From f1d3be4bdaaa1d6fcf7119d075b41cdc57ea31eb Mon Sep 17 00:00:00 2001
From: Kevin Ahrendt <kevin.ahrendt@openhomefoundation.org>
Date: Thu, 30 Apr 2026 12:03:40 -0400
Subject: [PATCH 01/19] [core] Simplify RAMAllocator and add internal fallback
 to external mode (#16171)

---
 esphome/core/helpers.h | 61 ++++++++++++++++++++++++++++--------------
 1 file changed, 41 insertions(+), 20 deletions(-)

diff --git a/esphome/core/helpers.h b/esphome/core/helpers.h
index 355db6c7f4..07bcb7a74f 100644
--- a/esphome/core/helpers.h
+++ b/esphome/core/helpers.h
@@ -2045,7 +2045,8 @@ void delay_microseconds_safe(uint32_t us);
  * Returns `nullptr` in case no memory is available.
  *
  * By setting flags, it can be configured to:
- * - perform external allocation falling back to main memory if SPI RAM is full or unavailable
+ * - perform external allocation falling back to internal memory if SPI RAM is full or unavailable (default)
+ * - perform internal allocation falling back to external memory (with PREFER_INTERNAL)
  * - perform external allocation only
  * - perform internal allocation only
  */
@@ -2054,16 +2055,26 @@ template<class T> class RAMAllocator {
   using value_type = T;
 
   enum Flags {
-    NONE = 0,                 // Perform external allocation and fall back to internal memory
-    ALLOC_EXTERNAL = 1 << 0,  // Perform external allocation only.
-    ALLOC_INTERNAL = 1 << 1,  // Perform internal allocation only.
-    ALLOW_FAILURE = 1 << 2,   // Does nothing. Kept for compatibility.
+    NONE = 0,                  // Perform external allocation and fall back to internal memory
+    ALLOC_EXTERNAL = 1 << 0,   // Perform external allocation only.
+    ALLOC_INTERNAL = 1 << 1,   // Perform internal allocation only.
+    ALLOW_FAILURE = 1 << 2,    // Does nothing. Kept for compatibility.
+    PREFER_INTERNAL = 1 << 3,  // Perform internal allocation and fall back to external memory
   };
 
   constexpr RAMAllocator() = default;
-  constexpr RAMAllocator(uint8_t flags)
-      : flags_((flags & (ALLOC_INTERNAL | ALLOC_EXTERNAL)) != 0 ? (flags & (ALLOC_INTERNAL | ALLOC_EXTERNAL))
-                                                                : (ALLOC_INTERNAL | ALLOC_EXTERNAL)) {}
+  constexpr RAMAllocator(uint8_t flags) {
+    if (flags & PREFER_INTERNAL) {
+      this->flags_ = ALLOC_INTERNAL | ALLOC_EXTERNAL | PREFER_INTERNAL;
+      return;
+    }
+    const uint8_t alloc_bits = flags & (ALLOC_INTERNAL | ALLOC_EXTERNAL);
+    if (alloc_bits != 0) {
+      this->flags_ = alloc_bits;
+      return;
+    }
+    this->flags_ = ALLOC_INTERNAL | ALLOC_EXTERNAL;
+  }
   template<class U> constexpr RAMAllocator(const RAMAllocator<U> &other) : flags_{other.flags_} {}
 
   T *allocate(size_t n) { return this->allocate(n, sizeof(T)); }
@@ -2072,12 +2083,8 @@ template<class T> class RAMAllocator {
     size_t size = n * manual_size;
     T *ptr = nullptr;
 #ifdef USE_ESP32
-    if (this->flags_ & Flags::ALLOC_EXTERNAL) {
-      ptr = static_cast<T *>(heap_caps_malloc(size, MALLOC_CAP_SPIRAM | MALLOC_CAP_8BIT));
-    }
-    if (ptr == nullptr && this->flags_ & Flags::ALLOC_INTERNAL) {
-      ptr = static_cast<T *>(heap_caps_malloc(size, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT));
-    }
+    const auto caps = this->get_caps_();
+    ptr = static_cast<T *>(heap_caps_malloc_prefer(size, 2, caps[0], caps[1]));
 #else
     // Ignore ALLOC_EXTERNAL/ALLOC_INTERNAL flags if external allocation is not supported
     ptr = static_cast<T *>(malloc(size));  // NOLINT(cppcoreguidelines-owning-memory,cppcoreguidelines-no-malloc)
@@ -2091,12 +2098,8 @@ template<class T> class RAMAllocator {
     size_t size = n * manual_size;
     T *ptr = nullptr;
 #ifdef USE_ESP32
-    if (this->flags_ & Flags::ALLOC_EXTERNAL) {
-      ptr = static_cast<T *>(heap_caps_realloc(p, size, MALLOC_CAP_SPIRAM | MALLOC_CAP_8BIT));
-    }
-    if (ptr == nullptr && this->flags_ & Flags::ALLOC_INTERNAL) {
-      ptr = static_cast<T *>(heap_caps_realloc(p, size, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT));
-    }
+    const auto caps = this->get_caps_();
+    ptr = static_cast<T *>(heap_caps_realloc_prefer(p, size, 2, caps[0], caps[1]));
 #else
     // Ignore ALLOC_EXTERNAL/ALLOC_INTERNAL flags if external allocation is not supported
     ptr = static_cast<T *>(realloc(p, size));  // NOLINT(cppcoreguidelines-owning-memory,cppcoreguidelines-no-malloc)
@@ -2147,6 +2150,24 @@ template<class T> class RAMAllocator {
   }
 
  private:
+#ifdef USE_ESP32
+  /// Returns {primary_caps, fallback_caps} for heap_caps_*_prefer based on the configured flags.
+  /// PREFER_INTERNAL implies both regions are enabled (enforced by the constructor), so when it is set
+  /// the primary is internal and the fallback is external. Otherwise the primary is whichever region
+  /// is enabled (external preferred when both are enabled), and the fallback is the other region (or
+  /// the same region when only one is enabled, making the second attempt a no-op).
+  std::array<uint32_t, 2> get_caps_() const {
+    constexpr uint32_t external_caps = MALLOC_CAP_SPIRAM | MALLOC_CAP_8BIT;
+    constexpr uint32_t internal_caps = MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT;
+    if (this->flags_ & PREFER_INTERNAL) {
+      return {internal_caps, external_caps};
+    }
+    const uint32_t primary = (this->flags_ & ALLOC_EXTERNAL) ? external_caps : internal_caps;
+    const uint32_t fallback = (this->flags_ & ALLOC_INTERNAL) ? internal_caps : external_caps;
+    return {primary, fallback};
+  }
+#endif
+
   uint8_t flags_{ALLOC_INTERNAL | ALLOC_EXTERNAL};
 };
 

From d48aad8c4d42431beb4f7f89f35bf42628128a04 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 12:27:54 -0500
Subject: [PATCH 02/19] [esp32] Replace 512B stack buffer in printf wraps with
 picolibc cookie FILE (#16170)

---
 esphome/components/esp32/printf_stubs.cpp | 77 +++++++++++++++++++----
 1 file changed, 64 insertions(+), 13 deletions(-)

diff --git a/esphome/components/esp32/printf_stubs.cpp b/esphome/components/esp32/printf_stubs.cpp
index 386fbbd79d..908b4023ea 100644
--- a/esphome/components/esp32/printf_stubs.cpp
+++ b/esphome/components/esp32/printf_stubs.cpp
@@ -13,14 +13,21 @@
  * and printf() calls in SDK components are only in debug/assert paths
  * (gpio_dump_io_configuration, ringbuf diagnostics) that are either
  * GC'd or never called. Crash backtraces and panic output are
- * unaffected — they use esp_rom_printf() which is a ROM function
+ * unaffected; they use esp_rom_printf() which is a ROM function
  * and does not go through libc.
  *
- * These stubs redirect through vsnprintf() (which uses _svfprintf_r
- * already in the binary) and fwrite(), allowing the linker to
- * dead-code eliminate _vfprintf_r.
+ * On picolibc (default for IDF >= 5 on RISC-V, IDF >= 6 everywhere) we
+ * route output through a stack-allocated cookie FILE that forwards each
+ * byte to the real target stream via fputc(). Picolibc's tinystdio
+ * vfprintf walks the FILE::put callback one character at a time, so this
+ * costs ~32 bytes of stack for the cookie struct vs. a 512-byte format
+ * buffer. The buffered path overflows the loopTask stack on IDF 6.
  *
- * Saves ~11 KB of flash.
+ * On newlib (IDF <= 5 on Xtensa) we keep the original snprintf-then-fwrite
+ * path because that loopTask stack budget has plenty of headroom for the
+ * 512-byte buffer; the picolibc-only crash above does not affect it.
+ *
+ * Saves ~11 KB of flash on newlib, ~2.8 KB on picolibc.
  *
  * To disable these wraps, set enable_full_printf: true in the esp32
  * advanced config section.
@@ -30,10 +37,55 @@
 #include <cstdarg>
 #include <cstdio>
 
+#ifndef __PICOLIBC__
 #include "esp_system.h"
+#endif
 
 namespace esphome::esp32 {}
 
+// NOLINTBEGIN(bugprone-reserved-identifier,cert-dcl37-c,cert-dcl51-cpp,readability-identifier-naming)
+extern "C" {
+
+#ifdef __PICOLIBC__
+
+#include <cstddef>
+#include <type_traits>
+
+extern int __real_vfprintf(FILE *stream, const char *fmt, va_list ap);
+
+namespace {
+
+struct CookieFile {
+  FILE base;
+  FILE *target;
+};
+
+// cookie_put() recovers CookieFile* from FILE* via reinterpret_cast, which is
+// only well-defined when FILE is the first member at offset 0 and CookieFile
+// is standard-layout.
+static_assert(offsetof(CookieFile, base) == 0, "FILE must be the first member of CookieFile");
+static_assert(std::is_standard_layout<CookieFile>::value, "CookieFile must be standard-layout");
+
+int cookie_put(char c, FILE *stream) {
+  auto *cookie = reinterpret_cast<CookieFile *>(stream);
+  return fputc(static_cast<unsigned char>(c), cookie->target);
+}
+
+const FILE COOKIE_FILE_TEMPLATE = FDEV_SETUP_STREAM(cookie_put, nullptr, nullptr, _FDEV_SETUP_WRITE);
+
+}  // namespace
+
+int __wrap_vfprintf(FILE *stream, const char *fmt, va_list ap) {
+  CookieFile cookie;
+  cookie.base = COOKIE_FILE_TEMPLATE;
+  cookie.target = stream;
+  return __real_vfprintf(&cookie.base, fmt, ap);
+}
+
+int __wrap_vprintf(const char *fmt, va_list ap) { return __wrap_vfprintf(stdout, fmt, ap); }
+
+#else  // !__PICOLIBC__
+
 static constexpr size_t PRINTF_BUFFER_SIZE = 512;
 
 // These stubs are essentially dead code at runtime — ESPHome replaces the
@@ -55,14 +107,18 @@ static int write_printf_buffer(FILE *stream, char *buf, int len) {
   return len;
 }
 
-// NOLINTBEGIN(bugprone-reserved-identifier,cert-dcl37-c,cert-dcl51-cpp,readability-identifier-naming)
-extern "C" {
-
 int __wrap_vprintf(const char *fmt, va_list ap) {
   char buf[PRINTF_BUFFER_SIZE];
   return write_printf_buffer(stdout, buf, vsnprintf(buf, sizeof(buf), fmt, ap));
 }
 
+int __wrap_vfprintf(FILE *stream, const char *fmt, va_list ap) {
+  char buf[PRINTF_BUFFER_SIZE];
+  return write_printf_buffer(stream, buf, vsnprintf(buf, sizeof(buf), fmt, ap));
+}
+
+#endif  // __PICOLIBC__
+
 int __wrap_printf(const char *fmt, ...) {
   va_list ap;
   va_start(ap, fmt);
@@ -71,11 +127,6 @@ int __wrap_printf(const char *fmt, ...) {
   return len;
 }
 
-int __wrap_vfprintf(FILE *stream, const char *fmt, va_list ap) {
-  char buf[PRINTF_BUFFER_SIZE];
-  return write_printf_buffer(stream, buf, vsnprintf(buf, sizeof(buf), fmt, ap));
-}
-
 int __wrap_fprintf(FILE *stream, const char *fmt, ...) {
   va_list ap;
   va_start(ap, fmt);

From 61261b4a592fa449a1e820a6392c4696fbc3d49a Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 12:33:22 -0500
Subject: [PATCH 03/19] [libretiny] Move HAL bodies into
 components/libretiny/hal.cpp + inline trivial dispatches (#16113)

---
 esphome/components/libretiny/core.cpp | 53 +--------------------------
 esphome/components/libretiny/hal.cpp  | 53 +++++++++++++++++++++++++++
 esphome/core/hal/hal_libretiny.h      | 18 +++++++--
 3 files changed, 69 insertions(+), 55 deletions(-)
 create mode 100644 esphome/components/libretiny/hal.cpp

diff --git a/esphome/components/libretiny/core.cpp b/esphome/components/libretiny/core.cpp
index f46abe3b81..8686a41e64 100644
--- a/esphome/components/libretiny/core.cpp
+++ b/esphome/components/libretiny/core.cpp
@@ -1,55 +1,6 @@
 #ifdef USE_LIBRETINY
 
-#include "core.h"
-#include "esphome/core/defines.h"
-#include "esphome/core/hal.h"
-#include "esphome/core/helpers.h"
-#include "preferences.h"
-
-#include <FreeRTOS.h>
-#include <task.h>
-
-void setup();
-void loop();
-
-namespace esphome {
-
-// yield(), delay(), micros(), millis(), millis_64() inlined in hal.h.
-void IRAM_ATTR HOT delayMicroseconds(uint32_t us) { ::delayMicroseconds(us); }
-
-void arch_init() {
-  libretiny::setup_preferences();
-  lt_wdt_enable(10000L);
-#ifdef USE_BK72XX
-  // BK72xx SDK creates the main Arduino task at priority 3, which is lower than
-  // all WiFi (4-5), LwIP (4), and TCP/IP (7) tasks. This causes ~100ms loop
-  // stalls whenever WiFi background processing runs, because the main task
-  // cannot resume until every higher-priority task finishes.
-  //
-  // By contrast, RTL87xx creates the main task at osPriorityRealtime (highest).
-  //
-  // Raise to priority 6: above WiFi/LwIP tasks (4-5) so they don't preempt the
-  // main loop, but below the TCP/IP thread (7) so packet processing keeps priority.
-  // This is safe because ESPHome yields voluntarily via wakeable_delay() and
-  // the Arduino mainTask yield() after each loop() iteration.
-  static constexpr UBaseType_t MAIN_TASK_PRIORITY = 6;
-  static_assert(MAIN_TASK_PRIORITY < configMAX_PRIORITIES, "MAIN_TASK_PRIORITY must be less than configMAX_PRIORITIES");
-  vTaskPrioritySet(nullptr, MAIN_TASK_PRIORITY);
-#endif
-#if LT_GPIO_RECOVER
-  lt_gpio_recover();
-#endif
-}
-
-void arch_restart() {
-  lt_reboot();
-  while (1) {
-  }
-}
-void HOT arch_feed_wdt() { lt_wdt_feed(); }
-uint32_t arch_get_cpu_cycle_count() { return lt_cpu_get_cycle_count(); }
-uint32_t arch_get_cpu_freq_hz() { return lt_cpu_get_freq(); }
-
-}  // namespace esphome
+// HAL functions live in hal.cpp. core.cpp is intentionally empty for
+// libretiny — there is no extra component bootstrap to keep here.
 
 #endif  // USE_LIBRETINY
diff --git a/esphome/components/libretiny/hal.cpp b/esphome/components/libretiny/hal.cpp
new file mode 100644
index 0000000000..e6dbb7296c
--- /dev/null
+++ b/esphome/components/libretiny/hal.cpp
@@ -0,0 +1,53 @@
+#ifdef USE_LIBRETINY
+
+#include "core.h"
+#include "esphome/core/hal.h"
+#include "preferences.h"
+
+#include <FreeRTOS.h>
+#include <task.h>
+
+// Empty libretiny namespace block to satisfy ci-custom's lint_namespace check.
+// HAL functions live in namespace esphome (root) — they are not part of the
+// libretiny component's API.
+namespace esphome::libretiny {}  // namespace esphome::libretiny
+
+namespace esphome {
+
+// yield(), delay(), micros(), millis(), millis_64(), delayMicroseconds(),
+// arch_feed_wdt(), arch_get_cpu_cycle_count(), arch_get_cpu_freq_hz()
+// inlined in core/hal/hal_libretiny.h.
+
+void arch_init() {
+  libretiny::setup_preferences();
+  lt_wdt_enable(10000L);
+#ifdef USE_BK72XX
+  // BK72xx SDK creates the main Arduino task at priority 3, which is lower than
+  // all WiFi (4-5), LwIP (4), and TCP/IP (7) tasks. This causes ~100ms loop
+  // stalls whenever WiFi background processing runs, because the main task
+  // cannot resume until every higher-priority task finishes.
+  //
+  // By contrast, RTL87xx creates the main task at osPriorityRealtime (highest).
+  //
+  // Raise to priority 6: above WiFi/LwIP tasks (4-5) so they don't preempt the
+  // main loop, but below the TCP/IP thread (7) so packet processing keeps priority.
+  // This is safe because ESPHome yields voluntarily via wakeable_delay() and
+  // the Arduino mainTask yield() after each loop() iteration.
+  static constexpr UBaseType_t MAIN_TASK_PRIORITY = 6;
+  static_assert(MAIN_TASK_PRIORITY < configMAX_PRIORITIES, "MAIN_TASK_PRIORITY must be less than configMAX_PRIORITIES");
+  vTaskPrioritySet(nullptr, MAIN_TASK_PRIORITY);
+#endif
+#if LT_GPIO_RECOVER
+  lt_gpio_recover();
+#endif
+}
+
+void arch_restart() {
+  lt_reboot();
+  while (1) {
+  }
+}
+
+}  // namespace esphome
+
+#endif  // USE_LIBRETINY
diff --git a/esphome/core/hal/hal_libretiny.h b/esphome/core/hal/hal_libretiny.h
index ecfe830fe3..db0fc11bfb 100644
--- a/esphome/core/hal/hal_libretiny.h
+++ b/esphome/core/hal/hal_libretiny.h
@@ -51,8 +51,16 @@ extern "C" void yield(void);
 extern "C" void delay(unsigned long ms);
 extern "C" unsigned long micros(void);
 extern "C" unsigned long millis(void);
+extern "C" void delayMicroseconds(unsigned int us);
 // NOLINTEND(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
 
+// Forward decls from libretiny's <lt_api.h> family for the inline arch_*
+// wrappers below. Pulling the full header would drag in the rest of the
+// LibreTiny C API.
+extern "C" void lt_wdt_feed(void);
+extern "C" uint32_t lt_cpu_get_cycle_count(void);
+extern "C" uint32_t lt_cpu_get_freq(void);
+
 namespace esphome {
 
 /// Returns true when executing inside an interrupt handler.
@@ -88,11 +96,13 @@ __attribute__((always_inline)) inline uint32_t millis() { return static_cast<uin
 #endif
 __attribute__((always_inline)) inline uint64_t millis_64() { return Millis64Impl::compute(millis()); }
 
-void delayMicroseconds(uint32_t us);  // NOLINT(readability-identifier-naming)
-void arch_feed_wdt();
-uint32_t arch_get_cpu_cycle_count();
+// NOLINTNEXTLINE(readability-identifier-naming)
+__attribute__((always_inline)) inline void delayMicroseconds(uint32_t us) { ::delayMicroseconds(us); }
+__attribute__((hot, always_inline)) inline void arch_feed_wdt() { lt_wdt_feed(); }
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_cycle_count() { return lt_cpu_get_cycle_count(); }
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_freq_hz() { return lt_cpu_get_freq(); }
+
 void arch_init();
-uint32_t arch_get_cpu_freq_hz();
 
 }  // namespace esphome
 

From 2fbfb4c385506070b2ab23d4114ad221d48da544 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:10:53 -0500
Subject: [PATCH 04/19] [ci] Split integration tests into 3 buckets when count
 is more than 10 (#16152)

---
 .github/workflows/ci.yml            | 29 +++++-----
 script/determine-jobs.py            | 66 ++++++++++++++++++++--
 tests/script/test_determine_jobs.py | 88 +++++++++++++++++++++++++++--
 3 files changed, 156 insertions(+), 27 deletions(-)

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 57053c3645..3af1709774 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -199,8 +199,7 @@ jobs:
       - common
     outputs:
       integration-tests: ${{ steps.determine.outputs.integration-tests }}
-      integration-tests-run-all: ${{ steps.determine.outputs.integration-tests-run-all }}
-      integration-test-files: ${{ steps.determine.outputs.integration-test-files }}
+      integration-test-buckets: ${{ steps.determine.outputs.integration-test-buckets }}
       clang-tidy: ${{ steps.determine.outputs.clang-tidy }}
       clang-tidy-mode: ${{ steps.determine.outputs.clang-tidy-mode }}
       python-linters: ${{ steps.determine.outputs.python-linters }}
@@ -243,8 +242,7 @@ jobs:
 
           # Extract individual fields
           echo "integration-tests=$(echo "$output" | jq -r '.integration_tests')" >> $GITHUB_OUTPUT
-          echo "integration-tests-run-all=$(echo "$output" | jq -r '.integration_tests_run_all')" >> $GITHUB_OUTPUT
-          echo "integration-test-files=$(echo "$output" | jq -c '.integration_test_files')" >> $GITHUB_OUTPUT
+          echo "integration-test-buckets=$(echo "$output" | jq -c '.integration_test_buckets')" >> $GITHUB_OUTPUT
           echo "clang-tidy=$(echo "$output" | jq -r '.clang_tidy')" >> $GITHUB_OUTPUT
           echo "clang-tidy-mode=$(echo "$output" | jq -r '.clang_tidy_mode')" >> $GITHUB_OUTPUT
           echo "python-linters=$(echo "$output" | jq -r '.python_linters')" >> $GITHUB_OUTPUT
@@ -267,12 +265,16 @@ jobs:
           key: components-graph-${{ hashFiles('esphome/components/**/*.py') }}
 
   integration-tests:
-    name: Run integration tests
+    name: Run integration tests (${{ matrix.bucket.name }})
     runs-on: ubuntu-latest
     needs:
       - common
       - determine-jobs
     if: needs.determine-jobs.outputs.integration-tests == 'true'
+    strategy:
+      fail-fast: false
+      matrix:
+        bucket: ${{ fromJson(needs.determine-jobs.outputs.integration-test-buckets) }}
     steps:
       - name: Check out code from GitHub
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -299,19 +301,14 @@ jobs:
         run: echo "::add-matcher::.github/workflows/matchers/pytest.json"
       - name: Run integration tests
         env:
-          INTEGRATION_TEST_FILES: ${{ needs.determine-jobs.outputs.integration-test-files }}
-          INTEGRATION_TESTS_RUN_ALL: ${{ needs.determine-jobs.outputs.integration-tests-run-all }}
+          # JSON array of test paths; parsed into a bash array below to avoid
+          # shell word-splitting / glob hazards.
+          BUCKET_TESTS: ${{ toJson(matrix.bucket.tests) }}
         run: |
           . venv/bin/activate
-          if [[ "$INTEGRATION_TESTS_RUN_ALL" == "true" ]]; then
-            echo "Running all integration tests"
-            pytest -vv --no-cov --tb=native -n auto tests/integration/
-          else
-            # Parse JSON array into bash array to avoid shell expansion issues
-            mapfile -t test_files < <(echo "$INTEGRATION_TEST_FILES" | jq -r '.[]')
-            echo "Running ${#test_files[@]} specific integration tests"
-            pytest -vv --no-cov --tb=native -n auto "${test_files[@]}"
-          fi
+          mapfile -t test_files < <(echo "$BUCKET_TESTS" | jq -r '.[]')
+          echo "Bucket ${{ matrix.bucket.name }}: running ${#test_files[@]} integration tests"
+          pytest -vv --no-cov --tb=native -n auto "${test_files[@]}"
 
   cpp-unit-tests:
     name: Run C++ unit tests
diff --git a/script/determine-jobs.py b/script/determine-jobs.py
index 6fd7ab297c..c0cf8ecbdc 100755
--- a/script/determine-jobs.py
+++ b/script/determine-jobs.py
@@ -6,8 +6,7 @@ what files have changed. It outputs JSON with the following structure:
 
 {
   "integration_tests": true/false,
-  "integration_tests_run_all": true/false,
-  "integration_test_files": ["tests/integration/test_foo.py", ...],
+  "integration_test_buckets": [{"name": "1/3", "tests": ["tests/integration/test_foo.py", ...]}, ...],
   "clang_tidy": true/false,
   "clang_format": true/false,
   "python_linters": true/false,
@@ -81,6 +80,62 @@ CLANG_TIDY_SPLIT_THRESHOLD = 65
 # Isolated components count as 10x, groupable components count as 1x
 COMPONENT_TEST_BATCH_SIZE = 40
 
+# Integration test bucketing: when more than the threshold tests are scheduled,
+# fan out across this many parallel jobs. Below the threshold, a single job runs.
+INTEGRATION_TESTS_SPLIT_THRESHOLD = 10
+INTEGRATION_TESTS_SPLIT_BUCKETS = 3
+
+
+def _split_list(items: list[str], n: int) -> list[list[str]]:
+    """Split a list into n roughly-equal contiguous parts (matches script/clang-tidy)."""
+    k, m = divmod(len(items), n)
+    return [items[i * k + min(i, m) : (i + 1) * k + min(i + 1, m)] for i in range(n)]
+
+
+def _all_integration_test_files() -> list[str]:
+    """Return all integration test file paths, sorted, relative to repo root."""
+    return sorted(
+        str(p.relative_to(root_path))
+        for p in (Path(root_path) / "tests" / "integration").glob("test_*.py")
+    )
+
+
+def _compute_integration_test_buckets(
+    integration_run_all: bool,
+    integration_test_files: list[str],
+) -> tuple[bool, list[dict[str, Any]]]:
+    """Compute (run_integration, buckets) from the determine_integration_tests result.
+
+    Pure function for unit testing — no I/O beyond `_all_integration_test_files`
+    when `integration_run_all` is set.
+
+    `buckets` is a list of `{name, tests}` dicts where `tests` is a JSON-friendly
+    list of file paths so the workflow can build a bash array via jq, avoiding
+    shell word-splitting / glob hazards.
+    """
+    if integration_run_all:
+        files = _all_integration_test_files()
+    else:
+        files = sorted(integration_test_files)
+
+    # Empty list (e.g. run_all expansion with no files on disk) would otherwise
+    # cause the workflow to invoke pytest with no path argument and collect
+    # tests outside tests/integration/. Suppress the run instead.
+    if not files:
+        return False, []
+
+    if len(files) > INTEGRATION_TESTS_SPLIT_THRESHOLD:
+        parts = [
+            part for part in _split_list(files, INTEGRATION_TESTS_SPLIT_BUCKETS) if part
+        ]
+        buckets = [
+            {"name": f"{i + 1}/{len(parts)}", "tests": part}
+            for i, part in enumerate(parts)
+        ]
+    else:
+        buckets = [{"name": "1/1", "tests": files}]
+    return True, buckets
+
 
 class Platform(StrEnum):
     """Platform identifiers for memory impact analysis."""
@@ -812,7 +867,9 @@ def main() -> None:
     integration_run_all, integration_test_files = determine_integration_tests(
         args.branch
     )
-    run_integration = integration_run_all or bool(integration_test_files)
+    run_integration, integration_test_buckets = _compute_integration_test_buckets(
+        integration_run_all, integration_test_files
+    )
     run_clang_tidy = should_run_clang_tidy(args.branch)
     run_clang_format = should_run_clang_format(args.branch)
     run_python_linters = should_run_python_linters(args.branch)
@@ -944,8 +1001,7 @@ def main() -> None:
 
     output: dict[str, Any] = {
         "integration_tests": run_integration,
-        "integration_tests_run_all": integration_run_all,
-        "integration_test_files": integration_test_files,
+        "integration_test_buckets": integration_test_buckets,
         "clang_tidy": run_clang_tidy,
         "clang_tidy_mode": clang_tidy_mode,
         "clang_format": run_clang_format,
diff --git a/tests/script/test_determine_jobs.py b/tests/script/test_determine_jobs.py
index 44c110b689..e85f1757b0 100644
--- a/tests/script/test_determine_jobs.py
+++ b/tests/script/test_determine_jobs.py
@@ -122,10 +122,19 @@ def test_main_all_tests_should_run(
         "esphome/helpers.py",
     ]
 
+    # Stable, deterministic stand-in for the tests/integration/ glob so the
+    # bucket assertions don't drift with the real test count.
+    fake_test_files = [f"tests/integration/test_{i:03d}.py" for i in range(15)]
+
     # Run main function with mocked argv
     with (
         patch("sys.argv", ["determine-jobs.py"]),
         patch.object(determine_jobs, "_is_clang_tidy_full_scan", return_value=False),
+        patch.object(
+            determine_jobs,
+            "_all_integration_test_files",
+            return_value=fake_test_files,
+        ),
         patch.object(
             determine_jobs,
             "get_changed_components",
@@ -161,8 +170,24 @@ def test_main_all_tests_should_run(
     output = json.loads(captured.out)
 
     assert output["integration_tests"] is True
-    assert output["integration_tests_run_all"] is True
-    assert output["integration_test_files"] == []
+    # run_all=True expands to the full glob and pre-buckets into 3 parts.
+    # Each bucket's `tests` is a JSON list of file paths.
+    assert isinstance(output["integration_test_buckets"], list)
+    assert len(output["integration_test_buckets"]) == 3
+    assert [b["name"] for b in output["integration_test_buckets"]] == [
+        "1/3",
+        "2/3",
+        "3/3",
+    ]
+    for bucket in output["integration_test_buckets"]:
+        assert isinstance(bucket["tests"], list)
+        for path in bucket["tests"]:
+            assert isinstance(path, str)
+    bucket_files = [f for b in output["integration_test_buckets"] for f in b["tests"]]
+    assert bucket_files == fake_test_files
+    # Bucket sizes are balanced (max-min difference at most 1).
+    sizes = [len(b["tests"]) for b in output["integration_test_buckets"]]
+    assert max(sizes) - min(sizes) <= 1
     assert output["clang_tidy"] is True
     assert output["clang_tidy_mode"] in ["nosplit", "split"]
     assert output["clang_format"] is True
@@ -247,8 +272,7 @@ def test_main_no_tests_should_run(
     output = json.loads(captured.out)
 
     assert output["integration_tests"] is False
-    assert output["integration_tests_run_all"] is False
-    assert output["integration_test_files"] == []
+    assert output["integration_test_buckets"] == []
     assert output["clang_tidy"] is False
     assert output["clang_tidy_mode"] == "disabled"
     assert output["clang_format"] is False
@@ -332,8 +356,7 @@ def test_main_with_branch_argument(
     output = json.loads(captured.out)
 
     assert output["integration_tests"] is False
-    assert output["integration_tests_run_all"] is False
-    assert output["integration_test_files"] == []
+    assert output["integration_test_buckets"] == []
     assert output["clang_tidy"] is True
     assert output["clang_tidy_mode"] in ["nosplit", "split"]
     assert output["clang_format"] is False
@@ -357,6 +380,59 @@ def test_main_with_branch_argument(
     assert output["cpp_unit_tests_components"] == ["mqtt"]
 
 
+def test_compute_integration_test_buckets_empty() -> None:
+    """No integration tests scheduled => (False, [])."""
+    run, buckets = determine_jobs._compute_integration_test_buckets(False, [])
+    assert run is False
+    assert buckets == []
+
+
+def test_compute_integration_test_buckets_below_threshold() -> None:
+    """A small explicit list (<= threshold) => single 1/1 bucket with that list."""
+    files = [f"tests/integration/test_{name}.py" for name in ("c", "a", "b")]
+    run, buckets = determine_jobs._compute_integration_test_buckets(False, files)
+    assert run is True
+    assert buckets == [{"name": "1/1", "tests": sorted(files)}]
+
+
+def test_compute_integration_test_buckets_at_threshold_stays_single() -> None:
+    """Exactly INTEGRATION_TESTS_SPLIT_THRESHOLD files => still one bucket
+    (the split kicks in only when count is strictly greater than threshold)."""
+    files = [
+        f"tests/integration/test_{i:02d}.py"
+        for i in range(determine_jobs.INTEGRATION_TESTS_SPLIT_THRESHOLD)
+    ]
+    run, buckets = determine_jobs._compute_integration_test_buckets(False, files)
+    assert run is True
+    assert len(buckets) == 1
+    assert buckets[0]["name"] == "1/1"
+    assert buckets[0]["tests"] == sorted(files)
+
+
+def test_compute_integration_test_buckets_just_over_threshold_splits() -> None:
+    """One file over the threshold triggers the 3-bucket fan-out, balanced."""
+    n = determine_jobs.INTEGRATION_TESTS_SPLIT_THRESHOLD + 1
+    files = [f"tests/integration/test_{i:02d}.py" for i in range(n)]
+    run, buckets = determine_jobs._compute_integration_test_buckets(False, files)
+    assert run is True
+    assert [b["name"] for b in buckets] == ["1/3", "2/3", "3/3"]
+    union = [path for b in buckets for path in b["tests"]]
+    assert union == sorted(files)
+    sizes = [len(b["tests"]) for b in buckets]
+    assert max(sizes) - min(sizes) <= 1
+
+
+def test_compute_integration_test_buckets_run_all_with_empty_glob_disables_run() -> (
+    None
+):
+    """run_all=True but glob returns no files => run suppressed (otherwise
+    pytest would collect tests outside tests/integration/)."""
+    with patch.object(determine_jobs, "_all_integration_test_files", return_value=[]):
+        run, buckets = determine_jobs._compute_integration_test_buckets(True, [])
+    assert run is False
+    assert buckets == []
+
+
 def test_determine_integration_tests(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:

From e085cb50d98a7a784050d5e2948754247d07fb9a Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:11:30 -0500
Subject: [PATCH 05/19] [sensor] Drop Component from filter classes, use
 self-keyed scheduler (#16132)

---
 esphome/components/sensor/__init__.py | 19 ++++++-------------
 esphome/components/sensor/filter.cpp  | 21 +++++++--------------
 esphome/components/sensor/filter.h    | 15 +++++----------
 3 files changed, 18 insertions(+), 37 deletions(-)

diff --git a/esphome/components/sensor/__init__.py b/esphome/components/sensor/__init__.py
index 48b7d25d4d..c18aa32f37 100644
--- a/esphome/components/sensor/__init__.py
+++ b/esphome/components/sensor/__init__.py
@@ -266,7 +266,7 @@ StreamingMovingAverageFilter = sensor_ns.class_("StreamingMovingAverageFilter",
 ExponentialMovingAverageFilter = sensor_ns.class_(
     "ExponentialMovingAverageFilter", Filter
 )
-ThrottleAverageFilter = sensor_ns.class_("ThrottleAverageFilter", Filter, cg.Component)
+ThrottleAverageFilter = sensor_ns.class_("ThrottleAverageFilter", Filter)
 LambdaFilter = sensor_ns.class_("LambdaFilter", Filter)
 StatelessLambdaFilter = sensor_ns.class_("StatelessLambdaFilter", Filter)
 OffsetFilter = sensor_ns.class_("OffsetFilter", Filter)
@@ -283,8 +283,8 @@ ThrottleWithPriorityNanFilter = sensor_ns.class_(
 TimeoutFilterBase = sensor_ns.class_("TimeoutFilterBase", Filter, cg.Component)
 TimeoutFilterLast = sensor_ns.class_("TimeoutFilterLast", TimeoutFilterBase)
 TimeoutFilterConfigured = sensor_ns.class_("TimeoutFilterConfigured", TimeoutFilterBase)
-DebounceFilter = sensor_ns.class_("DebounceFilter", Filter, cg.Component)
-HeartbeatFilter = sensor_ns.class_("HeartbeatFilter", Filter, cg.Component)
+DebounceFilter = sensor_ns.class_("DebounceFilter", Filter)
+HeartbeatFilter = sensor_ns.class_("HeartbeatFilter", Filter)
 DeltaFilter = sensor_ns.class_("DeltaFilter", Filter)
 OrFilter = sensor_ns.class_("OrFilter", Filter)
 CalibrateLinearFilter = sensor_ns.class_("CalibrateLinearFilter", Filter)
@@ -567,9 +567,7 @@ async def exponential_moving_average_filter_to_code(config, filter_id):
     "throttle_average", ThrottleAverageFilter, cv.positive_time_period_milliseconds
 )
 async def throttle_average_filter_to_code(config, filter_id):
-    var = cg.new_Pvariable(filter_id, config)
-    await cg.register_component(var, {})
-    return var
+    return cg.new_Pvariable(filter_id, config)
 
 
 @FILTER_REGISTRY.register("lambda", LambdaFilter, cv.returning_lambda)
@@ -698,13 +696,10 @@ HEARTBEAT_SCHEMA = cv.Schema(
 async def heartbeat_filter_to_code(config, filter_id):
     if isinstance(config, dict):
         var = cg.new_Pvariable(filter_id, config[CONF_PERIOD])
-        await cg.register_component(var, {})
         cg.add(var.set_optimistic(config[CONF_OPTIMISTIC]))
         return var
 
-    var = cg.new_Pvariable(filter_id, config)
-    await cg.register_component(var, {})
-    return var
+    return cg.new_Pvariable(filter_id, config)
 
 
 TIMEOUT_SCHEMA = cv.maybe_simple_value(
@@ -738,9 +733,7 @@ async def timeout_filter_to_code(config, filter_id):
     "debounce", DebounceFilter, cv.positive_time_period_milliseconds
 )
 async def debounce_filter_to_code(config, filter_id):
-    var = cg.new_Pvariable(filter_id, config)
-    await cg.register_component(var, {})
-    return var
+    return cg.new_Pvariable(filter_id, config)
 
 
 CONF_DATAPOINTS = "datapoints"
diff --git a/esphome/components/sensor/filter.cpp b/esphome/components/sensor/filter.cpp
index 4896757d3f..5f7f19769a 100644
--- a/esphome/components/sensor/filter.cpp
+++ b/esphome/components/sensor/filter.cpp
@@ -13,11 +13,6 @@ namespace esphome::sensor {
 
 static const char *const TAG = "sensor.filter";
 
-// Filter scheduler IDs.
-// Each filter is its own Component instance, so the scheduler scopes
-// IDs by component pointer — no risk of collisions between instances.
-constexpr uint32_t FILTER_ID = 0;
-
 // Filter
 void Filter::input(float value) {
   ESP_LOGVV(TAG, "Filter(%p)::input(%f)", this, value);
@@ -185,8 +180,9 @@ optional<float> ThrottleAverageFilter::new_value(float value) {
   }
   return {};
 }
-void ThrottleAverageFilter::setup() {
-  this->set_interval(FILTER_ID, this->time_period_, [this]() {
+void ThrottleAverageFilter::initialize(Sensor *parent, Filter *next) {
+  Filter::initialize(parent, next);
+  App.scheduler.set_interval(this, this->time_period_, [this]() {
     ESP_LOGVV(TAG, "ThrottleAverageFilter(%p)::interval(sum=%f, n=%i)", this, this->sum_, this->n_);
     if (this->n_ == 0) {
       if (this->have_nan_)
@@ -199,7 +195,6 @@ void ThrottleAverageFilter::setup() {
     this->have_nan_ = false;
   });
 }
-float ThrottleAverageFilter::get_setup_priority() const { return setup_priority::HARDWARE; }
 
 // LambdaFilter
 LambdaFilter::LambdaFilter(lambda_filter_t lambda_filter) : lambda_filter_(std::move(lambda_filter)) {}
@@ -362,13 +357,12 @@ optional<float> TimeoutFilterConfigured::new_value(float value) {
 
 // DebounceFilter
 optional<float> DebounceFilter::new_value(float value) {
-  this->set_timeout(FILTER_ID, this->time_period_, [this, value]() { this->output(value); });
+  App.scheduler.set_timeout(this, this->time_period_, [this, value]() { this->output(value); });
 
   return {};
 }
 
 DebounceFilter::DebounceFilter(uint32_t time_period) : time_period_(time_period) {}
-float DebounceFilter::get_setup_priority() const { return setup_priority::HARDWARE; }
 
 // HeartbeatFilter
 HeartbeatFilter::HeartbeatFilter(uint32_t time_period) : time_period_(time_period), last_input_(NAN) {}
@@ -384,8 +378,9 @@ optional<float> HeartbeatFilter::new_value(float value) {
   return {};
 }
 
-void HeartbeatFilter::setup() {
-  this->set_interval(FILTER_ID, this->time_period_, [this]() {
+void HeartbeatFilter::initialize(Sensor *parent, Filter *next) {
+  Filter::initialize(parent, next);
+  App.scheduler.set_interval(this, this->time_period_, [this]() {
     ESP_LOGVV(TAG, "HeartbeatFilter(%p)::interval(has_value=%s, last_input=%f)", this, YESNO(this->has_value_),
               this->last_input_);
     if (!this->has_value_)
@@ -395,8 +390,6 @@ void HeartbeatFilter::setup() {
   });
 }
 
-float HeartbeatFilter::get_setup_priority() const { return setup_priority::HARDWARE; }
-
 optional<float> calibrate_linear_compute(const std::array<float, 3> *functions, size_t count, float value) {
   for (size_t i = 0; i < count; i++) {
     if (!std::isfinite(functions[i][2]) || value < functions[i][2])
diff --git a/esphome/components/sensor/filter.h b/esphome/components/sensor/filter.h
index 917a1ce7d5..d61df11d9b 100644
--- a/esphome/components/sensor/filter.h
+++ b/esphome/components/sensor/filter.h
@@ -254,16 +254,14 @@ class ExponentialMovingAverageFilter : public Filter {
  *
  * It takes the average of all the values received in a period of time.
  */
-class ThrottleAverageFilter : public Filter, public Component {
+class ThrottleAverageFilter : public Filter {
  public:
   explicit ThrottleAverageFilter(uint32_t time_period);
 
-  void setup() override;
+  void initialize(Sensor *parent, Filter *next) override;
 
   optional<float> new_value(float value) override;
 
-  float get_setup_priority() const override;
-
  protected:
   float sum_{0.0f};
   unsigned int n_{0};
@@ -454,25 +452,22 @@ class TimeoutFilterConfigured : public TimeoutFilterBase {
   // Total: 8 (base) + 4 = 12 bytes + vtable ptr + Component overhead
 };
 
-class DebounceFilter : public Filter, public Component {
+class DebounceFilter : public Filter {
  public:
   explicit DebounceFilter(uint32_t time_period);
 
   optional<float> new_value(float value) override;
 
-  float get_setup_priority() const override;
-
  protected:
   uint32_t time_period_;
 };
 
-class HeartbeatFilter : public Filter, public Component {
+class HeartbeatFilter : public Filter {
  public:
   explicit HeartbeatFilter(uint32_t time_period);
 
-  void setup() override;
+  void initialize(Sensor *parent, Filter *next) override;
   optional<float> new_value(float value) override;
-  float get_setup_priority() const override;
 
   void set_optimistic(bool optimistic) { this->optimistic_ = optimistic; }
 

From 2f3e16b4823870cbf701c9273529d185b37a34eb Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:12:06 -0500
Subject: [PATCH 06/19] [bk72xx] Apply CFG_SUPPORT_BLE=0 SDK option to BK7238
 (#16181)

---
 .../components/beken_spi_led_strip/light.py   |  1 +
 esphome/components/libretiny/__init__.py      | 20 +++++++++++--------
 2 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/esphome/components/beken_spi_led_strip/light.py b/esphome/components/beken_spi_led_strip/light.py
index 31572cd800..9093b08b62 100644
--- a/esphome/components/beken_spi_led_strip/light.py
+++ b/esphome/components/beken_spi_led_strip/light.py
@@ -62,6 +62,7 @@ CONF_IS_WRGB = "is_wrgb"
 SUPPORTED_PINS = {
     libretiny.const.FAMILY_BK7231N: [16],
     libretiny.const.FAMILY_BK7231T: [16],
+    libretiny.const.FAMILY_BK7238: [16],
     libretiny.const.FAMILY_BK7251: [16],
 }
 
diff --git a/esphome/components/libretiny/__init__.py b/esphome/components/libretiny/__init__.py
index 40b8c8dc6c..74ac51d200 100644
--- a/esphome/components/libretiny/__init__.py
+++ b/esphome/components/libretiny/__init__.py
@@ -37,6 +37,7 @@ from .const import (
     CONF_UART_PORT,
     FAMILIES,
     FAMILY_BK7231N,
+    FAMILY_BK7238,
     FAMILY_COMPONENT,
     FAMILY_FRIENDLY,
     FAMILY_RTL8710B,
@@ -56,19 +57,22 @@ CODEOWNERS = ["@kuba2k2"]
 AUTO_LOAD = ["preferences"]
 IS_TARGET_PLATFORM = True
 
-# BK7231N SDK options to disable unused features.
+# BLE 5.x BK SDK options to disable unused features.
 # Disabling BLE saves ~21KB RAM and ~200KB Flash because BLE init code is
 # called unconditionally by the SDK. ESPHome doesn't use BLE on LibreTiny.
 #
-# This only works on BK7231N (BLE 5.x). Other BK72XX chips using BLE 4.2
-# (BK7231T, BK7231Q, BK7251; BK7252 boards use the BK7251 family) have a bug
-# where the BLE library still links and references undefined symbols when
-# CFG_SUPPORT_BLE=0.
+# This only works on BLE 5.x BK chips (BK7231N, BK7238). Other BK72XX chips
+# using BLE 4.2 (BK7231T, BK7231Q, BK7251; BK7252 boards use the BK7251 family)
+# have a bug where the BLE library still links and references undefined symbols
+# when CFG_SUPPORT_BLE=0.
+#
+# On BK7238 the SDK also hangs at WiFi STA enable when BLE init runs, so
+# disabling it is required for reliable boot, not just an optimization.
 #
 # Other options like CFG_TX_EVM_TEST, CFG_RX_SENSITIVITY_TEST, CFG_SUPPORT_BKREG,
 # CFG_SUPPORT_OTA_HTTP, and CFG_USE_SPI_SLAVE were evaluated but provide no  # NOLINT
 # measurable benefit - the linker already strips unreferenced code via -gc-sections.
-_BK7231N_SYS_CONFIG_OPTIONS = [
+_BLE5_BK_SYS_CONFIG_OPTIONS = [
     "CFG_SUPPORT_BLE=0",
 ]
 
@@ -549,9 +553,9 @@ async def component_to_code(config):
         cg.add_platformio_option("custom_fw_version", __version__)
 
     # Apply chip-specific SDK options to save RAM/Flash
-    if config[CONF_FAMILY] == FAMILY_BK7231N:
+    if config[CONF_FAMILY] in (FAMILY_BK7231N, FAMILY_BK7238):
         cg.add_platformio_option(
-            "custom_options.sys_config#h", _BK7231N_SYS_CONFIG_OPTIONS
+            "custom_options.sys_config#h", _BLE5_BK_SYS_CONFIG_OPTIONS
         )
 
     # Tune lwIP for ESPHome's actual needs.

From 3b3e003aa3312635ce22454703d386d49563fdca Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:13:10 -0500
Subject: [PATCH 07/19] [sensor] Pack ThrottleAverageFilter have_nan_ into n_
 bitfield (-4 B/instance) (#16169)

---
 esphome/components/sensor/__init__.py | 7 ++++++-
 esphome/components/sensor/filter.h    | 7 +++++--
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/esphome/components/sensor/__init__.py b/esphome/components/sensor/__init__.py
index c18aa32f37..ed02cc2543 100644
--- a/esphome/components/sensor/__init__.py
+++ b/esphome/components/sensor/__init__.py
@@ -564,7 +564,12 @@ async def exponential_moving_average_filter_to_code(config, filter_id):
 
 
 @FILTER_REGISTRY.register(
-    "throttle_average", ThrottleAverageFilter, cv.positive_time_period_milliseconds
+    "throttle_average",
+    ThrottleAverageFilter,
+    cv.All(
+        cv.positive_time_period_milliseconds,
+        cv.Range(max=cv.TimePeriod(hours=24)),
+    ),
 )
 async def throttle_average_filter_to_code(config, filter_id):
     return cg.new_Pvariable(filter_id, config)
diff --git a/esphome/components/sensor/filter.h b/esphome/components/sensor/filter.h
index d61df11d9b..57a2386a7f 100644
--- a/esphome/components/sensor/filter.h
+++ b/esphome/components/sensor/filter.h
@@ -264,9 +264,12 @@ class ThrottleAverageFilter : public Filter {
 
  protected:
   float sum_{0.0f};
-  unsigned int n_{0};
   uint32_t time_period_;
-  bool have_nan_{false};
+  // Sample count packed with NaN-seen flag in a single 32-bit word.
+  // n_ is bounded by YAML cap on time_period_ (24 h) × max plausible source
+  // rate (1 kHz) = 86.4M ≪ 2^31, so 31 bits has 25x headroom.
+  uint32_t n_ : 31 {0};
+  uint32_t have_nan_ : 1 {0};
 };
 
 using lambda_filter_t = std::function<optional<float>(float)>;

From 45e78e4114a9dba7628fca434163c8e4795670b8 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:13:54 -0500
Subject: [PATCH 08/19] [core] Inline loop gate expression to avoid stale local
 reuse (#16167)

---
 esphome/core/application.h | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/esphome/core/application.h b/esphome/core/application.h
index 4a18714d0d..5baf570e62 100644
--- a/esphome/core/application.h
+++ b/esphome/core/application.h
@@ -637,10 +637,12 @@ inline void ESPHOME_ALWAYS_INLINE Application::loop() {
   // flag preserves it. wake_request_take() exchange-clears the flag; wakes
   // that arrive during Phase B re-set it and run Phase B again on the next
   // iteration.
-  const bool high_frequency = HighFrequencyLoopRequester::is_high_frequency();
-  const uint32_t elapsed = now - this->last_loop_;
-  const bool woke = esphome::wake_request_take();
-  const bool do_component_phase = high_frequency || woke || (elapsed >= this->loop_interval_);
+  //
+  // wake_request_take() must always be called first since it does an
+  // atomic exchange to clear the flag, and we want to run the component phase
+  // if either the flag was set or the scheduler requested a high-frequency loop.
+  const bool do_component_phase = esphome::wake_request_take() || HighFrequencyLoopRequester::is_high_frequency() ||
+                                  (now - this->last_loop_ >= this->loop_interval_);
 
   if (do_component_phase) {
     ComponentPhaseGuard phase_guard{*this};

From 148d478decdbc07c69442162799d4a28a5e41c8d Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:14:20 -0500
Subject: [PATCH 09/19] [api] Add encode/decode benchmarks for Z-Wave, IR/RF,
 and serial proxy messages (#16157)

---
 tests/benchmarks/components/api/__init__.py   |  14 +-
 .../components/api/bench_proto_proxy.cpp      | 280 ++++++++++++++++++
 .../esphome/components/infrared/infrared.h    |  45 +++
 .../radio_frequency/radio_frequency.h         |  51 ++++
 .../components/serial_proxy/serial_proxy.h    |  46 +++
 .../components/zwave_proxy/zwave_proxy.h      |  29 ++
 6 files changed, 462 insertions(+), 3 deletions(-)
 create mode 100644 tests/benchmarks/components/api/bench_proto_proxy.cpp
 create mode 100644 tests/benchmarks/stubs/esphome/components/infrared/infrared.h
 create mode 100644 tests/benchmarks/stubs/esphome/components/radio_frequency/radio_frequency.h
 create mode 100644 tests/benchmarks/stubs/esphome/components/serial_proxy/serial_proxy.h
 create mode 100644 tests/benchmarks/stubs/esphome/components/zwave_proxy/zwave_proxy.h

diff --git a/tests/benchmarks/components/api/__init__.py b/tests/benchmarks/components/api/__init__.py
index eb86492964..0d02e0b054 100644
--- a/tests/benchmarks/components/api/__init__.py
+++ b/tests/benchmarks/components/api/__init__.py
@@ -11,11 +11,19 @@ def override_manifest(manifest: ComponentManifestOverride) -> None:
 
     async def to_code(config):
         await original_to_code(config)
-        # Enable BLE proto message types for benchmarks.  The real
-        # bluetooth_proxy component is ESP32-only; a lightweight stub
-        # header in tests/benchmarks/stubs/ satisfies the include.
+        # Enable proxy proto message types for benchmarks.  The real
+        # components have hardware dependencies (BLE/UART/RMT); lightweight
+        # stub headers in tests/benchmarks/stubs/ satisfy the includes.
         cg.add_define("USE_BLUETOOTH_PROXY")
         cg.add_define("BLUETOOTH_PROXY_MAX_CONNECTIONS", 3)
         cg.add_define("BLUETOOTH_PROXY_ADVERTISEMENT_BATCH_SIZE", 16)
+        cg.add_define("USE_ZWAVE_PROXY")
+        cg.add_define("USE_INFRARED")
+        cg.add_define("USE_IR_RF")
+        cg.add_define("USE_RADIO_FREQUENCY")
+        cg.add_define("USE_SERIAL_PROXY")
+        cg.add_define("SERIAL_PROXY_COUNT", 0)
+        cg.add_define("ESPHOME_ENTITY_INFRARED_COUNT", 0)
+        cg.add_define("ESPHOME_ENTITY_RADIO_FREQUENCY_COUNT", 0)
 
     manifest.to_code = to_code
diff --git a/tests/benchmarks/components/api/bench_proto_proxy.cpp b/tests/benchmarks/components/api/bench_proto_proxy.cpp
new file mode 100644
index 0000000000..fa3191a969
--- /dev/null
+++ b/tests/benchmarks/components/api/bench_proto_proxy.cpp
@@ -0,0 +1,280 @@
+// Encode/decode microbenchmarks for proxy message families that carry
+// high-volume traffic (Z-Wave, IR/RF, serial). Mirrors the existing
+// BluetoothLERawAdvertisementsResponse benchmarks in bench_proto_encode.cpp.
+
+#include <benchmark/benchmark.h>
+
+#include <cstring>
+
+#include "esphome/components/api/api_pb2.h"
+#include "esphome/components/api/api_buffer.h"
+
+namespace esphome::api::benchmarks {
+
+static constexpr int kInnerIterations = 2000;
+
+// Encodes `src` into `out`. Caller owns `out` and must keep it alive across
+// the decode loop (decoded messages may store pointers back into its bytes).
+template<typename T> static void encode_into(APIBuffer &out, const T &src) {
+  out.resize(src.calculate_size());
+  ProtoWriteBuffer writer(&out, 0);
+  src.encode(writer);
+}
+
+// --- ZWaveProxyFrame (Z-Wave frame, ~16 bytes payload) ---
+
+#ifdef USE_ZWAVE_PROXY
+
+static const uint8_t kZWaveFrameData[] = {0x01, 0x09, 0x00, 0x13, 0x01, 0x02, 0x00, 0x00,
+                                          0x25, 0x00, 0x05, 0xC4, 0x00, 0x00, 0x00, 0x00};
+
+static void Encode_ZWaveProxyFrame(benchmark::State &state) {
+  ZWaveProxyFrame msg;
+  msg.data = kZWaveFrameData;
+  msg.data_len = sizeof(kZWaveFrameData);
+  APIBuffer buffer;
+  buffer.resize(msg.calculate_size());
+
+  for (auto _ : state) {
+    for (int i = 0; i < kInnerIterations; i++) {
+      ProtoWriteBuffer writer(&buffer, 0);
+      msg.encode(writer);
+    }
+    benchmark::DoNotOptimize(buffer.data());
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(Encode_ZWaveProxyFrame);
+
+static void Decode_ZWaveProxyFrame(benchmark::State &state) {
+  ZWaveProxyFrame source;
+  source.data = kZWaveFrameData;
+  source.data_len = sizeof(kZWaveFrameData);
+  APIBuffer encoded;
+  encode_into(encoded, source);
+  const uint8_t *data = encoded.data();
+  size_t size = encoded.size();
+
+  for (auto _ : state) {
+    for (int i = 0; i < kInnerIterations; i++) {
+      ZWaveProxyFrame msg;
+      msg.decode(data, size);
+      benchmark::DoNotOptimize(msg);
+    }
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(Decode_ZWaveProxyFrame);
+
+static const uint8_t kZWaveRequestData[] = {0xDE, 0xAD, 0xBE, 0xEF};
+
+static void Decode_ZWaveProxyRequest(benchmark::State &state) {
+  ZWaveProxyRequest source;
+  source.type = enums::ZWAVE_PROXY_REQUEST_TYPE_HOME_ID_CHANGE;
+  source.data = kZWaveRequestData;
+  source.data_len = sizeof(kZWaveRequestData);
+  APIBuffer encoded;
+  encode_into(encoded, source);
+  const uint8_t *data = encoded.data();
+  size_t size = encoded.size();
+
+  for (auto _ : state) {
+    for (int i = 0; i < kInnerIterations; i++) {
+      ZWaveProxyRequest msg;
+      msg.decode(data, size);
+      benchmark::DoNotOptimize(msg);
+    }
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(Decode_ZWaveProxyRequest);
+
+#endif  // USE_ZWAVE_PROXY
+
+// --- SerialProxyDataReceived encode + SerialProxyWriteRequest decode ---
+//
+// SerialProxyWriteRequest is decode-only (SOURCE_CLIENT) but has the same
+// wire layout as SerialProxyDataReceived, so we encode via the latter and
+// decode as the former.
+
+#ifdef USE_SERIAL_PROXY
+
+static constexpr size_t kSerialPayloadSize = 64;
+static const uint8_t kSerialPayload[kSerialPayloadSize] = {
+    0x55, 0xAA, 0x12, 0x34, 0x56, 0x78, 0x9A, 0xBC, 0xDE, 0xF0, 0x01, 0x23, 0x45, 0x67, 0x89, 0xAB,
+    0xCD, 0xEF, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77, 0x88, 0x99, 0xAA, 0xBB, 0xCC, 0xDD, 0xEE,
+    0xFF, 0x00, 0x10, 0x20, 0x30, 0x40, 0x50, 0x60, 0x70, 0x80, 0x90, 0xA0, 0xB0, 0xC0, 0xD0, 0xE0,
+    0xF0, 0x0F, 0x1F, 0x2F, 0x3F, 0x4F, 0x5F, 0x6F, 0x7F, 0x8F, 0x9F, 0xAF, 0xBF, 0xCF, 0xDF, 0xEF};
+
+static void Encode_SerialProxyDataReceived(benchmark::State &state) {
+  SerialProxyDataReceived msg;
+  msg.instance = 0;
+  msg.set_data(kSerialPayload, kSerialPayloadSize);
+  APIBuffer buffer;
+  buffer.resize(msg.calculate_size());
+
+  for (auto _ : state) {
+    for (int i = 0; i < kInnerIterations; i++) {
+      ProtoWriteBuffer writer(&buffer, 0);
+      msg.encode(writer);
+    }
+    benchmark::DoNotOptimize(buffer.data());
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(Encode_SerialProxyDataReceived);
+
+static void Decode_SerialProxyWriteRequest(benchmark::State &state) {
+  SerialProxyDataReceived source;
+  source.instance = 0;
+  source.set_data(kSerialPayload, kSerialPayloadSize);
+  APIBuffer encoded;
+  encode_into(encoded, source);
+  const uint8_t *data = encoded.data();
+  size_t size = encoded.size();
+
+  for (auto _ : state) {
+    for (int i = 0; i < kInnerIterations; i++) {
+      SerialProxyWriteRequest msg;
+      msg.decode(data, size);
+      benchmark::DoNotOptimize(msg);
+    }
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(Decode_SerialProxyWriteRequest);
+
+#endif  // USE_SERIAL_PROXY
+
+// --- InfraredRFReceiveEvent encode (100 sint32 timings) +
+//     InfraredRFTransmitRawTimingsRequest decode (hand-built wire bytes) ---
+
+#if defined(USE_IR_RF) || defined(USE_RADIO_FREQUENCY)
+
+// Mark/space pairs simulating a typical RC-5 / NEC capture (100 timings).
+static std::vector<int32_t> make_ir_timings_100() {
+  std::vector<int32_t> v;
+  v.reserve(100);
+  for (int i = 0; i < 100; i++) {
+    v.push_back((i % 2 == 0) ? 560 : -560);
+  }
+  return v;
+}
+
+static const std::vector<int32_t> &get_ir_timings_100() {
+  static const std::vector<int32_t> timings = make_ir_timings_100();
+  return timings;
+}
+
+static void Encode_InfraredRFReceiveEvent(benchmark::State &state) {
+  InfraredRFReceiveEvent msg;
+  msg.key = 0xDEADBEEF;
+  msg.timings = &get_ir_timings_100();
+  APIBuffer buffer;
+  buffer.resize(msg.calculate_size());
+
+  for (auto _ : state) {
+    for (int i = 0; i < kInnerIterations; i++) {
+      ProtoWriteBuffer writer(&buffer, 0);
+      msg.encode(writer);
+    }
+    benchmark::DoNotOptimize(buffer.data());
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(Encode_InfraredRFReceiveEvent);
+
+static void CalculateSize_InfraredRFReceiveEvent(benchmark::State &state) {
+  InfraredRFReceiveEvent msg;
+  msg.key = 0xDEADBEEF;
+  msg.timings = &get_ir_timings_100();
+
+  for (auto _ : state) {
+    uint32_t result = 0;
+    for (int i = 0; i < kInnerIterations; i++) {
+      result += msg.calculate_size();
+    }
+    benchmark::DoNotOptimize(result);
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(CalculateSize_InfraredRFReceiveEvent);
+
+// Hand-built wire bytes for InfraredRFTransmitRawTimingsRequest (decode-only,
+// no sister message with identical wire layout).
+//   field 2 (key, fixed32):           tag=0x15, 4 LE bytes
+//   field 3 (carrier_frequency):      tag=0x18, varint
+//   field 4 (repeat_count):           tag=0x20, varint
+//   field 5 (timings, packed sint32): tag=0x2A, length varint, packed payload
+//   field 6 (modulation):             tag=0x30, varint
+static APIBuffer build_infrared_rf_transmit_wire() {
+  uint8_t bytes[256];
+  size_t len = 0;
+
+  auto put_byte = [&](uint8_t b) { bytes[len++] = b; };
+  auto put_varint = [&](uint32_t v) {
+    while (v >= 0x80) {
+      bytes[len++] = static_cast<uint8_t>((v & 0x7F) | 0x80);
+      v >>= 7;
+    }
+    bytes[len++] = static_cast<uint8_t>(v);
+  };
+  auto encode_zigzag = [](int32_t v) -> uint32_t {
+    return (static_cast<uint32_t>(v) << 1) ^ static_cast<uint32_t>(v >> 31);
+  };
+
+  put_byte(0x15);
+  put_byte(0xEF);
+  put_byte(0xBE);
+  put_byte(0xAD);
+  put_byte(0xDE);
+  put_byte(0x18);
+  put_varint(38000);
+  put_byte(0x20);
+  put_varint(2);
+
+  uint8_t packed[200];
+  size_t packed_len = 0;
+  for (int i = 0; i < 100; i++) {
+    int32_t value = (i % 2 == 0) ? 560 : -560;
+    uint32_t zz = encode_zigzag(value);
+    while (zz >= 0x80) {
+      packed[packed_len++] = static_cast<uint8_t>((zz & 0x7F) | 0x80);
+      zz >>= 7;
+    }
+    packed[packed_len++] = static_cast<uint8_t>(zz);
+  }
+  put_byte(0x2A);
+  put_varint(static_cast<uint32_t>(packed_len));
+  std::memcpy(bytes + len, packed, packed_len);
+  len += packed_len;
+  // field 6: modulation = 1 (non-zero so it's actually emitted and exercises
+  // decode_varint for this field, matching the documented layout above).
+  put_byte(0x30);
+  put_varint(1);
+
+  APIBuffer buf;
+  buf.resize(len);
+  std::memcpy(buf.data(), bytes, len);
+  return buf;
+}
+
+static void Decode_InfraredRFTransmitRawTimingsRequest(benchmark::State &state) {
+  auto encoded = build_infrared_rf_transmit_wire();
+  const uint8_t *data = encoded.data();
+  size_t size = encoded.size();
+
+  for (auto _ : state) {
+    for (int i = 0; i < kInnerIterations; i++) {
+      InfraredRFTransmitRawTimingsRequest msg;
+      msg.decode(data, size);
+      benchmark::DoNotOptimize(msg);
+    }
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+BENCHMARK(Decode_InfraredRFTransmitRawTimingsRequest);
+
+#endif  // USE_IR_RF || USE_RADIO_FREQUENCY
+
+}  // namespace esphome::api::benchmarks
diff --git a/tests/benchmarks/stubs/esphome/components/infrared/infrared.h b/tests/benchmarks/stubs/esphome/components/infrared/infrared.h
new file mode 100644
index 0000000000..874e7a270b
--- /dev/null
+++ b/tests/benchmarks/stubs/esphome/components/infrared/infrared.h
@@ -0,0 +1,45 @@
+// Stub for benchmark builds — provides the minimal interface that
+// api_connection.cpp and Application need when USE_INFRARED is defined,
+// without pulling in the real remote_base/RMT dependencies.
+#pragma once
+
+#include <cstdint>
+#include "esphome/core/component.h"
+#include "esphome/core/entity_base.h"
+
+namespace esphome::infrared {
+
+class Infrared;
+
+class InfraredCall {
+ public:
+  explicit InfraredCall(Infrared *parent) : parent_(parent) {}
+  InfraredCall &set_carrier_frequency(uint32_t /*frequency*/) { return *this; }
+  InfraredCall &set_raw_timings_packed(const uint8_t * /*data*/, uint16_t /*length*/, uint16_t /*count*/) {
+    return *this;
+  }
+  InfraredCall &set_repeat_count(uint32_t /*count*/) { return *this; }
+  void perform() {}
+
+ protected:
+  Infrared *parent_;
+};
+
+class InfraredTraits {
+ public:
+  uint32_t get_receiver_frequency_hz() const { return 0; }
+};
+
+class Infrared : public Component, public EntityBase {
+ public:
+  Infrared() = default;
+  InfraredTraits &get_traits() { return this->traits_; }
+  const InfraredTraits &get_traits() const { return this->traits_; }
+  InfraredCall make_call() { return InfraredCall(this); }
+  uint32_t get_capability_flags() const { return 0; }
+
+ protected:
+  InfraredTraits traits_;
+};
+
+}  // namespace esphome::infrared
diff --git a/tests/benchmarks/stubs/esphome/components/radio_frequency/radio_frequency.h b/tests/benchmarks/stubs/esphome/components/radio_frequency/radio_frequency.h
new file mode 100644
index 0000000000..72fc08034b
--- /dev/null
+++ b/tests/benchmarks/stubs/esphome/components/radio_frequency/radio_frequency.h
@@ -0,0 +1,51 @@
+// Stub for benchmark builds — provides the minimal interface that
+// api_connection.cpp and Application need when USE_RADIO_FREQUENCY is defined.
+#pragma once
+
+#include <cstdint>
+#include "esphome/core/component.h"
+#include "esphome/core/entity_base.h"
+
+namespace esphome::radio_frequency {
+
+enum RadioFrequencyModulation : uint32_t {
+  RADIO_FREQUENCY_MODULATION_OOK = 0,
+};
+
+class RadioFrequency;
+
+class RadioFrequencyCall {
+ public:
+  explicit RadioFrequencyCall(RadioFrequency *parent) : parent_(parent) {}
+  RadioFrequencyCall &set_frequency(uint32_t /*frequency*/) { return *this; }
+  RadioFrequencyCall &set_modulation(RadioFrequencyModulation /*mod*/) { return *this; }
+  RadioFrequencyCall &set_repeat_count(uint32_t /*count*/) { return *this; }
+  RadioFrequencyCall &set_raw_timings_packed(const uint8_t * /*data*/, uint16_t /*length*/, uint16_t /*count*/) {
+    return *this;
+  }
+  void perform() {}
+
+ protected:
+  RadioFrequency *parent_;
+};
+
+class RadioFrequencyTraits {
+ public:
+  uint32_t get_frequency_min_hz() const { return 0; }
+  uint32_t get_frequency_max_hz() const { return 0; }
+  uint32_t get_supported_modulations() const { return 0; }
+};
+
+class RadioFrequency : public Component, public EntityBase {
+ public:
+  RadioFrequency() = default;
+  RadioFrequencyTraits &get_traits() { return this->traits_; }
+  const RadioFrequencyTraits &get_traits() const { return this->traits_; }
+  RadioFrequencyCall make_call() { return RadioFrequencyCall(this); }
+  uint32_t get_capability_flags() const { return 0; }
+
+ protected:
+  RadioFrequencyTraits traits_;
+};
+
+}  // namespace esphome::radio_frequency
diff --git a/tests/benchmarks/stubs/esphome/components/serial_proxy/serial_proxy.h b/tests/benchmarks/stubs/esphome/components/serial_proxy/serial_proxy.h
new file mode 100644
index 0000000000..bab27549e7
--- /dev/null
+++ b/tests/benchmarks/stubs/esphome/components/serial_proxy/serial_proxy.h
@@ -0,0 +1,46 @@
+// Stub for benchmark builds — provides the minimal interface that
+// api_connection.cpp and Application need when USE_SERIAL_PROXY is defined,
+// without pulling in the real UART implementation.
+#pragma once
+
+#include <cstdint>
+#include <cstddef>
+#include "esphome/components/api/api_pb2.h"
+
+namespace esphome {
+
+namespace api {
+class APIConnection;
+}  // namespace api
+
+namespace uart {
+enum class UARTFlushResult : uint8_t {
+  UART_FLUSH_RESULT_SUCCESS,
+  UART_FLUSH_RESULT_ASSUMED_SUCCESS,
+  UART_FLUSH_RESULT_TIMEOUT,
+  UART_FLUSH_RESULT_FAILED,
+};
+}  // namespace uart
+
+namespace serial_proxy {
+
+class SerialProxy {
+ public:
+  void set_instance_index(uint32_t index) { this->instance_index_ = index; }
+  uint32_t get_instance_index() const { return this->instance_index_; }
+  const char *get_name() const { return ""; }
+  api::enums::SerialProxyPortType get_port_type() const { return {}; }
+  api::APIConnection *get_api_connection() { return nullptr; }
+  void serial_proxy_request(api::APIConnection *conn, api::enums::SerialProxyRequestType type) {}
+  void configure(uint32_t baudrate, bool flow_control, uint8_t parity, uint32_t stop_bits, uint32_t data_size) {}
+  void write_from_client(const uint8_t *data, size_t len) {}
+  void set_modem_pins(uint32_t line_states) {}
+  uint32_t get_modem_pins() const { return 0; }
+  uart::UARTFlushResult flush_port() { return uart::UARTFlushResult::UART_FLUSH_RESULT_SUCCESS; }
+
+ protected:
+  uint32_t instance_index_{0};
+};
+
+}  // namespace serial_proxy
+}  // namespace esphome
diff --git a/tests/benchmarks/stubs/esphome/components/zwave_proxy/zwave_proxy.h b/tests/benchmarks/stubs/esphome/components/zwave_proxy/zwave_proxy.h
new file mode 100644
index 0000000000..ba97e81236
--- /dev/null
+++ b/tests/benchmarks/stubs/esphome/components/zwave_proxy/zwave_proxy.h
@@ -0,0 +1,29 @@
+// Stub for benchmark builds — provides the minimal interface that
+// api_connection.cpp needs when USE_ZWAVE_PROXY is defined,
+// without pulling in the real UART-based ZWaveProxy implementation.
+#pragma once
+
+#include "esphome/components/api/api_pb2.h"
+
+namespace esphome {
+namespace api {
+class APIConnection;
+}  // namespace api
+
+namespace zwave_proxy {
+
+class ZWaveProxy {
+ public:
+  api::APIConnection *get_api_connection() { return nullptr; }
+  void zwave_proxy_request(api::APIConnection *conn, api::enums::ZWaveProxyRequestType type) {}
+  void send_frame(const uint8_t *data, size_t length) {}
+  void api_connection_authenticated(api::APIConnection *conn) {}
+  uint32_t get_feature_flags() const { return 0; }
+  uint32_t get_home_id() { return 0; }
+};
+
+// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
+extern ZWaveProxy *global_zwave_proxy;
+
+}  // namespace zwave_proxy
+}  // namespace esphome

From b708d1a8260b55acd9c0c7c66af3fbadc6b4c564 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:14:34 -0500
Subject: [PATCH 10/19] [core] Drop unused DELAY_ACTION from
 InternalSchedulerID enum (#16151)

---
 esphome/core/component.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/esphome/core/component.h b/esphome/core/component.h
index 185d51ab37..5baf795ca6 100644
--- a/esphome/core/component.h
+++ b/esphome/core/component.h
@@ -65,7 +65,6 @@ inline constexpr uint32_t SCHEDULER_DONT_RUN = 4294967295UL;
 /// with component-level NUMERIC_ID values, even if the uint32_t values overlap.
 enum class InternalSchedulerID : uint32_t {
   POLLING_UPDATE = 0,  // PollingComponent interval
-  DELAY_ACTION = 1,    // DelayAction timeout
 };
 
 // Forward declaration

From ba7c06785a03f1f7c570b572e5aa677d0ed9d073 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:14:55 -0500
Subject: [PATCH 11/19] [mdns] Broadcast config_hash TXT record on
 _esphomelib._tcp (#16145)

---
 esphome/components/mdns/mdns_component.cpp | 40 ++++++++++++++++++++--
 esphome/components/mdns/mdns_component.h   | 33 +++++-------------
 esphome/components/mdns/mdns_host.cpp      |  7 +++-
 3 files changed, 52 insertions(+), 28 deletions(-)

diff --git a/esphome/components/mdns/mdns_component.cpp b/esphome/components/mdns/mdns_component.cpp
index e05373ac5d..9bf27e71e4 100644
--- a/esphome/components/mdns/mdns_component.cpp
+++ b/esphome/components/mdns/mdns_component.cpp
@@ -39,7 +39,39 @@ MDNS_STATIC_CONST_CHAR(SERVICE_TCP, "_tcp");
 // Wrap build-time defines into flash storage
 MDNS_STATIC_CONST_CHAR(VALUE_VERSION, ESPHOME_VERSION);
 
-void MDNSComponent::compile_records_(StaticVector<MDNSService, MDNS_SERVICE_COUNT> &services, char *mac_address_buf) {
+void MDNSComponent::setup_buffers_and_register_(PlatformRegisterFn platform_register) {
+#ifdef USE_MDNS_STORE_SERVICES
+  auto &services = this->services_;
+#else
+  StaticVector<MDNSService, MDNS_SERVICE_COUNT> services_storage;
+  auto &services = services_storage;
+#endif
+
+#ifdef USE_API
+#ifdef USE_MDNS_STORE_SERVICES
+  get_mac_address_into_buffer(this->mac_address_);
+  char *mac_ptr = this->mac_address_;
+  format_hex_to(this->config_hash_str_, App.get_config_hash());
+  char *cfg_ptr = this->config_hash_str_;
+#else
+  char mac_address[MAC_ADDRESS_BUFFER_SIZE];
+  char config_hash_str[CONFIG_HASH_STR_SIZE];
+  get_mac_address_into_buffer(mac_address);
+  format_hex_to(config_hash_str, App.get_config_hash());
+  char *mac_ptr = mac_address;
+  char *cfg_ptr = config_hash_str;
+#endif
+#else
+  char *mac_ptr = nullptr;
+  char *cfg_ptr = nullptr;
+#endif
+
+  this->compile_records_(services, mac_ptr, cfg_ptr);
+  platform_register(this, services);
+}
+
+void MDNSComponent::compile_records_(StaticVector<MDNSService, MDNS_SERVICE_COUNT> &services, char *mac_address_buf,
+                                     char *config_hash_buf) {
   // IMPORTANT: The #ifdef blocks below must match COMPONENTS_WITH_MDNS_SERVICES
   // in mdns/__init__.py. If you add a new service here, update both locations.
 
@@ -47,6 +79,7 @@ void MDNSComponent::compile_records_(StaticVector<MDNSService, MDNS_SERVICE_COUN
   MDNS_STATIC_CONST_CHAR(SERVICE_ESPHOMELIB, "_esphomelib");
   MDNS_STATIC_CONST_CHAR(TXT_FRIENDLY_NAME, "friendly_name");
   MDNS_STATIC_CONST_CHAR(TXT_VERSION, "version");
+  MDNS_STATIC_CONST_CHAR(TXT_CONFIG_HASH, "config_hash");
   MDNS_STATIC_CONST_CHAR(TXT_MAC, "mac");
   MDNS_STATIC_CONST_CHAR(TXT_PLATFORM, "platform");
   MDNS_STATIC_CONST_CHAR(TXT_BOARD, "board");
@@ -63,7 +96,7 @@ void MDNSComponent::compile_records_(StaticVector<MDNSService, MDNS_SERVICE_COUN
     bool friendly_name_empty = friendly_name.empty();
 
     // Calculate exact capacity for txt_records
-    size_t txt_count = 3;  // version, mac, board (always present)
+    size_t txt_count = 4;  // version, config_hash, mac, board (always present)
     if (!friendly_name_empty) {
       txt_count++;  // friendly_name
     }
@@ -91,6 +124,9 @@ void MDNSComponent::compile_records_(StaticVector<MDNSService, MDNS_SERVICE_COUN
     }
     txt_records.push_back({MDNS_STR(TXT_VERSION), MDNS_STR(VALUE_VERSION)});
 
+    // Config hash: passed from caller (either member buffer or stack buffer depending on USE_MDNS_STORE_SERVICES)
+    txt_records.push_back({MDNS_STR(TXT_CONFIG_HASH), MDNS_STR(config_hash_buf)});
+
     // MAC address: passed from caller (either member buffer or stack buffer depending on USE_MDNS_STORE_SERVICES)
     txt_records.push_back({MDNS_STR(TXT_MAC), MDNS_STR(mac_address_buf)});
 
diff --git a/esphome/components/mdns/mdns_component.h b/esphome/components/mdns/mdns_component.h
index 798af0e0bf..e981a65941 100644
--- a/esphome/components/mdns/mdns_component.h
+++ b/esphome/components/mdns/mdns_component.h
@@ -70,6 +70,9 @@ class MDNSComponent final : public Component
   void setup() override;
   void dump_config() override;
 
+  /// Size of buffer required for config hash hex string (8 hex chars + null terminator)
+  static constexpr size_t CONFIG_HASH_STR_SIZE = format_hex_size(sizeof(uint32_t));
+
 #ifdef USE_MDNS_EVENT_DRIVEN_POLLING
   // LEAmDNS has meaningful work only during the probe+announce phase (3×250ms probes +
   // 8×1000ms announces, ~9s). Afterwards every internal timer is resetToNeverExpires()
@@ -124,30 +127,7 @@ class MDNSComponent final : public Component
   /// Helper to set up services and MAC buffers, then call platform-specific registration
   using PlatformRegisterFn = void (*)(MDNSComponent *, StaticVector<MDNSService, MDNS_SERVICE_COUNT> &);
 
-  void setup_buffers_and_register_(PlatformRegisterFn platform_register) {
-#ifdef USE_MDNS_STORE_SERVICES
-    auto &services = this->services_;
-#else
-    StaticVector<MDNSService, MDNS_SERVICE_COUNT> services_storage;
-    auto &services = services_storage;
-#endif
-
-#ifdef USE_API
-#ifdef USE_MDNS_STORE_SERVICES
-    get_mac_address_into_buffer(this->mac_address_);
-    char *mac_ptr = this->mac_address_;
-#else
-    char mac_address[MAC_ADDRESS_BUFFER_SIZE];
-    get_mac_address_into_buffer(mac_address);
-    char *mac_ptr = mac_address;
-#endif
-#else
-    char *mac_ptr = nullptr;
-#endif
-
-    this->compile_records_(services, mac_ptr);
-    platform_register(this, services);
-  }
+  void setup_buffers_and_register_(PlatformRegisterFn platform_register);
 
 #ifdef USE_MDNS_DYNAMIC_TXT
   /// Storage for runtime-generated TXT values from user lambdas
@@ -159,6 +139,8 @@ class MDNSComponent final : public Component
 #if defined(USE_API) && defined(USE_MDNS_STORE_SERVICES)
   /// Fixed buffer for MAC address (only needed when services are stored)
   char mac_address_[MAC_ADDRESS_BUFFER_SIZE];
+  /// Fixed buffer for config hash hex string (only needed when services are stored)
+  char config_hash_str_[CONFIG_HASH_STR_SIZE];
 #endif
 #ifdef USE_MDNS_STORE_SERVICES
   StaticVector<MDNSService, MDNS_SERVICE_COUNT> services_{};
@@ -167,7 +149,8 @@ class MDNSComponent final : public Component
   // RP2040 defers MDNS.begin() until the first IP-up event; this tracks that.
   bool initialized_{false};
 #endif
-  void compile_records_(StaticVector<MDNSService, MDNS_SERVICE_COUNT> &services, char *mac_address_buf);
+  void compile_records_(StaticVector<MDNSService, MDNS_SERVICE_COUNT> &services, char *mac_address_buf,
+                        char *config_hash_buf);
 };
 
 }  // namespace esphome::mdns
diff --git a/esphome/components/mdns/mdns_host.cpp b/esphome/components/mdns/mdns_host.cpp
index 4d902319b8..1e66a10df0 100644
--- a/esphome/components/mdns/mdns_host.cpp
+++ b/esphome/components/mdns/mdns_host.cpp
@@ -3,6 +3,8 @@
 
 #include "esphome/components/network/ip_address.h"
 #include "esphome/components/network/util.h"
+#include "esphome/core/application.h"
+#include "esphome/core/helpers.h"
 #include "esphome/core/log.h"
 #include "mdns_component.h"
 
@@ -13,10 +15,13 @@ void MDNSComponent::setup() {
 #ifdef USE_API
   get_mac_address_into_buffer(this->mac_address_);
   char *mac_ptr = this->mac_address_;
+  format_hex_to(this->config_hash_str_, App.get_config_hash());
+  char *cfg_ptr = this->config_hash_str_;
 #else
   char *mac_ptr = nullptr;
+  char *cfg_ptr = nullptr;
 #endif
-  this->compile_records_(this->services_, mac_ptr);
+  this->compile_records_(this->services_, mac_ptr, cfg_ptr);
 #endif
   // Host platform doesn't have actual mDNS implementation
 }

From 550444dc34da939cd3c1da77042d7d07c88fe766 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:15:18 -0500
Subject: [PATCH 12/19] [binary_sensor] Drop Component from filter classes, use
 self-keyed scheduler (#16131)

---
 esphome/components/binary_sensor/__init__.py | 15 +++------
 esphome/components/binary_sensor/filter.cpp  | 34 +++++++-------------
 esphome/components/binary_sensor/filter.h    | 18 +++--------
 3 files changed, 22 insertions(+), 45 deletions(-)

diff --git a/esphome/components/binary_sensor/__init__.py b/esphome/components/binary_sensor/__init__.py
index 1456e5bc66..db82290750 100644
--- a/esphome/components/binary_sensor/__init__.py
+++ b/esphome/components/binary_sensor/__init__.py
@@ -143,15 +143,15 @@ BinarySensorCondition = binary_sensor_ns.class_("BinarySensorCondition", Conditi
 
 # Filters
 Filter = binary_sensor_ns.class_("Filter")
-TimeoutFilter = binary_sensor_ns.class_("TimeoutFilter", Filter, cg.Component)
-DelayedOnOffFilter = binary_sensor_ns.class_("DelayedOnOffFilter", Filter, cg.Component)
-DelayedOnFilter = binary_sensor_ns.class_("DelayedOnFilter", Filter, cg.Component)
-DelayedOffFilter = binary_sensor_ns.class_("DelayedOffFilter", Filter, cg.Component)
+TimeoutFilter = binary_sensor_ns.class_("TimeoutFilter", Filter)
+DelayedOnOffFilter = binary_sensor_ns.class_("DelayedOnOffFilter", Filter)
+DelayedOnFilter = binary_sensor_ns.class_("DelayedOnFilter", Filter)
+DelayedOffFilter = binary_sensor_ns.class_("DelayedOffFilter", Filter)
 InvertFilter = binary_sensor_ns.class_("InvertFilter", Filter)
 AutorepeatFilter = binary_sensor_ns.class_("AutorepeatFilter", Filter, cg.Component)
 LambdaFilter = binary_sensor_ns.class_("LambdaFilter", Filter)
 StatelessLambdaFilter = binary_sensor_ns.class_("StatelessLambdaFilter", Filter)
-SettleFilter = binary_sensor_ns.class_("SettleFilter", Filter, cg.Component)
+SettleFilter = binary_sensor_ns.class_("SettleFilter", Filter)
 
 _LOGGER = getLogger(__name__)
 
@@ -175,7 +175,6 @@ async def invert_filter_to_code(config, filter_id):
 )
 async def timeout_filter_to_code(config, filter_id):
     var = cg.new_Pvariable(filter_id)
-    await cg.register_component(var, {})
     template_ = await cg.templatable(config, [], cg.uint32)
     cg.add(var.set_timeout_value(template_))
     return var
@@ -203,7 +202,6 @@ async def timeout_filter_to_code(config, filter_id):
 )
 async def delayed_on_off_filter_to_code(config, filter_id):
     var = cg.new_Pvariable(filter_id)
-    await cg.register_component(var, {})
     if isinstance(config, dict):
         template_ = await cg.templatable(config[CONF_TIME_ON], [], cg.uint32)
         cg.add(var.set_on_delay(template_))
@@ -221,7 +219,6 @@ async def delayed_on_off_filter_to_code(config, filter_id):
 )
 async def delayed_on_filter_to_code(config, filter_id):
     var = cg.new_Pvariable(filter_id)
-    await cg.register_component(var, {})
     template_ = await cg.templatable(config, [], cg.uint32)
     cg.add(var.set_delay(template_))
     return var
@@ -234,7 +231,6 @@ async def delayed_on_filter_to_code(config, filter_id):
 )
 async def delayed_off_filter_to_code(config, filter_id):
     var = cg.new_Pvariable(filter_id)
-    await cg.register_component(var, {})
     template_ = await cg.templatable(config, [], cg.uint32)
     cg.add(var.set_delay(template_))
     return var
@@ -306,7 +302,6 @@ async def lambda_filter_to_code(config, filter_id):
 )
 async def settle_filter_to_code(config, filter_id):
     var = cg.new_Pvariable(filter_id)
-    await cg.register_component(var, {})
     template_ = await cg.templatable(config, [], cg.uint32)
     cg.add(var.set_delay(template_))
     return var
diff --git a/esphome/components/binary_sensor/filter.cpp b/esphome/components/binary_sensor/filter.cpp
index 914060ce13..0a463ee9a9 100644
--- a/esphome/components/binary_sensor/filter.cpp
+++ b/esphome/components/binary_sensor/filter.cpp
@@ -4,16 +4,14 @@
 #include "filter.h"
 
 #include "binary_sensor.h"
+#include "esphome/core/application.h"
 
 namespace esphome::binary_sensor {
 
 static const char *const TAG = "sensor.filter";
 
-// Timeout IDs for filter classes.
-// Each filter is its own Component instance, so the scheduler scopes
-// IDs by component pointer — no risk of collisions between instances.
-constexpr uint32_t FILTER_TIMEOUT_ID = 0;
-// AutorepeatFilter needs two distinct IDs (both timeouts on the same component)
+// AutorepeatFilter still inherits Component (it schedules two distinct timer
+// purposes), so it keeps the (Component *, id) scheduler API.
 constexpr uint32_t AUTOREPEAT_TIMING_ID = 0;
 constexpr uint32_t AUTOREPEAT_ON_OFF_ID = 1;
 
@@ -34,46 +32,40 @@ void Filter::input(bool value) {
 }
 
 void TimeoutFilter::input(bool value) {
-  this->set_timeout(FILTER_TIMEOUT_ID, this->timeout_delay_.value(), [this]() { this->parent_->invalidate_state(); });
+  App.scheduler.set_timeout(this, this->timeout_delay_.value(), [this]() { this->parent_->invalidate_state(); });
   // we do not de-dup here otherwise changes from invalid to valid state will not be output
   this->output(value);
 }
 
 optional<bool> DelayedOnOffFilter::new_value(bool value) {
   if (value) {
-    this->set_timeout(FILTER_TIMEOUT_ID, this->on_delay_.value(), [this]() { this->output(true); });
+    App.scheduler.set_timeout(this, this->on_delay_.value(), [this]() { this->output(true); });
   } else {
-    this->set_timeout(FILTER_TIMEOUT_ID, this->off_delay_.value(), [this]() { this->output(false); });
+    App.scheduler.set_timeout(this, this->off_delay_.value(), [this]() { this->output(false); });
   }
   return {};
 }
 
-float DelayedOnOffFilter::get_setup_priority() const { return setup_priority::HARDWARE; }
-
 optional<bool> DelayedOnFilter::new_value(bool value) {
   if (value) {
-    this->set_timeout(FILTER_TIMEOUT_ID, this->delay_.value(), [this]() { this->output(true); });
+    App.scheduler.set_timeout(this, this->delay_.value(), [this]() { this->output(true); });
     return {};
   } else {
-    this->cancel_timeout(FILTER_TIMEOUT_ID);
+    App.scheduler.cancel_timeout(this);
     return false;
   }
 }
 
-float DelayedOnFilter::get_setup_priority() const { return setup_priority::HARDWARE; }
-
 optional<bool> DelayedOffFilter::new_value(bool value) {
   if (!value) {
-    this->set_timeout(FILTER_TIMEOUT_ID, this->delay_.value(), [this]() { this->output(false); });
+    App.scheduler.set_timeout(this, this->delay_.value(), [this]() { this->output(false); });
     return {};
   } else {
-    this->cancel_timeout(FILTER_TIMEOUT_ID);
+    App.scheduler.cancel_timeout(this);
     return true;
   }
 }
 
-float DelayedOffFilter::get_setup_priority() const { return setup_priority::HARDWARE; }
-
 optional<bool> InvertFilter::new_value(bool value) { return !value; }
 
 // AutorepeatFilterBase
@@ -118,20 +110,18 @@ optional<bool> LambdaFilter::new_value(bool value) { return this->f_(value); }
 
 optional<bool> SettleFilter::new_value(bool value) {
   if (!this->steady_) {
-    this->set_timeout(FILTER_TIMEOUT_ID, this->delay_.value(), [this, value]() {
+    App.scheduler.set_timeout(this, this->delay_.value(), [this, value]() {
       this->steady_ = true;
       this->output(value);
     });
     return {};
   } else {
     this->steady_ = false;
-    this->set_timeout(FILTER_TIMEOUT_ID, this->delay_.value(), [this]() { this->steady_ = true; });
+    App.scheduler.set_timeout(this, this->delay_.value(), [this]() { this->steady_ = true; });
     return value;
   }
 }
 
-float SettleFilter::get_setup_priority() const { return setup_priority::HARDWARE; }
-
 }  // namespace esphome::binary_sensor
 
 #endif  // USE_BINARY_SENSOR_FILTER
diff --git a/esphome/components/binary_sensor/filter.h b/esphome/components/binary_sensor/filter.h
index 2e45554f81..8ff57cab0c 100644
--- a/esphome/components/binary_sensor/filter.h
+++ b/esphome/components/binary_sensor/filter.h
@@ -29,7 +29,7 @@ class Filter {
   Deduplicator<bool> dedup_;
 };
 
-class TimeoutFilter : public Filter, public Component {
+class TimeoutFilter : public Filter {
  public:
   optional<bool> new_value(bool value) override { return value; }
   void input(bool value) override;
@@ -39,12 +39,10 @@ class TimeoutFilter : public Filter, public Component {
   TemplatableFn<uint32_t> timeout_delay_{};
 };
 
-class DelayedOnOffFilter final : public Filter, public Component {
+class DelayedOnOffFilter final : public Filter {
  public:
   optional<bool> new_value(bool value) override;
 
-  float get_setup_priority() const override;
-
   template<typename T> void set_on_delay(T delay) { this->on_delay_ = delay; }
   template<typename T> void set_off_delay(T delay) { this->off_delay_ = delay; }
 
@@ -53,24 +51,20 @@ class DelayedOnOffFilter final : public Filter, public Component {
   TemplatableFn<uint32_t> off_delay_{};
 };
 
-class DelayedOnFilter : public Filter, public Component {
+class DelayedOnFilter : public Filter {
  public:
   optional<bool> new_value(bool value) override;
 
-  float get_setup_priority() const override;
-
   template<typename T> void set_delay(T delay) { this->delay_ = delay; }
 
  protected:
   TemplatableFn<uint32_t> delay_{};
 };
 
-class DelayedOffFilter : public Filter, public Component {
+class DelayedOffFilter : public Filter {
  public:
   optional<bool> new_value(bool value) override;
 
-  float get_setup_priority() const override;
-
   template<typename T> void set_delay(T delay) { this->delay_ = delay; }
 
  protected:
@@ -146,12 +140,10 @@ class StatelessLambdaFilter : public Filter {
   optional<bool> (*f_)(bool);
 };
 
-class SettleFilter : public Filter, public Component {
+class SettleFilter : public Filter {
  public:
   optional<bool> new_value(bool value) override;
 
-  float get_setup_priority() const override;
-
   template<typename T> void set_delay(T delay) { this->delay_ = delay; }
 
  protected:

From 24fdfcf1a18be08c720a674384cd15712168d7d5 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:15:41 -0500
Subject: [PATCH 13/19] [rp2040] Move HAL bodies into components/rp2040/hal.cpp
 + inline trivial dispatches (#16114)

---
 esphome/components/rp2040/core.cpp | 39 ++--------------------------
 esphome/components/rp2040/hal.cpp  | 41 ++++++++++++++++++++++++++++++
 esphome/core/hal/hal_rp2040.h      | 19 +++++++++++---
 3 files changed, 59 insertions(+), 40 deletions(-)
 create mode 100644 esphome/components/rp2040/hal.cpp

diff --git a/esphome/components/rp2040/core.cpp b/esphome/components/rp2040/core.cpp
index d3dc1cf2bb..11f23ccfef 100644
--- a/esphome/components/rp2040/core.cpp
+++ b/esphome/components/rp2040/core.cpp
@@ -1,41 +1,6 @@
 #ifdef USE_RP2040
 
-#include "core.h"
-#include "esphome/core/defines.h"
-#ifdef USE_RP2040_CRASH_HANDLER
-#include "crash_handler.h"
-#endif
-#include "esphome/core/hal.h"
-#include "esphome/core/helpers.h"
-
-#include "hardware/timer.h"
-#include "hardware/watchdog.h"
-
-namespace esphome {
-
-// yield(), delay(), micros(), millis(), millis_64() inlined in hal.h.
-void HOT delayMicroseconds(uint32_t us) { delay_microseconds_safe(us); }
-void arch_restart() {
-  watchdog_reboot(0, 0, 10);
-  while (1) {
-    continue;
-  }
-}
-
-void arch_init() {
-#ifdef USE_RP2040_CRASH_HANDLER
-  rp2040::crash_handler_read_and_clear();
-#endif
-#if USE_RP2040_WATCHDOG_TIMEOUT > 0
-  watchdog_enable(USE_RP2040_WATCHDOG_TIMEOUT, false);
-#endif
-}
-
-void HOT arch_feed_wdt() { watchdog_update(); }
-
-uint32_t HOT arch_get_cpu_cycle_count() { return ulMainGetRunTimeCounterValue(); }
-uint32_t arch_get_cpu_freq_hz() { return RP2040::f_cpu(); }
-
-}  // namespace esphome
+// HAL functions live in hal.cpp. core.cpp is intentionally empty for
+// rp2040 — there is no extra component bootstrap to keep here.
 
 #endif  // USE_RP2040
diff --git a/esphome/components/rp2040/hal.cpp b/esphome/components/rp2040/hal.cpp
new file mode 100644
index 0000000000..7475205d60
--- /dev/null
+++ b/esphome/components/rp2040/hal.cpp
@@ -0,0 +1,41 @@
+#ifdef USE_RP2040
+
+#include "core.h"
+#include "esphome/core/defines.h"
+#include "esphome/core/hal.h"
+#ifdef USE_RP2040_CRASH_HANDLER
+#include "crash_handler.h"
+#endif
+
+#include "hardware/watchdog.h"
+
+// Empty rp2040 namespace block to satisfy ci-custom's lint_namespace check.
+// HAL functions live in namespace esphome (root) — they are not part of the
+// rp2040 component's API.
+namespace esphome::rp2040 {}  // namespace esphome::rp2040
+
+namespace esphome {
+
+// yield(), delay(), micros(), millis(), millis_64(), delayMicroseconds(),
+// arch_feed_wdt(), arch_get_cpu_cycle_count() inlined in core/hal/hal_rp2040.h.
+void arch_restart() {
+  watchdog_reboot(0, 0, 10);
+  while (1) {
+    continue;
+  }
+}
+
+void arch_init() {
+#ifdef USE_RP2040_CRASH_HANDLER
+  rp2040::crash_handler_read_and_clear();
+#endif
+#if USE_RP2040_WATCHDOG_TIMEOUT > 0
+  watchdog_enable(USE_RP2040_WATCHDOG_TIMEOUT, false);
+#endif
+}
+
+uint32_t arch_get_cpu_freq_hz() { return RP2040::f_cpu(); }
+
+}  // namespace esphome
+
+#endif  // USE_RP2040
diff --git a/esphome/core/hal/hal_rp2040.h b/esphome/core/hal/hal_rp2040.h
index 46f6e421cd..27a9b23c0b 100644
--- a/esphome/core/hal/hal_rp2040.h
+++ b/esphome/core/hal/hal_rp2040.h
@@ -20,8 +20,17 @@ extern "C" unsigned long millis(void);
 // Forward decl from <pico/time.h>.
 extern "C" uint64_t time_us_64(void);
 
+// Forward decls from pico-sdk / FreeRTOS port for the inline arch_*
+// wrappers below.
+extern "C" void watchdog_update(void);
+extern "C" unsigned long ulMainGetRunTimeCounterValue(void);
+
 namespace esphome {
 
+// Forward decl from helpers.h.
+// NOLINTNEXTLINE(readability-redundant-declaration)
+void delay_microseconds_safe(uint32_t us);
+
 /// Returns true when executing inside an interrupt handler.
 __attribute__((always_inline)) inline bool in_isr_context() {
   uint32_t ipsr;
@@ -35,9 +44,13 @@ __attribute__((always_inline)) inline uint32_t micros() { return static_cast<uin
 __attribute__((always_inline)) inline uint32_t millis() { return micros_to_millis(::time_us_64()); }
 __attribute__((always_inline)) inline uint64_t millis_64() { return micros_to_millis<uint64_t>(::time_us_64()); }
 
-void delayMicroseconds(uint32_t us);  // NOLINT(readability-identifier-naming)
-void arch_feed_wdt();
-uint32_t arch_get_cpu_cycle_count();
+// NOLINTNEXTLINE(readability-identifier-naming)
+__attribute__((always_inline)) inline void delayMicroseconds(uint32_t us) { delay_microseconds_safe(us); }
+__attribute__((always_inline)) inline void arch_feed_wdt() { watchdog_update(); }
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_cycle_count() {
+  return static_cast<uint32_t>(ulMainGetRunTimeCounterValue());
+}
+
 void arch_init();
 uint32_t arch_get_cpu_freq_hz();
 

From 3d69169141984487a28d5b1e323a3740748a1408 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:16:16 -0500
Subject: [PATCH 14/19] [climate] Fold ControlAction fields into a single
 stateless lambda (#16044)

---
 esphome/components/climate/__init__.py        | 87 +++++++++++-------
 esphome/components/climate/automation.h       | 35 ++-----
 .../fixtures/climate_control_action.yaml      | 92 +++++++++++++++++++
 .../test_climate_control_action.py            | 84 +++++++++++++++++
 4 files changed, 238 insertions(+), 60 deletions(-)
 create mode 100644 tests/integration/fixtures/climate_control_action.yaml
 create mode 100644 tests/integration/test_climate_control_action.py

diff --git a/esphome/components/climate/__init__.py b/esphome/components/climate/__init__.py
index 0fdb18a92c..7c9002d6dc 100644
--- a/esphome/components/climate/__init__.py
+++ b/esphome/components/climate/__init__.py
@@ -48,13 +48,13 @@ from esphome.const import (
     CONF_VISUAL,
     CONF_WEB_SERVER,
 )
-from esphome.core import CORE, CoroPriority, coroutine_with_priority
+from esphome.core import CORE, CoroPriority, Lambda, coroutine_with_priority
 from esphome.core.entity_helpers import (
     entity_duplicate_validator,
     queue_entity_register,
     setup_entity,
 )
-from esphome.cpp_generator import MockObjClass
+from esphome.cpp_generator import LambdaExpression, MockObjClass
 
 IS_PLATFORM_COMPONENT = True
 
@@ -487,38 +487,57 @@ CLIMATE_CONTROL_ACTION_SCHEMA = cv.Schema(
 )
 async def climate_control_to_code(config, action_id, template_arg, args):
     paren = await cg.get_variable(config[CONF_ID])
-    var = cg.new_Pvariable(action_id, template_arg, paren)
-    if (mode := config.get(CONF_MODE)) is not None:
-        template_ = await cg.templatable(mode, args, ClimateMode)
-        cg.add(var.set_mode(template_))
-    if (target_temp := config.get(CONF_TARGET_TEMPERATURE)) is not None:
-        template_ = await cg.templatable(target_temp, args, cg.float_)
-        cg.add(var.set_target_temperature(template_))
-    if (target_temp_low := config.get(CONF_TARGET_TEMPERATURE_LOW)) is not None:
-        template_ = await cg.templatable(target_temp_low, args, cg.float_)
-        cg.add(var.set_target_temperature_low(template_))
-    if (target_temp_high := config.get(CONF_TARGET_TEMPERATURE_HIGH)) is not None:
-        template_ = await cg.templatable(target_temp_high, args, cg.float_)
-        cg.add(var.set_target_temperature_high(template_))
-    if (target_humidity := config.get(CONF_TARGET_HUMIDITY)) is not None:
-        template_ = await cg.templatable(target_humidity, args, cg.float_)
-        cg.add(var.set_target_humidity(template_))
-    if (fan_mode := config.get(CONF_FAN_MODE)) is not None:
-        template_ = await cg.templatable(fan_mode, args, ClimateFanMode)
-        cg.add(var.set_fan_mode(template_))
-    if (custom_fan_mode := config.get(CONF_CUSTOM_FAN_MODE)) is not None:
-        template_ = await cg.templatable(custom_fan_mode, args, cg.std_string)
-        cg.add(var.set_custom_fan_mode(template_))
-    if (preset := config.get(CONF_PRESET)) is not None:
-        template_ = await cg.templatable(preset, args, ClimatePreset)
-        cg.add(var.set_preset(template_))
-    if (custom_preset := config.get(CONF_CUSTOM_PRESET)) is not None:
-        template_ = await cg.templatable(custom_preset, args, cg.std_string)
-        cg.add(var.set_custom_preset(template_))
-    if (swing_mode := config.get(CONF_SWING_MODE)) is not None:
-        template_ = await cg.templatable(swing_mode, args, ClimateSwingMode)
-        cg.add(var.set_swing_mode(template_))
-    return var
+
+    # All configured fields are folded into a single stateless lambda whose
+    # constants live in flash; the action stores only a function pointer.
+    # For custom_fan_mode/custom_preset the static-string path emits the
+    # (const char *, size_t) overload of set_fan_mode/set_preset to avoid
+    # constructing a std::string and calling runtime strlen.
+    FIELDS = (
+        (CONF_MODE, "set_mode", ClimateMode),
+        (CONF_TARGET_TEMPERATURE, "set_target_temperature", cg.float_),
+        (CONF_TARGET_TEMPERATURE_LOW, "set_target_temperature_low", cg.float_),
+        (CONF_TARGET_TEMPERATURE_HIGH, "set_target_temperature_high", cg.float_),
+        (CONF_TARGET_HUMIDITY, "set_target_humidity", cg.float_),
+        (CONF_FAN_MODE, "set_fan_mode", ClimateFanMode),
+        (CONF_CUSTOM_FAN_MODE, "set_fan_mode", cg.std_string),
+        (CONF_PRESET, "set_preset", ClimatePreset),
+        (CONF_CUSTOM_PRESET, "set_preset", cg.std_string),
+        (CONF_SWING_MODE, "set_swing_mode", ClimateSwingMode),
+    )
+
+    fwd_args = ", ".join(name for _, name in args)
+    body_lines: list[str] = []
+
+    for conf_key, setter, type_ in FIELDS:
+        if (value := config.get(conf_key)) is None:
+            continue
+        if isinstance(value, Lambda):
+            inner = await cg.process_lambda(value, args, return_type=type_)
+            body_lines.append(f"call.{setter}(({inner})({fwd_args}));")
+        elif type_ is cg.std_string:
+            # Static custom strings: emit a flash literal and pass the
+            # UTF-8 byte length to skip the runtime strlen inside
+            # set_fan_mode/set_preset.
+            literal = cg.safe_exp(value)
+            body_lines.append(
+                f"call.{setter}({literal}, {len(value.encode('utf-8'))});"
+            )
+        else:
+            body_lines.append(f"call.{setter}({cg.safe_exp(value)});")
+
+    # Match ControlAction::ApplyFn signature: const Ts &... for trigger args.
+    apply_args = [
+        (ClimateCall.operator("ref"), "call"),
+        *((t.operator("const").operator("ref"), n) for t, n in args),
+    ]
+    apply_lambda = LambdaExpression(
+        ["\n".join(body_lines)],
+        apply_args,
+        capture="",
+        return_type=cg.void,
+    )
+    return cg.new_Pvariable(action_id, template_arg, paren, apply_lambda)
 
 
 @coroutine_with_priority(CoroPriority.CORE)
diff --git a/esphome/components/climate/automation.h b/esphome/components/climate/automation.h
index fac56d9d9e..71d23fd6b6 100644
--- a/esphome/components/climate/automation.h
+++ b/esphome/components/climate/automation.h
@@ -5,42 +5,25 @@
 
 namespace esphome::climate {
 
+// All configured fields are baked into a single stateless lambda whose
+// constants live in flash. The action only stores one function pointer
+// plus one parent pointer, regardless of how many fields the user set.
+// Trigger args are forwarded to the apply function so user lambdas
+// (e.g. `target_temperature: !lambda "return x;"`) keep working.
 template<typename... Ts> class ControlAction : public Action<Ts...> {
  public:
-  explicit ControlAction(Climate *climate) : climate_(climate) {}
-
-  TEMPLATABLE_VALUE(ClimateMode, mode)
-  TEMPLATABLE_VALUE(float, target_temperature)
-  TEMPLATABLE_VALUE(float, target_temperature_low)
-  TEMPLATABLE_VALUE(float, target_temperature_high)
-  TEMPLATABLE_VALUE(float, target_humidity)
-  TEMPLATABLE_VALUE(bool, away)
-  TEMPLATABLE_VALUE(ClimateFanMode, fan_mode)
-  TEMPLATABLE_VALUE(std::string, custom_fan_mode)
-  TEMPLATABLE_VALUE(ClimatePreset, preset)
-  TEMPLATABLE_VALUE(std::string, custom_preset)
-  TEMPLATABLE_VALUE(ClimateSwingMode, swing_mode)
+  using ApplyFn = void (*)(ClimateCall &, const Ts &...);
+  ControlAction(Climate *climate, ApplyFn apply) : climate_(climate), apply_(apply) {}
 
   void play(const Ts &...x) override {
     auto call = this->climate_->make_call();
-    call.set_mode(this->mode_.optional_value(x...));
-    call.set_target_temperature(this->target_temperature_.optional_value(x...));
-    call.set_target_temperature_low(this->target_temperature_low_.optional_value(x...));
-    call.set_target_temperature_high(this->target_temperature_high_.optional_value(x...));
-    call.set_target_humidity(this->target_humidity_.optional_value(x...));
-    if (away_.has_value()) {
-      call.set_preset(away_.value(x...) ? CLIMATE_PRESET_AWAY : CLIMATE_PRESET_HOME);
-    }
-    call.set_fan_mode(this->fan_mode_.optional_value(x...));
-    call.set_fan_mode(this->custom_fan_mode_.optional_value(x...));
-    call.set_preset(this->preset_.optional_value(x...));
-    call.set_preset(this->custom_preset_.optional_value(x...));
-    call.set_swing_mode(this->swing_mode_.optional_value(x...));
+    this->apply_(call, x...);
     call.perform();
   }
 
  protected:
   Climate *climate_;
+  ApplyFn apply_;
 };
 
 class ControlTrigger : public Trigger<ClimateCall &> {
diff --git a/tests/integration/fixtures/climate_control_action.yaml b/tests/integration/fixtures/climate_control_action.yaml
new file mode 100644
index 0000000000..1dd300fcc2
--- /dev/null
+++ b/tests/integration/fixtures/climate_control_action.yaml
@@ -0,0 +1,92 @@
+esphome:
+  name: climate-control-action-test
+host:
+api:
+logger:
+  level: DEBUG
+
+globals:
+  - id: test_target_temp
+    type: float
+    initial_value: "21.5"
+
+sensor:
+  - platform: template
+    id: temp_sensor
+    name: "Temp"
+    lambda: 'return 20.0;'
+    update_interval: 60s
+
+climate:
+  - platform: thermostat
+    id: test_climate
+    name: "Test Climate"
+    sensor: temp_sensor
+    min_idle_time: 30s
+    min_heating_off_time: 300s
+    min_heating_run_time: 300s
+    min_cooling_off_time: 300s
+    min_cooling_run_time: 300s
+    heat_action:
+      - logger.log: heating
+    idle_action:
+      - logger.log: idle
+    cool_action:
+      - logger.log: cooling
+    heat_cool_mode:
+      - logger.log: heat_cool
+    preset:
+      - name: Default
+        default_target_temperature_low: 18 °C
+        default_target_temperature_high: 22 °C
+    visual:
+      min_temperature: 10 °C
+      max_temperature: 30 °C
+
+button:
+  # mode only
+  - platform: template
+    id: btn_mode
+    name: "Set Mode Heat"
+    on_press:
+      - climate.control:
+          id: test_climate
+          mode: HEAT
+
+  # mode + target_temperature_low + target_temperature_high
+  - platform: template
+    id: btn_mode_temps
+    name: "Set Mode Temps"
+    on_press:
+      - climate.control:
+          id: test_climate
+          mode: HEAT_COOL
+          target_temperature_low: 19.0 °C
+          target_temperature_high: 23.0 °C
+
+  # target_temperature_low only
+  - platform: template
+    id: btn_low_only
+    name: "Set Low Only"
+    on_press:
+      - climate.control:
+          id: test_climate
+          target_temperature_low: 17.5 °C
+
+  # Lambda path: target_temperature_high computed at runtime
+  - platform: template
+    id: btn_lambda_high
+    name: "Lambda High"
+    on_press:
+      - climate.control:
+          id: test_climate
+          target_temperature_high: !lambda "return id(test_target_temp);"
+
+  # mode only — turn off via mode
+  - platform: template
+    id: btn_off
+    name: "Set Off"
+    on_press:
+      - climate.control:
+          id: test_climate
+          mode: "OFF"
diff --git a/tests/integration/test_climate_control_action.py b/tests/integration/test_climate_control_action.py
new file mode 100644
index 0000000000..2b0293b209
--- /dev/null
+++ b/tests/integration/test_climate_control_action.py
@@ -0,0 +1,84 @@
+"""Integration test for climate ControlAction.
+
+Tests that climate.control automation actions work correctly with the
+single stateless apply lambda/function pointer implementation. Exercises
+multiple field combinations and the lambda path.
+"""
+
+from __future__ import annotations
+
+import asyncio
+
+from aioesphomeapi import (
+    ButtonInfo,
+    ClimateInfo,
+    ClimateMode,
+    ClimateState,
+    EntityState,
+)
+import pytest
+
+from .state_utils import InitialStateHelper, require_entity
+from .types import APIClientConnectedFactory, RunCompiledFunction
+
+
+@pytest.mark.asyncio
+async def test_climate_control_action(
+    yaml_config: str,
+    run_compiled: RunCompiledFunction,
+    api_client_connected: APIClientConnectedFactory,
+) -> None:
+    """Test climate ControlAction with constants and lambdas."""
+    loop = asyncio.get_running_loop()
+    async with run_compiled(yaml_config), api_client_connected() as client:
+        climate_state_future: asyncio.Future[ClimateState] | None = None
+
+        def on_state(state: EntityState) -> None:
+            if (
+                isinstance(state, ClimateState)
+                and climate_state_future is not None
+                and not climate_state_future.done()
+            ):
+                climate_state_future.set_result(state)
+
+        async def wait_for_climate_state(timeout: float = 5.0) -> ClimateState:
+            nonlocal climate_state_future
+            climate_state_future = loop.create_future()
+            try:
+                return await asyncio.wait_for(climate_state_future, timeout)
+            finally:
+                climate_state_future = None
+
+        entities, _ = await client.list_entities_services()
+        initial_state_helper = InitialStateHelper(entities)
+        client.subscribe_states(initial_state_helper.on_state_wrapper(on_state))
+        await initial_state_helper.wait_for_initial_states()
+
+        require_entity(entities, "test_climate", ClimateInfo)
+
+        async def press_and_wait(name: str) -> ClimateState:
+            btn = require_entity(entities, name.lower().replace(" ", "_"), ButtonInfo)
+            client.button_command(btn.key)
+            return await wait_for_climate_state()
+
+        # mode only — set HEAT
+        state = await press_and_wait("Set Mode Heat")
+        assert state.mode == ClimateMode.HEAT
+
+        # mode + target_temperature_low + target_temperature_high
+        state = await press_and_wait("Set Mode Temps")
+        assert state.mode == ClimateMode.HEAT_COOL
+        assert state.target_temperature_low == pytest.approx(19.0, abs=0.5)
+        assert state.target_temperature_high == pytest.approx(23.0, abs=0.5)
+
+        # target_temperature_low only
+        state = await press_and_wait("Set Low Only")
+        assert state.target_temperature_low == pytest.approx(17.5, abs=0.5)
+
+        # lambda path: target_temperature_high computed at runtime
+        state = await press_and_wait("Lambda High")
+        assert state.target_temperature_high == pytest.approx(21.5, abs=0.5)
+
+        # mode only — turn off via mode
+        state = await press_and_wait("Set Off")
+        assert state.mode == ClimateMode.OFF

From 92aa98f680c8ea40ff424dfd6a529fd3f6313d7f Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 19:42:38 -0500
Subject: [PATCH 15/19] [host] Move HAL bodies into components/host/hal.cpp +
 inline trivial dispatches (#16115)

---
 esphome/components/host/core.cpp | 60 +----------------------------
 esphome/components/host/hal.cpp  | 65 ++++++++++++++++++++++++++++++++
 esphome/core/hal/hal_host.h      | 12 +++---
 3 files changed, 73 insertions(+), 64 deletions(-)
 create mode 100644 esphome/components/host/hal.cpp

diff --git a/esphome/components/host/core.cpp b/esphome/components/host/core.cpp
index b067ebbf6e..9123975884 100644
--- a/esphome/components/host/core.cpp
+++ b/esphome/components/host/core.cpp
@@ -1,74 +1,16 @@
 #ifdef USE_HOST
 
 #include "esphome/core/application.h"
-#include "esphome/core/hal.h"
-#include "esphome/core/helpers.h"
 #include "preferences.h"
 
 #include <csignal>
-#include <sched.h>
-#include <time.h>
-#include <cstdlib>
 
 namespace {
 volatile sig_atomic_t s_signal_received = 0;  // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
 void signal_handler(int signal) { s_signal_received = signal; }
 }  // namespace
 
-namespace esphome {
-
-void HOT yield() { ::sched_yield(); }
-uint32_t IRAM_ATTR HOT millis() {
-  struct timespec spec;
-  clock_gettime(CLOCK_MONOTONIC, &spec);
-  return static_cast<uint32_t>(spec.tv_sec * 1000ULL + spec.tv_nsec / 1000000);
-}
-uint64_t millis_64() {
-  struct timespec spec;
-  clock_gettime(CLOCK_MONOTONIC, &spec);
-  return static_cast<uint64_t>(spec.tv_sec) * 1000ULL + static_cast<uint64_t>(spec.tv_nsec) / 1000000ULL;
-}
-void HOT delay(uint32_t ms) {
-  struct timespec ts;
-  ts.tv_sec = ms / 1000;
-  ts.tv_nsec = (ms % 1000) * 1000000;
-  int res;
-  do {
-    res = nanosleep(&ts, &ts);
-  } while (res != 0 && errno == EINTR);
-}
-uint32_t IRAM_ATTR HOT micros() {
-  struct timespec spec;
-  clock_gettime(CLOCK_MONOTONIC, &spec);
-  return static_cast<uint32_t>(spec.tv_sec * 1000000ULL + spec.tv_nsec / 1000);
-}
-void IRAM_ATTR HOT delayMicroseconds(uint32_t us) {
-  struct timespec ts;
-  ts.tv_sec = us / 1000000U;
-  ts.tv_nsec = (us % 1000000U) * 1000U;
-  int res;
-  do {
-    res = nanosleep(&ts, &ts);
-  } while (res != 0 && errno == EINTR);
-}
-void arch_restart() { exit(0); }
-void arch_init() {
-  // pass
-}
-void HOT arch_feed_wdt() {
-  // pass
-}
-
-uint32_t arch_get_cpu_cycle_count() {
-  struct timespec spec;
-  clock_gettime(CLOCK_MONOTONIC, &spec);
-  time_t seconds = spec.tv_sec;
-  uint32_t us = spec.tv_nsec;
-  return ((uint32_t) seconds) * 1000000000U + us;
-}
-uint32_t arch_get_cpu_freq_hz() { return 1000000000U; }
-
-}  // namespace esphome
+// HAL functions live in hal.cpp.
 
 void setup();
 void loop();
diff --git a/esphome/components/host/hal.cpp b/esphome/components/host/hal.cpp
new file mode 100644
index 0000000000..256a12ac62
--- /dev/null
+++ b/esphome/components/host/hal.cpp
@@ -0,0 +1,65 @@
+#ifdef USE_HOST
+
+#include "esphome/core/hal.h"
+#include "esphome/core/helpers.h"
+
+#include <time.h>
+#include <cerrno>
+#include <cstdlib>
+
+// Empty host namespace block to satisfy ci-custom's lint_namespace check.
+// HAL functions live in namespace esphome (root) — they are not part of the
+// host component's API.
+namespace esphome::host {}  // namespace esphome::host
+
+namespace esphome {
+
+// yield(), arch_init(), arch_feed_wdt(), arch_get_cpu_freq_hz() inlined in
+// core/hal/hal_host.h.
+
+uint32_t IRAM_ATTR HOT millis() {
+  struct timespec spec;
+  clock_gettime(CLOCK_MONOTONIC, &spec);
+  return static_cast<uint32_t>(spec.tv_sec * 1000ULL + spec.tv_nsec / 1000000);
+}
+uint64_t millis_64() {
+  struct timespec spec;
+  clock_gettime(CLOCK_MONOTONIC, &spec);
+  return static_cast<uint64_t>(spec.tv_sec) * 1000ULL + static_cast<uint64_t>(spec.tv_nsec) / 1000000ULL;
+}
+void HOT delay(uint32_t ms) {
+  struct timespec ts;
+  ts.tv_sec = ms / 1000;
+  ts.tv_nsec = (ms % 1000) * 1000000;
+  int res;
+  do {
+    res = nanosleep(&ts, &ts);
+  } while (res != 0 && errno == EINTR);
+}
+uint32_t IRAM_ATTR HOT micros() {
+  struct timespec spec;
+  clock_gettime(CLOCK_MONOTONIC, &spec);
+  return static_cast<uint32_t>(spec.tv_sec * 1000000ULL + spec.tv_nsec / 1000);
+}
+void IRAM_ATTR HOT delayMicroseconds(uint32_t us) {
+  struct timespec ts;
+  ts.tv_sec = us / 1000000U;
+  ts.tv_nsec = (us % 1000000U) * 1000U;
+  int res;
+  do {
+    res = nanosleep(&ts, &ts);
+  } while (res != 0 && errno == EINTR);
+}
+void arch_restart() { exit(0); }
+
+uint32_t arch_get_cpu_cycle_count() {
+  struct timespec spec;
+  clock_gettime(CLOCK_MONOTONIC, &spec);
+  time_t seconds = spec.tv_sec;
+  uint32_t ns = static_cast<uint32_t>(spec.tv_nsec);
+  return static_cast<uint32_t>(seconds) * 1000000000U + ns;
+}
+
+}  // namespace esphome
+
+#endif  // USE_HOST
diff --git a/esphome/core/hal/hal_host.h b/esphome/core/hal/hal_host.h
index a8896fdf63..d7f317176e 100644
--- a/esphome/core/hal/hal_host.h
+++ b/esphome/core/hal/hal_host.h
@@ -3,6 +3,7 @@
 #ifdef USE_HOST
 
 #include <cstdint>
+#include <sched.h>
 
 #define IRAM_ATTR
 #define PROGMEM
@@ -13,17 +14,18 @@ namespace esphome {
 /// Host has no ISR concept.
 __attribute__((always_inline)) inline bool in_isr_context() { return false; }
 
-void yield();
+__attribute__((always_inline)) inline void yield() { ::sched_yield(); }
+
 void delay(uint32_t ms);
 uint32_t micros();
 uint32_t millis();
 uint64_t millis_64();
-
 void delayMicroseconds(uint32_t us);  // NOLINT(readability-identifier-naming)
-void arch_feed_wdt();
 uint32_t arch_get_cpu_cycle_count();
-void arch_init();
-uint32_t arch_get_cpu_freq_hz();
+
+__attribute__((always_inline)) inline void arch_init() {}
+__attribute__((always_inline)) inline void arch_feed_wdt() {}
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_freq_hz() { return 1000000000U; }
 
 }  // namespace esphome
 

From 9999913d072b930bdf330ba421e403126d7b65a7 Mon Sep 17 00:00:00 2001
From: "J. Nick Koston" <nick@koston.org>
Date: Thu, 30 Apr 2026 20:10:51 -0500
Subject: [PATCH 16/19] [zephyr] Move HAL bodies into components/zephyr/hal.cpp
 + inline trivial dispatches (#16116)

---
 esphome/components/zephyr/core.cpp | 52 +-----------------------
 esphome/components/zephyr/hal.cpp  | 63 ++++++++++++++++++++++++++++++
 esphome/core/hal/hal_zephyr.h      | 20 ++++++----
 3 files changed, 76 insertions(+), 59 deletions(-)
 create mode 100644 esphome/components/zephyr/hal.cpp

diff --git a/esphome/components/zephyr/core.cpp b/esphome/components/zephyr/core.cpp
index 93a9a1ae8e..d1bdaee02d 100644
--- a/esphome/components/zephyr/core.cpp
+++ b/esphome/components/zephyr/core.cpp
@@ -1,8 +1,6 @@
 #ifdef USE_ZEPHYR
 
 #include <zephyr/kernel.h>
-#include <zephyr/drivers/watchdog.h>
-#include <zephyr/sys/reboot.h>
 #include <zephyr/random/random.h>
 #include "esphome/core/hal.h"
 #include "esphome/core/helpers.h"
@@ -10,55 +8,7 @@
 
 namespace esphome {
 
-#ifdef CONFIG_WATCHDOG
-static int wdt_channel_id = -1;  // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
-static const device *const WDT = DEVICE_DT_GET(DT_ALIAS(watchdog0));
-#endif
-
-void yield() { ::k_yield(); }
-uint32_t millis() { return static_cast<uint32_t>(millis_64()); }
-uint64_t millis_64() { return static_cast<uint64_t>(k_uptime_get()); }
-uint32_t micros() { return k_ticks_to_us_floor32(k_uptime_ticks()); }
-void delayMicroseconds(uint32_t us) { ::k_usleep(us); }
-void delay(uint32_t ms) { ::k_msleep(ms); }
-
-void arch_init() {
-#ifdef CONFIG_WATCHDOG
-  if (device_is_ready(WDT)) {
-    static wdt_timeout_cfg wdt_config{};
-    wdt_config.flags = WDT_FLAG_RESET_SOC;
-#ifdef USE_ZIGBEE
-    // zboss thread use a lot of cpu cycles during start
-    wdt_config.window.max = 10000;
-#else
-    wdt_config.window.max = 2000;
-#endif
-    wdt_channel_id = wdt_install_timeout(WDT, &wdt_config);
-    if (wdt_channel_id >= 0) {
-      uint8_t options = 0;
-#ifdef USE_DEBUG
-      options |= WDT_OPT_PAUSE_HALTED_BY_DBG;
-#endif
-#ifdef USE_DEEP_SLEEP
-      options |= WDT_OPT_PAUSE_IN_SLEEP;
-#endif
-      wdt_setup(WDT, options);
-    }
-  }
-#endif
-}
-
-void arch_feed_wdt() {
-#ifdef CONFIG_WATCHDOG
-  if (wdt_channel_id >= 0) {
-    wdt_feed(WDT, wdt_channel_id);
-  }
-#endif
-}
-
-void arch_restart() { sys_reboot(SYS_REBOOT_COLD); }
-uint32_t arch_get_cpu_cycle_count() { return k_cycle_get_32(); }
-uint32_t arch_get_cpu_freq_hz() { return sys_clock_hw_cycles_per_sec(); }
+// HAL functions live in hal.cpp.
 
 Mutex::Mutex() {
   auto *mutex = new k_mutex();
diff --git a/esphome/components/zephyr/hal.cpp b/esphome/components/zephyr/hal.cpp
new file mode 100644
index 0000000000..5c08ed2519
--- /dev/null
+++ b/esphome/components/zephyr/hal.cpp
@@ -0,0 +1,63 @@
+#ifdef USE_ZEPHYR
+
+#include "esphome/core/defines.h"
+#include "esphome/core/hal.h"
+
+#include <zephyr/drivers/watchdog.h>
+#include <zephyr/sys/reboot.h>
+
+// Empty zephyr namespace block to satisfy ci-custom's lint_namespace check.
+// HAL functions live in namespace esphome (root) — they are not part of the
+// zephyr component's API.
+namespace esphome::zephyr {}  // namespace esphome::zephyr
+
+namespace esphome {
+
+#ifdef CONFIG_WATCHDOG
+static int wdt_channel_id = -1;  // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
+static const device *const WDT = DEVICE_DT_GET(DT_ALIAS(watchdog0));
+#endif
+
+// yield(), delay(), micros(), millis(), millis_64(), delayMicroseconds(),
+// arch_get_cpu_cycle_count(), arch_get_cpu_freq_hz() inlined in
+// core/hal/hal_zephyr.h.
+
+void arch_init() {
+#ifdef CONFIG_WATCHDOG
+  if (device_is_ready(WDT)) {
+    static wdt_timeout_cfg wdt_config{};
+    wdt_config.flags = WDT_FLAG_RESET_SOC;
+#ifdef USE_ZIGBEE
+    // zboss thread uses a lot of CPU cycles during startup
+    wdt_config.window.max = 10000;
+#else
+    wdt_config.window.max = 2000;
+#endif
+    wdt_channel_id = wdt_install_timeout(WDT, &wdt_config);
+    if (wdt_channel_id >= 0) {
+      uint8_t options = 0;
+#ifdef USE_DEBUG
+      options |= WDT_OPT_PAUSE_HALTED_BY_DBG;
+#endif
+#ifdef USE_DEEP_SLEEP
+      options |= WDT_OPT_PAUSE_IN_SLEEP;
+#endif
+      wdt_setup(WDT, options);
+    }
+  }
+#endif
+}
+
+void arch_feed_wdt() {
+#ifdef CONFIG_WATCHDOG
+  if (wdt_channel_id >= 0) {
+    wdt_feed(WDT, wdt_channel_id);
+  }
+#endif
+}
+
+void arch_restart() { sys_reboot(SYS_REBOOT_COLD); }
+
+}  // namespace esphome
+
+#endif  // USE_ZEPHYR
diff --git a/esphome/core/hal/hal_zephyr.h b/esphome/core/hal/hal_zephyr.h
index d4b37b5eb6..613b3911c1 100644
--- a/esphome/core/hal/hal_zephyr.h
+++ b/esphome/core/hal/hal_zephyr.h
@@ -4,6 +4,8 @@
 
 #include <cstdint>
 
+#include <zephyr/kernel.h>
+
 #define IRAM_ATTR
 #define PROGMEM
 
@@ -13,17 +15,19 @@ namespace esphome {
 /// Zephyr/nRF52: not currently consulted — wake path is platform-specific.
 __attribute__((always_inline)) inline bool in_isr_context() { return false; }
 
-void yield();
-void delay(uint32_t ms);
-uint32_t micros();
-uint32_t millis();
-uint64_t millis_64();
+__attribute__((always_inline)) inline void yield() { ::k_yield(); }
+__attribute__((always_inline)) inline void delay(uint32_t ms) { ::k_msleep(ms); }
+__attribute__((always_inline)) inline uint32_t micros() { return k_ticks_to_us_floor32(k_uptime_ticks()); }
+__attribute__((always_inline)) inline uint64_t millis_64() { return static_cast<uint64_t>(k_uptime_get()); }
+__attribute__((always_inline)) inline uint32_t millis() { return static_cast<uint32_t>(millis_64()); }
+
+// NOLINTNEXTLINE(readability-identifier-naming)
+__attribute__((always_inline)) inline void delayMicroseconds(uint32_t us) { ::k_usleep(us); }
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_cycle_count() { return k_cycle_get_32(); }
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_freq_hz() { return sys_clock_hw_cycles_per_sec(); }
 
-void delayMicroseconds(uint32_t us);  // NOLINT(readability-identifier-naming)
 void arch_feed_wdt();
-uint32_t arch_get_cpu_cycle_count();
 void arch_init();
-uint32_t arch_get_cpu_freq_hz();
 
 }  // namespace esphome
 

From faa61696e036cb26e44ef0e4a28400e0636c5ca3 Mon Sep 17 00:00:00 2001
From: Kevin Ahrendt <kevin.ahrendt@openhomefoundation.org>
Date: Thu, 30 Apr 2026 21:43:24 -0400
Subject: [PATCH 17/19] [sendspin] Use sendspin-cpp to v0.4.0 to reduce
 stuttering (#16178)

---
 esphome/components/sendspin/__init__.py       | 30 ++++++++++++++++---
 .../sendspin/media_source/__init__.py         |  4 +++
 esphome/idf_component.yml                     |  2 +-
 .../sendspin/common-media_source.yaml         |  1 +
 4 files changed, 32 insertions(+), 5 deletions(-)

diff --git a/esphome/components/sendspin/__init__.py b/esphome/components/sendspin/__init__.py
index 58687ae838..1348bf2bfc 100644
--- a/esphome/components/sendspin/__init__.py
+++ b/esphome/components/sendspin/__init__.py
@@ -24,6 +24,7 @@ CONF_SENDSPIN_ID = "sendspin_id"
 
 CONF_INITIAL_STATIC_DELAY = "initial_static_delay"
 CONF_FIXED_DELAY = "fixed_delay"
+CONF_DECODE_MEMORY = "decode_memory"
 
 # sendspin-cpp library lives in the global `sendspin` namespace.
 sendspin_library_ns = cg.global_ns.namespace("sendspin")
@@ -39,6 +40,18 @@ CODEC_FORMAT_UNSUPPORTED = SendspinCodecFormat.enum("UNSUPPORTED")
 AudioSupportedFormatObject = sendspin_library_ns.struct("AudioSupportedFormatObject")
 PlayerRoleConfig = sendspin_library_ns.struct("PlayerRoleConfig")
 
+# MemoryLocation enum (from sendspin/types.h) controls SPIRAM-vs-internal-RAM placement
+# preference for the player role's transfer buffers.
+SendspinMemoryLocation = sendspin_library_ns.enum("MemoryLocation", is_class=True)
+
+MEMORY_PSRAM = "psram"
+MEMORY_INTERNAL = "internal"
+MEMORY_LOCATIONS = [MEMORY_PSRAM, MEMORY_INTERNAL]
+MEMORY_LOCATION_ENUM = {
+    MEMORY_PSRAM: SendspinMemoryLocation.PREFER_EXTERNAL,
+    MEMORY_INTERNAL: SendspinMemoryLocation.PREFER_INTERNAL,
+}
+
 # Trailing underscore avoids clashing with sendspin-cpp's global `sendspin` namespace.
 # Analysis tools strip the trailing underscore (same pattern as `template_`).
 sendspin_ns = cg.esphome_ns.namespace("sendspin_")
@@ -193,7 +206,7 @@ async def to_code(config: ConfigType) -> None:
         )
 
     # sendspin-cpp library
-    esp32.add_idf_component(name="sendspin/sendspin-cpp", ref="0.3.1")
+    esp32.add_idf_component(name="sendspin/sendspin-cpp", ref="0.4.0")
 
     cg.add_define("USE_SENDSPIN", True)  # for MDNS
 
@@ -249,14 +262,23 @@ async def to_code(config: ConfigType) -> None:
                 "CONFIG_SPIRAM_ALLOW_STACK_EXTERNAL_MEMORY", True
             )
 
-        player_config_struct = cg.StructInitializer(
-            PlayerRoleConfig,
+        # Library defaults: priority 18 (one above httpd_priority 17 so the decoder is not
+        # starved by the HTTP server during the initial encoded-audio burst at stream start),
+        # interpolation/decode buffer locations PREFER_EXTERNAL.
+        player_struct_fields = [
             ("audio_formats", audio_format_structs),
             ("audio_buffer_capacity", player_cfg[CONF_BUFFER_SIZE]),
             ("fixed_delay_us", player_cfg[CONF_FIXED_DELAY]),
             ("initial_static_delay_ms", player_cfg[CONF_INITIAL_STATIC_DELAY]),
             ("psram_stack", psram_stack),
-            ("priority", 2),
+        ]
+        if (decode_memory := player_cfg.get(CONF_DECODE_MEMORY)) is not None:
+            player_struct_fields.append(
+                ("decode_buffer_location", MEMORY_LOCATION_ENUM[decode_memory])
+            )
+        player_config_struct = cg.StructInitializer(
+            PlayerRoleConfig,
+            *player_struct_fields,
         )
         cg.add(var.set_player_config(player_config_struct))
     else:
diff --git a/esphome/components/sendspin/media_source/__init__.py b/esphome/components/sendspin/media_source/__init__.py
index 6d61a8a636..f689ab01cb 100644
--- a/esphome/components/sendspin/media_source/__init__.py
+++ b/esphome/components/sendspin/media_source/__init__.py
@@ -13,9 +13,11 @@ from esphome.cpp_generator import MockObj, TemplateArgsType
 from esphome.types import ConfigType
 
 from .. import (
+    CONF_DECODE_MEMORY,
     CONF_FIXED_DELAY,
     CONF_INITIAL_STATIC_DELAY,
     CONF_SENDSPIN_ID,
+    MEMORY_LOCATIONS,
     SendspinHub,
     _validate_task_stack_in_psram,
     register_player_config,
@@ -57,6 +59,7 @@ def _register(config: ConfigType) -> ConfigType:
             CONF_INITIAL_STATIC_DELAY: config[CONF_INITIAL_STATIC_DELAY],
             CONF_FIXED_DELAY: config[CONF_FIXED_DELAY],
             CONF_TASK_STACK_IN_PSRAM: config.get(CONF_TASK_STACK_IN_PSRAM, False),
+            CONF_DECODE_MEMORY: config.get(CONF_DECODE_MEMORY),
         }
     )
     return config
@@ -82,6 +85,7 @@ CONFIG_SCHEMA = cv.All(
             cv.Optional(CONF_SAMPLE_RATE, default=48000): cv.int_range(
                 min=16000, max=96000
             ),
+            cv.Optional(CONF_DECODE_MEMORY): cv.one_of(*MEMORY_LOCATIONS, lower=True),
         }
     ),
     cv.only_on_esp32,
diff --git a/esphome/idf_component.yml b/esphome/idf_component.yml
index 757fcf9dd7..d02c3adb8f 100644
--- a/esphome/idf_component.yml
+++ b/esphome/idf_component.yml
@@ -92,6 +92,6 @@ dependencies:
   esp32async/asynctcp:
     version: 3.4.91
   sendspin/sendspin-cpp:
-    version: 0.3.1
+    version: 0.4.0
   lvgl/lvgl:
     version: 9.5.0
diff --git a/tests/components/sendspin/common-media_source.yaml b/tests/components/sendspin/common-media_source.yaml
index 4a7cd79c67..5b33a54647 100644
--- a/tests/components/sendspin/common-media_source.yaml
+++ b/tests/components/sendspin/common-media_source.yaml
@@ -7,3 +7,4 @@ media_source:
     initial_static_delay: 5ms
     static_delay_adjustable: true
     fixed_delay: 480us
+    decode_memory: internal

From 08e5cb55764033d11e4feb35c2c6a7fc3fa2da50 Mon Sep 17 00:00:00 2001
From: Jonathan Swoboda <154711427+swoboda1337@users.noreply.github.com>
Date: Thu, 30 Apr 2026 21:47:22 -0400
Subject: [PATCH 18/19] [esp32_hosted] Bump esp_hosted to 2.12.6 and
 esp_wifi_remote to 1.5.1 (#16176)

---
 esphome/components/esp32_hosted/__init__.py |  7 ++++---
 esphome/idf_component.yml                   | 10 +++++++---
 2 files changed, 11 insertions(+), 6 deletions(-)

diff --git a/esphome/components/esp32_hosted/__init__.py b/esphome/components/esp32_hosted/__init__.py
index 1619a845d8..eca7c24b10 100644
--- a/esphome/components/esp32_hosted/__init__.py
+++ b/esphome/components/esp32_hosted/__init__.py
@@ -246,9 +246,10 @@ async def to_code(config):
     idf_ver = esp32.idf_version()
     os.environ["ESP_IDF_VERSION"] = f"{idf_ver.major}.{idf_ver.minor}"
     if idf_ver >= cv.Version(5, 5, 0):
-        esp32.add_idf_component(name="espressif/esp_wifi_remote", ref="1.4.0")
-        esp32.add_idf_component(name="espressif/eppp_link", ref="1.1.4")
-        esp32.add_idf_component(name="espressif/esp_hosted", ref="2.12.1")
+        esp32.add_idf_component(name="espressif/esp_wifi_remote", ref="1.5.1")
+        esp32.add_idf_component(name="espressif/wifi_remote_over_eppp", ref="0.3.2")
+        esp32.add_idf_component(name="espressif/eppp_link", ref="1.1.5")
+        esp32.add_idf_component(name="espressif/esp_hosted", ref="2.12.6")
     else:
         esp32.add_idf_component(name="espressif/esp_wifi_remote", ref="0.13.0")
         esp32.add_idf_component(name="espressif/eppp_link", ref="0.2.0")
diff --git a/esphome/idf_component.yml b/esphome/idf_component.yml
index d02c3adb8f..f5a8dd8c60 100644
--- a/esphome/idf_component.yml
+++ b/esphome/idf_component.yml
@@ -20,15 +20,19 @@ dependencies:
   espressif/mdns:
     version: 1.11.0
   espressif/esp_wifi_remote:
-    version: 1.4.0
+    version: 1.5.1
+    rules:
+      - if: "target in [esp32h2, esp32p4]"
+  espressif/wifi_remote_over_eppp:
+    version: 0.3.2
     rules:
       - if: "target in [esp32h2, esp32p4]"
   espressif/eppp_link:
-    version: 1.1.4
+    version: 1.1.5
     rules:
       - if: "target in [esp32h2, esp32p4]"
   espressif/esp_hosted:
-    version: 2.12.1
+    version: 2.12.6
     rules:
       - if: "target in [esp32h2, esp32p4]"
   zorxx/multipart-parser:

From f6e39d305d5410e7524ffb4d9d7fb0b388c861ae Mon Sep 17 00:00:00 2001
From: luar123 <49960470+luar123@users.noreply.github.com>
Date: Fri, 1 May 2026 04:08:55 +0200
Subject: [PATCH 19/19] [zigbee] Add newlib compatibility for zigbee sdk in idf
 6 (#16174)

---
 esphome/components/zigbee/zigbee_esp32.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/esphome/components/zigbee/zigbee_esp32.py b/esphome/components/zigbee/zigbee_esp32.py
index 1b98df6c0a..9081582c7b 100644
--- a/esphome/components/zigbee/zigbee_esp32.py
+++ b/esphome/components/zigbee/zigbee_esp32.py
@@ -9,6 +9,7 @@ from esphome.components.esp32 import (
     add_idf_component,
     add_idf_sdkconfig_option,
     add_partition,
+    idf_version,
     require_vfs_select,
 )
 import esphome.config_validation as cv
@@ -186,6 +187,10 @@ async def _zigbee_add_sdkconfigs(config: ConfigType) -> None:
     # The pre-built Zigbee library uses esp_log_default_level which requires
     # dynamic log level control to be enabled
     add_idf_sdkconfig_option("CONFIG_LOG_DYNAMIC_LEVEL_CONTROL", True)
+    # The pre-built Zigbee library is compiled against newlib which requires newlib
+    # reentrancy to be enabled with picolibc compatibility.
+    if idf_version() >= cv.Version(6, 0, 0):
+        add_idf_sdkconfig_option("CONFIG_LIBC_PICOLIBC_NEWLIB_COMPATIBILITY", True)
 
 
 async def attributes_to_code(