTENER
+      // Notify listeners that the previous session was aborted before the new one starts.
+      this->parent_->notify_state_deferred_(ota::OTA_ABORT, 0.0f, 0);
+#endif
+      this->ota_backend_.reset();
+    }
+
     // Initialize OTA on first call
     this->ota_init_(filename.c_str());
 
diff --git a/esphome/const.py b/esphome/const.py
index b4cd758330..d5d726d0db 100644
--- a/esphome/const.py
+++ b/esphome/const.py
@@ -4,7 +4,7 @@ from enum import Enum
 
 from esphome.enum import StrEnum
 
-__version__ = "2026.4.0b2"
+__version__ = "2026.4.0b3"
 
 ALLOWED_NAME_CHARS = "abcdefghijklmnopqrstuvwxyz0123456789-_"
 VALID_SUBSTITUTIONS_CHARACTERS = (
diff --git a/esphome/cpp_helpers.py b/esphome/cpp_helpers.py
index 479090016f..f2bd3b92a3 100644
--- a/esphome/cpp_helpers.py
+++ b/esphome/cpp_helpers.py
@@ -113,7 +113,8 @@ def _generate_source_table_code(
     entries = ", ".join(var_names)
     lines.append(f"static const char *const {table_var}[] PROGMEM = {{{entries}}};")
     lines.append(f"const LogString *{lookup_fn}(uint8_t index) {{")
-    lines.append(f'  if (index == 0 || index > {count}) return LOG_STR("<unknown>");')
+    cond = "index == 0" if count >= 255 else f"index == 0 || index > {count}"
+    lines.append(f'  if ({cond}) return LOG_STR("<unknown>");')
     lines.append("  return reinterpret_cast<const LogString *>(")
     lines.append(f"    progmem_read_ptr(&{table_var}[index - 1]));")
     lines.append("}")
diff --git a/esphome/platformio_api.py b/esphome/platformio_api.py
index e9719f7dcd..fc21977fdd 100644
--- a/esphome/platformio_api.py
+++ b/esphome/platformio_api.py
@@ -65,9 +65,6 @@ def run_platformio_cli(*args, **kwargs) -> str | int:
     os.environ.setdefault("UV_HTTP_RETRIES", "10")
     cmd = [sys.executable, "-m", "esphome.platformio_runner"] + list(args)
 
-    if not CORE.verbose:
-        kwargs["filter_lines"] = FILTER_PLATFORMIO_LINES
-
     return run_external_process(*cmd, **kwargs)
 
 
diff --git a/esphome/platformio_runner.py b/esphome/platformio_runner.py
index 408d49d1a6..92700d5c42 100644
--- a/esphome/platformio_runner.py
+++ b/esphome/platformio_runner.py
@@ -105,6 +105,36 @@ def main() -> int:
     patch_structhash()
     patch_file_downloader()
 
+    # Wrap stdout/stderr with RedirectText before PlatformIO runs:
+    #
+    # 1. RedirectText.isatty() unconditionally returns True. Click, tqdm, and
+    #    PlatformIO's own progress-bar code check ``stream.isatty()`` to
+    #    decide whether to emit TTY-format output (``\r`` cursor moves, ANSI
+    #    colors, fancy progress bars). With the wrapper in place they always
+    #    emit TTY format, even when our real stdout is a pipe to the parent
+    #    process. Downstream consumers (local terminals and the Home
+    #    Assistant dashboard log viewer) render the TTY control sequences
+    #    correctly, so the user sees real progress bars.
+    #
+    # 2. FILTER_PLATFORMIO_LINES is applied inside RedirectText.write() in
+    #    this subprocess, so noisy PlatformIO output is dropped before it
+    #    ever leaves the runner. This replaces the parent-side filtering
+    #    that was lost when we switched from in-process to subprocess — the
+    #    parent's ``subprocess.run`` uses ``.fileno()`` on RedirectText and
+    #    bypasses its ``write()`` path entirely.
+    #
+    # Filtering is disabled when the user passed -v / --verbose to
+    # ``esphome compile``, preserving the previous in-process behavior where
+    # verbose mode let all PlatformIO output through unfiltered.
+    from esphome.platformio_api import FILTER_PLATFORMIO_LINES
+    from esphome.util import RedirectText
+
+    is_verbose = any(arg in ("-v", "--verbose") for arg in sys.argv[1:])
+    filter_lines = None if is_verbose else FILTER_PLATFORMIO_LINES
+
+    sys.stdout = RedirectText(sys.stdout, filter_lines=filter_lines)
+    sys.stderr = RedirectText(sys.stderr, filter_lines=filter_lines)
+
     import platformio.__main__
 
     return platformio.__main__.main() or 0
diff --git a/platformio.ini b/platformio.ini
index 7d17628a8f..708d62afdc 100644
--- a/platformio.ini
+++ b/platformio.ini
@@ -133,7 +133,7 @@ extra_scripts = post:esphome/components/esp8266/post_build.py.script
 ; This are common settings for the ESP32 (all variants) using Arduino.
 [common:esp32-arduino]
 extends = common:arduino
-platform = https://github.com/pioarduino/platform-espressif32/releases/download/55.03.38/platform-espressif32.zip
+platform = https://github.com/pioarduino/platform-espressif32/releases/download/55.03.38-1/platform-espressif32.zip
 platform_packages =
     pioarduino/framework-arduinoespressif32@https://github.com/espressif/arduino-esp32/releases/download/3.3.8/esp32-core-3.3.8.tar.xz
     pioarduino/framework-espidf@https://github.com/pioarduino/esp-idf/releases/download/v5.5.4/esp-idf-v5.5.4.tar.xz
@@ -169,7 +169,7 @@ extra_scripts = post:esphome/components/esp32/post_build.py.script
 ; This are common settings for the ESP32 (all variants) using IDF.
 [common:esp32-idf]
 extends = common:idf
-platform = https://github.com/pioarduino/platform-espressif32/releases/download/55.03.38/platform-espressif32.zip
+platform = https://github.com/pioarduino/platform-espressif32/releases/download/55.03.38-1/platform-espressif32.zip
 platform_packages =
     pioarduino/framework-espidf@https://github.com/pioarduino/esp-idf/releases/download/v5.5.4/esp-idf-v5.5.4.tar.xz
 
diff --git a/requirements.txt b/requirements.txt
index e4a019a490..726a0a221a 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -12,7 +12,7 @@ platformio==6.1.19
 esptool==5.2.0
 click==8.3.2
 esphome-dashboard==20260408.1
-aioesphomeapi==44.13.3
+aioesphomeapi==44.15.0
 zeroconf==0.148.0
 puremagic==1.30
 ruamel.yaml==0.19.1 # dashboard_import
diff --git a/script/api_protobuf/api_protobuf.py b/script/api_protobuf/api_protobuf.py
index 39bfc865d0..73e0859d5e 100755
--- a/script/api_protobuf/api_protobuf.py
+++ b/script/api_protobuf/api_protobuf.py
@@ -1028,7 +1028,8 @@ class BytesType(TypeInfo):
         )
 
     def get_size_calculation(self, name: str, force: bool = False) -> str:
-        return f"size += ProtoSize::calc_length({self.calculate_field_id_size()}, this->{self.field_name}_len_);"
+        calc_fn = "calc_length_force" if force else "calc_length"
+        return f"size += ProtoSize::{calc_fn}({self.calculate_field_id_size()}, this->{self.field_name}_len_);"
 
     def get_estimated_size(self) -> int:
         return self.calculate_field_id_size() + 8  # field ID + 8 bytes typical bytes
@@ -1109,7 +1110,8 @@ class PointerToBytesBufferType(PointerToBufferTypeBase):
         )
 
     def get_size_calculation(self, name: str, force: bool = False) -> str:
-        return f"size += ProtoSize::calc_length({self.calculate_field_id_size()}, this->{self.field_name}_len);"
+        calc_fn = "calc_length_force" if force else "calc_length"
+        return f"size += ProtoSize::{calc_fn}({self.calculate_field_id_size()}, this->{self.field_name}_len);"
 
 
 class PointerToStringBufferType(PointerToBufferTypeBase):
@@ -2679,6 +2681,16 @@ def build_message_type(
         and get_opt(desc, inline_opt, False)
     )
 
+    # Check if this message wants speed-optimized encode/calculate_size.
+    # When set, __attribute__((optimize("O2"))) is added to the definitions
+    # so GCC inlines the small ProtoEncode helpers even under -Os.
+    is_speed_optimized = get_opt(desc, pb.speed_optimized, False)
+    speed_attr = (
+        '__attribute__((optimize("O2")))  // NOLINT(clang-diagnostic-unknown-attributes)\n'
+        if is_speed_optimized
+        else ""
+    )
+
     # Only generate encode method if this message needs encoding and has fields
     if needs_encode and encode and not is_inline_only:
         # Add PROTO_ENCODE_DEBUG_ARG after pos in all proto_* calls
@@ -2688,7 +2700,7 @@ def build_message_type(
             )
             for line in encode
         ]
-        o = f"uint8_t *{desc.name}::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {{\n"
+        o = f"{speed_attr}uint8_t *{desc.name}::encode(ProtoWriteBuffer &buffer PROTO_ENCODE_DEBUG_PARAM) const {{\n"
         o += "  uint8_t *__restrict__ pos = buffer.get_pos();\n"
         o += indent("\n".join(encode_debug)) + "\n"
         o += "  return pos;\n"
@@ -2702,7 +2714,7 @@ def build_message_type(
 
     # Add calculate_size method only if this message needs encoding and has fields
     if needs_encode and size_calc and not is_inline_only:
-        o = f"uint32_t {desc.name}::calculate_size() const {{\n"
+        o = f"{speed_attr}uint32_t {desc.name}::calculate_size() const {{\n"
         o += "  uint32_t size = 0;\n"
         o += indent("\n".join(size_calc)) + "\n"
         o += "  return size;\n"
diff --git a/tests/component_tests/packages/test_packages.py b/tests/component_tests/packages/test_packages.py
index 0b828d757e..cd91c4d8cb 100644
--- a/tests/component_tests/packages/test_packages.py
+++ b/tests/component_tests/packages/test_packages.py
@@ -1106,6 +1106,51 @@ def test_packages_invalid_type_raises() -> None:
         do_packages_pass(config)
 
 
+@patch("esphome.components.packages.resolve_include")
+def test_packages_include_file_resolves_to_list(mock_resolve_include) -> None:
+    """When packages: is an IncludeFile that resolves to a list, it is processed correctly."""
+    include_file = MagicMock(spec=IncludeFile)
+    package_content = {CONF_WIFI: {CONF_SSID: TEST_PACKAGE_WIFI_SSID}}
+    mock_resolve_include.return_value = ([package_content], None)
+
+    config = {CONF_PACKAGES: include_file}
+    result = do_packages_pass(config)
+    result = merge_packages(result)
+
+    assert result == {CONF_WIFI: {CONF_SSID: TEST_PACKAGE_WIFI_SSID}}
+
+
+@patch("esphome.components.packages.resolve_include")
+def test_packages_include_file_resolves_to_dict(mock_resolve_include) -> None:
+    """When packages: is an IncludeFile that resolves to a dict, it is processed correctly."""
+    include_file = MagicMock(spec=IncludeFile)
+    package_content = {CONF_WIFI: {CONF_SSID: TEST_PACKAGE_WIFI_SSID}}
+    mock_resolve_include.return_value = ({"network": package_content}, None)
+
+    config = {CONF_PACKAGES: include_file}
+    result = do_packages_pass(config)
+    result = merge_packages(result)
+
+    assert result == {CONF_WIFI: {CONF_SSID: TEST_PACKAGE_WIFI_SSID}}
+
+
+@patch("esphome.components.packages.resolve_include")
+def test_packages_include_file_resolves_to_invalid_type_raises(
+    mock_resolve_include,
+) -> None:
+    """When packages: is an IncludeFile that resolves to an invalid type, cv.Invalid is raised."""
+    include_file = MagicMock(spec=IncludeFile)
+    mock_resolve_include.return_value = ("not_a_dict_or_list", None)
+
+    config = {CONF_PACKAGES: include_file}
+    with pytest.raises(
+        cv.Invalid, match="Packages must be a key to value mapping or list"
+    ) as exc_info:
+        do_packages_pass(config)
+
+    assert exc_info.value.path == [CONF_PACKAGES]
+
+
 @pytest.mark.parametrize(
     "invalid_package",
     [
diff --git a/tests/components/globals/common.yaml b/tests/components/globals/common.yaml
index 35dca0624f..6d5721d3be 100644
--- a/tests/components/globals/common.yaml
+++ b/tests/components/globals/common.yaml
@@ -4,6 +4,14 @@ esphome:
       - globals.set:
           id: glob_int
           value: "10"
+      # Set a float global with an integer literal - must emit the correct
+      # return type so TemplatableFn stores a direct function pointer.
+      - globals.set:
+          id: glob_float
+          value: "102"
+      - globals.set:
+          id: glob_float
+          value: !lambda "return 42;"
 
 globals:
   - id: glob_int
diff --git a/tests/integration/fixtures/addressable_light_transition.yaml b/tests/integration/fixtures/addressable_light_transition.yaml
new file mode 100644
index 0000000000..7b847dd803
--- /dev/null
+++ b/tests/integration/fixtures/addressable_light_transition.yaml
@@ -0,0 +1,29 @@
+esphome:
+  name: addr-light-transition
+host:
+api:
+logger:
+  level: DEBUG
+
+external_components:
+  - source:
+      type: local
+      path: EXTERNAL_COMPONENT_PATH
+
+light:
+  - platform: mock_addressable_light
+    output_id: strip_output
+    id: strip
+    name: "Test Strip"
+    num_leds: 4
+    gamma_correct: 2.8
+    default_transition_length: 0s
+
+sensor:
+  - platform: template
+    name: "led0_red_raw"
+    id: led0_red_raw
+    update_interval: 10ms
+    accuracy_decimals: 0
+    lambda: |-
+      return (float) id(strip_output).get_raw_red(0);
diff --git a/tests/integration/fixtures/external_components/mock_addressable_light/__init__.py b/tests/integration/fixtures/external_components/mock_addressable_light/__init__.py
new file mode 100644
index 0000000000..e8cfff8e1f
--- /dev/null
+++ b/tests/integration/fixtures/external_components/mock_addressable_light/__init__.py
@@ -0,0 +1 @@
+CODEOWNERS = ["@esphome/tests"]
diff --git a/tests/integration/fixtures/external_components/mock_addressable_light/light.py b/tests/integration/fixtures/external_components/mock_addressable_light/light.py
new file mode 100644
index 0000000000..293d2854f4
--- /dev/null
+++ b/tests/integration/fixtures/external_components/mock_addressable_light/light.py
@@ -0,0 +1,23 @@
+import esphome.codegen as cg
+from esphome.components import light
+import esphome.config_validation as cv
+from esphome.const import CONF_NUM_LEDS, CONF_OUTPUT_ID
+from esphome.types import ConfigType
+
+mock_addressable_light_ns = cg.esphome_ns.namespace("mock_addressable_light")
+MockAddressableLight = mock_addressable_light_ns.class_(
+    "MockAddressableLight", light.AddressableLight
+)
+
+CONFIG_SCHEMA = light.ADDRESSABLE_LIGHT_SCHEMA.extend(
+    {
+        cv.GenerateID(CONF_OUTPUT_ID): cv.declare_id(MockAddressableLight),
+        cv.Optional(CONF_NUM_LEDS, default=4): cv.positive_not_null_int,
+    }
+)
+
+
+async def to_code(config: ConfigType) -> None:
+    var = cg.new_Pvariable(config[CONF_OUTPUT_ID], config[CONF_NUM_LEDS])
+    await light.register_light(var, config)
+    await cg.register_component(var, config)
diff --git a/tests/integration/fixtures/external_components/mock_addressable_light/mock_addressable_light.h b/tests/integration/fixtures/external_components/mock_addressable_light/mock_addressable_light.h
new file mode 100644
index 0000000000..c6b0d10601
--- /dev/null
+++ b/tests/integration/fixtures/external_components/mock_addressable_light/mock_addressable_light.h
@@ -0,0 +1,52 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+
+#include "esphome/components/light/addressable_light.h"
+#include "esphome/core/component.h"
+
+namespace esphome::mock_addressable_light {
+
+// In-memory addressable light for host-mode integration tests. Exposes the raw
+// per-LED byte buffer (post-gamma-correction, as the hardware would see it)
+// so tests can observe transition behavior without real hardware.
+class MockAddressableLight : public light::AddressableLight {
+ public:
+  explicit MockAddressableLight(uint16_t num_leds)
+      : num_leds_(num_leds), buf_(new uint8_t[num_leds * 4]()), effect_data_(new uint8_t[num_leds]()) {}
+
+  void setup() override {}
+  void write_state(light::LightState *state) override {}
+  int32_t size() const override { return this->num_leds_; }
+  void clear_effect_data() override {
+    for (uint16_t i = 0; i < this->num_leds_; i++)
+      this->effect_data_[i] = 0;
+  }
+  light::LightTraits get_traits() override {
+    auto traits = light::LightTraits();
+    traits.set_supported_color_modes({light::ColorMode::RGB});
+    return traits;
+  }
+
+  // Accessors for tests: return the raw stored byte (post gamma correction),
+  // which is what actual LED hardware would receive.
+  uint8_t get_raw_red(uint16_t index) const { return this->buf_[index * 4 + 0]; }
+  uint8_t get_raw_green(uint16_t index) const { return this->buf_[index * 4 + 1]; }
+  uint8_t get_raw_blue(uint16_t index) const { return this->buf_[index * 4 + 2]; }
+  uint8_t get_raw_white(uint16_t index) const { return this->buf_[index * 4 + 3]; }
+
+ protected:
+  light::ESPColorView get_view_internal(int32_t index) const override {
+    size_t pos = index * 4;
+    return {this->buf_.get() + pos + 0, this->buf_.get() + pos + 1,       this->buf_.get() + pos + 2,
+            this->buf_.get() + pos + 3, this->effect_data_.get() + index, &this->correction_};
+  }
+
+  uint16_t num_leds_;
+  std::unique_ptr<uint8_t[]> buf_;
+  std::unique_ptr<uint8_t[]> effect_data_;
+};
+
+}  // namespace esphome::mock_addressable_light
diff --git a/tests/integration/test_addressable_light_transition.py b/tests/integration/test_addressable_light_transition.py
new file mode 100644
index 0000000000..37fecde595
--- /dev/null
+++ b/tests/integration/test_addressable_light_transition.py
@@ -0,0 +1,119 @@
+"""Integration test for addressable light transitions with gamma correction.
+
+Regression test for a bug where a long turn-on transition on an addressable
+light with gamma correction (e.g. gamma_correct: 2.8) produced no visible
+output for ~90% of the transition duration, then jumped to the target in the
+final ~10%. Root cause: the transition algorithm read each LED's current value
+back through the 8-bit stored byte every step; at gamma 2.8 any pre-gamma value
+below ~27 rounds to stored byte 0, so the stored byte stalled at 0 until
+progress was high enough for a single step to produce a large-enough pre-gamma
+value to clear the gamma threshold.
+
+The fix interpolates against a cached start color when all LEDs started at the
+same value (the common case for plain turn_on/turn_off), avoiding the round-trip.
+
+This test uses a host-only mock addressable light that exposes the raw stored
+byte of each LED, so we can observe the transition directly.
+"""
+
+from __future__ import annotations
+
+import asyncio
+
+from aioesphomeapi import LightInfo, SensorInfo, SensorState
+import pytest
+
+from .state_utils import InitialStateHelper, require_entity
+from .types import APIClientConnectedFactory, RunCompiledFunction
+
+
+@pytest.mark.asyncio
+async def test_addressable_light_transition(
+    yaml_config: str,
+    run_compiled: RunCompiledFunction,
+    api_client_connected: APIClientConnectedFactory,
+) -> None:
+    """With gamma 2.8, the stored raw byte must rise visibly well before the end."""
+    async with run_compiled(yaml_config), api_client_connected() as client:
+        entities, _ = await client.list_entities_services()
+        light = require_entity(entities, "test_strip", LightInfo)
+        sensor = require_entity(entities, "led0_red_raw", SensorInfo)
+
+        # Track the raw-byte sensor. It polls every 10ms in the fixture, and
+        # ESPHome sensors publish on every change, so we collect a time series.
+        # Samples are stored as absolute (loop_time, value); we rebase to the
+        # command-issue time after the run so pre-command samples are strictly
+        # negative and reliably excluded.
+        loop = asyncio.get_running_loop()
+        samples: list[tuple[float, float]] = []
+
+        def on_state(state: object) -> None:
+            if not isinstance(state, SensorState) or state.key != sensor.key:
+                return
+            samples.append((loop.time(), state.state))
+
+        # InitialStateHelper swallows the first state ESPHome sends per entity
+        # on subscribe, so on_state only sees real post-subscribe updates.
+        initial_state_helper = InitialStateHelper(entities)
+        client.subscribe_states(initial_state_helper.on_state_wrapper(on_state))
+        await initial_state_helper.wait_for_initial_states()
+
+        # Start transition: off -> full white over 1 second. This is the
+        # scenario from the bug report, compressed in time.
+        transition_s = 1.0
+        command_time = loop.time()
+        client.light_command(
+            key=light.key,
+            state=True,
+            rgb=(1.0, 1.0, 1.0),
+            brightness=1.0,
+            transition_length=transition_s,
+        )
+
+        # Let the full transition run, plus margin for the final sample.
+        await asyncio.sleep(transition_s + 0.2)
+
+        # Rebase to command-issue time. Pre-command samples have t < 0 and are
+        # excluded; everything else is in seconds since the command was issued.
+        post_command = [
+            (t - command_time, v) for (t, v) in samples if t >= command_time
+        ]
+        assert post_command, "no sensor samples received after command was issued"
+
+        # Assertion 1: the transition is not stalled. With the bug, the raw
+        # byte stays at 0 until ~90% of the transition duration. With the fix,
+        # it becomes nonzero in the first ~30% (for gamma 2.8, pre-gamma 76
+        # clears the gamma threshold at progress ~0.30). Require the first
+        # nonzero sample to land well before 50% of the transition duration,
+        # measured from the command-issue time. The 50% bound (rather than
+        # 70%) leaves headroom for assertion 2's mid-window check.
+        first_nonzero = next(((t, v) for (t, v) in post_command if v > 0), None)
+        assert first_nonzero is not None, (
+            "raw byte never rose above 0 during the transition — the fade stalled"
+        )
+        assert first_nonzero[0] < transition_s * 0.5, (
+            f"raw byte only rose above 0 at t={first_nonzero[0]:.3f}s "
+            f"(>{transition_s * 0.5:.3f}s after command) — transition is stalling"
+        )
+
+        # Assertion 2: by mid-late transition, the raw byte should have reached
+        # a substantial fraction of its final value. Bound the window to
+        # [50%, 90%] of the transition so the post-transition settled value
+        # (which always reaches 255) can't satisfy this assertion — that would
+        # let "stays at 0 then jumps at 99%" regressions slip through.
+        mid_window = [
+            v
+            for (t, v) in post_command
+            if transition_s * 0.5 <= t <= transition_s * 0.9
+        ]
+        assert mid_window, "no samples captured in mid-transition window"
+        assert max(mid_window) >= 100, (
+            f"raw byte peaked at only {max(mid_window)} between 50%–90% of "
+            "transition (expected >= 100 for white target at gamma 2.8)"
+        )
+
+        # Assertion 3: final value reaches target. Gamma 2.8 of 255 is 255.
+        final_samples = [v for (_, v) in post_command[-5:]]
+        assert max(final_samples) >= 250, (
+            f"final raw byte was {max(final_samples)}, expected >= 250"
+        )
diff --git a/tests/unit_tests/fixtures/substitutions/13-packages_as_included_list.approved.yaml b/tests/unit_tests/fixtures/substitutions/13-packages_as_included_list.approved.yaml
new file mode 100644
index 0000000000..7863def190
--- /dev/null
+++ b/tests/unit_tests/fixtures/substitutions/13-packages_as_included_list.approved.yaml
@@ -0,0 +1,3 @@
+wifi:
+  password: pkg_password
+  ssid: main_ssid
diff --git a/tests/unit_tests/fixtures/substitutions/13-packages_as_included_list.input.yaml b/tests/unit_tests/fixtures/substitutions/13-packages_as_included_list.input.yaml
new file mode 100644
index 0000000000..7a3b4970db
--- /dev/null
+++ b/tests/unit_tests/fixtures/substitutions/13-packages_as_included_list.input.yaml
@@ -0,0 +1,4 @@
+packages: !include 13-packages_list.yaml
+
+wifi:
+  ssid: main_ssid
diff --git a/tests/unit_tests/fixtures/substitutions/13-packages_list.yaml b/tests/unit_tests/fixtures/substitutions/13-packages_list.yaml
new file mode 100644
index 0000000000..23161db3d3
--- /dev/null
+++ b/tests/unit_tests/fixtures/substitutions/13-packages_list.yaml
@@ -0,0 +1,2 @@
+- wifi:
+    password: pkg_password
diff --git a/tests/unit_tests/fixtures/substitutions/14-packages_as_included_dict.approved.yaml b/tests/unit_tests/fixtures/substitutions/14-packages_as_included_dict.approved.yaml
new file mode 100644
index 0000000000..7863def190
--- /dev/null
+++ b/tests/unit_tests/fixtures/substitutions/14-packages_as_included_dict.approved.yaml
@@ -0,0 +1,3 @@
+wifi:
+  password: pkg_password
+  ssid: main_ssid
diff --git a/tests/unit_tests/fixtures/substitutions/14-packages_as_included_dict.input.yaml b/tests/unit_tests/fixtures/substitutions/14-packages_as_included_dict.input.yaml
new file mode 100644
index 0000000000..8b9fc5ec3a
--- /dev/null
+++ b/tests/unit_tests/fixtures/substitutions/14-packages_as_included_dict.input.yaml
@@ -0,0 +1,4 @@
+packages: !include 14-packages_dict.yaml
+
+wifi:
+  ssid: main_ssid
diff --git a/tests/unit_tests/fixtures/substitutions/14-packages_dict.yaml b/tests/unit_tests/fixtures/substitutions/14-packages_dict.yaml
new file mode 100644
index 0000000000..55e8b38a43
--- /dev/null
+++ b/tests/unit_tests/fixtures/substitutions/14-packages_dict.yaml
@@ -0,0 +1,3 @@
+network:
+  wifi:
+    password: pkg_password
diff --git a/tests/unit_tests/test_main.py b/tests/unit_tests/test_main.py
index 85536d2f1c..e07b4accf2 100644
--- a/tests/unit_tests/test_main.py
+++ b/tests/unit_tests/test_main.py
@@ -1231,6 +1231,48 @@ def test_upload_using_esptool_path_conversion(
     assert partitions_path.endswith("partitions.bin")
 
 
+def test_upload_using_esptool_skips_missing_extra_flash_images(
+    tmp_path: Path,
+    mock_run_external_command_main: Mock,
+    mock_get_idedata: Mock,
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    """A non-existent path in extra_flash_images must be filtered out with a
+    warning, and must not appear in the esptool command line. Only the valid
+    images are flashed. Regression test for
+    https://github.com/esphome/esphome/issues/15634.
+    """
+    setup_core(platform=PLATFORM_ESP32, tmp_path=tmp_path, name="test")
+    CORE.data[KEY_ESP32] = {KEY_VARIANT: VARIANT_ESP32}
+
+    missing_path = tmp_path / "variants" / "tasmota" / "tinyuf2.bin"
+
+    mock_idedata = MagicMock(spec=platformio_api.IDEData)
+    mock_idedata.firmware_bin_path = tmp_path / "firmware.bin"
+    mock_idedata.extra_flash_images = [
+        platformio_api.FlashImage(path=tmp_path / "bootloader.bin", offset="0x1000"),
+        platformio_api.FlashImage(path=missing_path, offset="0x2d0000"),
+    ]
+    mock_get_idedata.return_value = mock_idedata
+
+    (tmp_path / "firmware.bin").touch()
+    (tmp_path / "bootloader.bin").touch()
+    # Intentionally do NOT create missing_path
+
+    config = {CONF_ESPHOME: {"platformio_options": {}}}
+
+    with caplog.at_level(logging.WARNING, logger="esphome.__main__"):
+        result = upload_using_esptool(config, "/dev/ttyUSB0", None, None)
+
+    assert result == 0
+    assert "Skipping missing flash image" in caplog.text
+    assert str(missing_path) in caplog.text
+
+    cmd_list = list(mock_run_external_command_main.call_args[0][1:])
+    assert str(missing_path) not in cmd_list
+    assert "0x2d0000" not in cmd_list
+
+
 def test_upload_using_esptool_with_file_path(
     tmp_path: Path,
     mock_run_external_command_main: Mock,