[esp32] Move HAL bodies from components/esp32/core.cpp into core/hal/

Per-platform follow-up to #15977.

Move out-of-line HAL bodies (millis(), arch_restart(), arch_get_cpu_freq_hz())
from components/esp32/core.cpp into a new esphome/core/hal/hal_esp32.cpp so
HAL implementation finally lives next to the HAL header.

Inline the trivial one-liners directly in hal_esp32.h:
  delayMicroseconds(us)        -> delay_microseconds_safe(us)
  arch_feed_wdt()              -> esp_task_wdt_reset()
  arch_get_cpu_cycle_count()   -> esp_cpu_get_cycle_count()

arch_init() stays in components/esp32/core.cpp because it uses
esp32::crash_handler_read_and_clear() from the component-private
crash_handler.h header — moving it would require an awkward layering
inversion of core/ -> components/esp32/.

config.py FILTER_SOURCE_FILES gains a hal/hal_esp32.cpp entry so only
ESP32 builds compile it (same pattern wake/wake_*.cpp uses).

Cross-platform decls of delayMicroseconds(), arch_feed_wdt(), and
arch_get_cpu_cycle_count() are dropped from the dispatcher hal.h and
moved into each per-platform header — that way subsequent per-platform
follow-up PRs (libretiny, rp2040, host, zephyr) only need to touch
their own platform header instead of accumulating gates in hal.h.

esphome/components/esp32/core.cpp

@@ -6,12 +6,8 @@
 #include "esphome/core/hal.h"
 #include "esphome/core/helpers.h"
 #include "preferences.h"
-#include <esp_clk_tree.h>
-#include <esp_cpu.h>
-#include <esp_idf_version.h>
 #include <esp_ota_ops.h>
 #include <esp_task_wdt.h>
-#include <esp_timer.h>
 #include <freertos/FreeRTOS.h>
 #include <freertos/task.h>
 
@@ -23,29 +19,8 @@ extern "C" __attribute__((weak)) void initArduino() {}
 namespace esphome {
 
 // yield(), delay(), micros(), millis_64() inlined in hal.h.
-// Use xTaskGetTickCount() when tick rate is 1 kHz (ESPHome's default via sdkconfig),
-// falling back to esp_timer for non-standard rates. IRAM_ATTR is required because
-// Wiegand and ZyAura call millis() from IRAM_ATTR ISR handlers on ESP32.
-// xTaskGetTickCountFromISR() is used in ISR context to satisfy the FreeRTOS API contract.
-uint32_t IRAM_ATTR HOT millis() {
-#if CONFIG_FREERTOS_HZ == 1000
-  if (xPortInIsrContext()) [[unlikely]] {
-    return xTaskGetTickCountFromISR();
-  }
-  return xTaskGetTickCount();
-#else
-  return micros_to_millis(static_cast<uint64_t>(esp_timer_get_time()));
-#endif
-}
-void IRAM_ATTR HOT delayMicroseconds(uint32_t us) { delay_microseconds_safe(us); }
-void arch_restart() {
-  esp_restart();
-  // restart() doesn't always end execution
-  while (true) {  // NOLINT(clang-diagnostic-unreachable-code)
-    yield();
-  }
-}
-
+// millis(), arch_restart(), arch_get_cpu_freq_hz() out-of-line in hal/hal_esp32.cpp.
+// delayMicroseconds(), arch_feed_wdt(), arch_get_cpu_cycle_count() inlined in hal/hal_esp32.h.
 void arch_init() {
 #ifdef USE_ESP32_CRASH_HANDLER
   // Read crash data from previous boot before anything else
@@ -61,14 +36,6 @@ void arch_init() {
   esp_ota_mark_app_valid_cancel_rollback();
 #endif
 }
-void HOT arch_feed_wdt() { esp_task_wdt_reset(); }
-
-uint32_t arch_get_cpu_cycle_count() { return esp_cpu_get_cycle_count(); }
-uint32_t arch_get_cpu_freq_hz() {
-  uint32_t freq = 0;
-  esp_clk_tree_src_get_freq_hz(SOC_MOD_CLK_CPU, ESP_CLK_TREE_SRC_FREQ_PRECISION_CACHED, &freq);
-  return freq;
-}
 
 TaskHandle_t loop_task_handle = nullptr;  // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
 static StaticTask_t loop_task_tcb;        // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)

esphome/core/config.py

@@ -815,6 +815,16 @@ FILTER_SOURCE_FILES = filter_source_files_from_platform(
         "wake/wake_zephyr.cpp": {
             PlatformFramework.NRF52_ZEPHYR,
         },
+        # Per-platform HAL out-of-line implementations — hal.h dispatches to
+        # exactly one platform header based on USE_*, and the matching .cpp
+        # provides the few HAL functions that stay out-of-line. Other
+        # platforms still keep their out-of-line bodies in
+        # components/<platform>/core.cpp until their per-platform follow-up
+        # PR moves them here.
+        "hal/hal_esp32.cpp": {
+            PlatformFramework.ESP32_ARDUINO,
+            PlatformFramework.ESP32_IDF,
+        },
         # Note: lock_free_queue.h and event_pool.h are header files and don't need to be filtered
         # as they are only included when needed by the preprocessor
     }

esphome/core/hal.h

@@ -30,15 +30,11 @@
 
 namespace esphome {
 
-// ESP8266 inlines delayMicroseconds() and arch_feed_wdt() in hal/hal_esp8266.h;
-// every other platform keeps them out-of-line in components/<platform>/core.cpp.
-#ifndef USE_ESP8266
-void delayMicroseconds(uint32_t us);  // NOLINT(readability-identifier-naming)
-void arch_feed_wdt();
-#endif
+// Cross-platform declarations. delayMicroseconds(), arch_feed_wdt(),
+// arch_get_cpu_cycle_count() vary per platform (some inline, some
+// out-of-line) so they live in hal/hal_<platform>.h.
 void __attribute__((noreturn)) arch_restart();
 void arch_init();
-uint32_t arch_get_cpu_cycle_count();
 uint32_t arch_get_cpu_freq_hz();
 
 #ifndef USE_ESP8266

esphome/core/hal/hal_esp32.cpp

@@ -0,0 +1,44 @@
+#ifdef USE_ESP32
+
+#include "esphome/core/hal.h"
+
+#include <esp_clk_tree.h>
+#include <esp_system.h>
+#include <esp_timer.h>
+#include <freertos/FreeRTOS.h>
+#include <freertos/task.h>
+
+namespace esphome {
+
+// Use xTaskGetTickCount() when tick rate is 1 kHz (ESPHome's default via sdkconfig),
+// falling back to esp_timer for non-standard rates. IRAM_ATTR is required because
+// Wiegand and ZyAura call millis() from IRAM_ATTR ISR handlers on ESP32.
+// xTaskGetTickCountFromISR() is used in ISR context to satisfy the FreeRTOS API contract.
+uint32_t IRAM_ATTR HOT millis() {
+#if CONFIG_FREERTOS_HZ == 1000
+  if (xPortInIsrContext()) [[unlikely]] {
+    return xTaskGetTickCountFromISR();
+  }
+  return xTaskGetTickCount();
+#else
+  return micros_to_millis(static_cast<uint64_t>(esp_timer_get_time()));
+#endif
+}
+
+void arch_restart() {
+  esp_restart();
+  // restart() doesn't always end execution
+  while (true) {  // NOLINT(clang-diagnostic-unreachable-code)
+    yield();
+  }
+}
+
+uint32_t arch_get_cpu_freq_hz() {
+  uint32_t freq = 0;
+  esp_clk_tree_src_get_freq_hz(SOC_MOD_CLK_CPU, ESP_CLK_TREE_SRC_FREQ_PRECISION_CACHED, &freq);
+  return freq;
+}
+
+}  // namespace esphome
+
+#endif  // USE_ESP32

esphome/core/hal/hal_esp32.h

@@ -4,6 +4,8 @@
 
 #include <cstdint>
 #include <esp_attr.h>
+#include <esp_cpu.h>
+#include <esp_task_wdt.h>
 #include <freertos/FreeRTOS.h>
 #include <freertos/task.h>
 
@@ -15,6 +17,11 @@
 
 namespace esphome {
 
+// Forward decl from helpers.h (esphome/core/helpers.h) — kept here so this
+// header does not need to pull the rest of 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() { return xPortInIsrContext() != 0; }
 
@@ -30,6 +37,11 @@ __attribute__((always_inline)) inline uint64_t millis_64() {
   return micros_to_millis<uint64_t>(static_cast<uint64_t>(esp_timer_get_time()));
 }
 
+// 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() { esp_task_wdt_reset(); }
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_cycle_count() { return esp_cpu_get_cycle_count(); }
+
 }  // namespace esphome
 
 #endif  // USE_ESP32

esphome/core/hal/hal_host.h

@@ -19,6 +19,10 @@ 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();
+
 }  // namespace esphome
 
 #endif  // USE_HOST

esphome/core/hal/hal_libretiny.h

@@ -88,6 +88,10 @@ __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();
+
 }  // namespace esphome
 
 #endif  // USE_LIBRETINY

esphome/core/hal/hal_rp2040.h

@@ -35,6 +35,10 @@ __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();
+
 }  // namespace esphome
 
 #endif  // USE_RP2040

esphome/core/hal/hal_zephyr.h

@@ -19,6 +19,10 @@ 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();
+
 }  // namespace esphome
 
 #endif  // USE_ZEPHYR