diff --git a/esphome/components/esp8266/core.cpp b/esphome/components/esp8266/core.cpp
index 9161ca6aaf5..b9ad4082e98 100644
--- a/esphome/components/esp8266/core.cpp
+++ b/esphome/components/esp8266/core.cpp
@@ -3,98 +3,12 @@
 #include "core.h"
 #include "esphome/core/defines.h"
 #include "esphome/core/hal.h"
-#include "esphome/core/time_64.h"
-#include "esphome/core/helpers.h"
-#include "preferences.h"
 #include <Arduino.h>
-#include <core_esp8266_features.h>
-
-extern "C" {
-#include <user_interface.h>
-}
 
 namespace esphome {
 
-// yield(), micros(), millis_64() inlined in hal.h.
-// Fast accumulator replacement for Arduino's millis() (~3.3 μs via 4× 64-bit
-// multiplies on the LX106). Tracks a running ms counter from 32-bit
-// system_get_time() deltas using pure 32-bit ops. Installed as __wrap_millis
-// (via -Wl,--wrap=millis) so Arduino libs and IRAM_ATTR ISR handlers (e.g.
-// Wiegand, ZyAura) also get the fast version. xt_rsil(15) guards the static
-// state against ISR re-entry; the critical section is bounded (≤10 while-loop
-// iterations, ~100 ns on the common path, or a constant-time /1000 ~2.5 μs on
-// the rare path — well under WiFi's ~10 μs ISR latency budget). NMIs (level
-// >15) are not masked, but the ESP8266 SDK's NMI handlers don't call millis().
-//
-// system_get_time() wraps every ~71.6 min; unsigned (now_us - last_us) handles
-// one wrap. The main loop calls millis() at 60+ Hz, so delta stays tiny — a
-// >71 min block would trip the watchdog long before it could matter here.
-static constexpr uint32_t MILLIS_RARE_PATH_THRESHOLD_US = 10000;
-static constexpr uint32_t US_PER_MS = 1000;
-
-uint32_t IRAM_ATTR HOT millis() {
-  // Struct packs the three statics so the compiler loads one base address
-  // instead of three separate literal pool entries (saves ~8 bytes IRAM).
-  static struct {
-    uint32_t cache;
-    uint32_t remainder;
-    uint32_t last_us;
-  } state = {0, 0, 0};
-  uint32_t ps = xt_rsil(15);
-  uint32_t now_us = system_get_time();
-  uint32_t delta = now_us - state.last_us;
-  state.last_us = now_us;
-  state.remainder += delta;
-  if (state.remainder >= MILLIS_RARE_PATH_THRESHOLD_US) {
-    // Rare path: large gap (WiFi scan, boot, long block). Constant-time
-    // conversion keeps the critical section bounded.
-    uint32_t ms = state.remainder / US_PER_MS;
-    state.cache += ms;
-    // Reuse ms instead of `remainder %= US_PER_MS` — `%` would compile to a
-    // second __umodsi3 call on the LX106 (no hardware divide).
-    state.remainder -= ms * US_PER_MS;
-  } else {
-    // Common path: small gap. At most ~10 iterations since remainder was
-    // < threshold (10 ms) on entry and delta adds at most one more threshold
-    // before exiting this branch.
-    while (state.remainder >= US_PER_MS) {
-      state.cache++;
-      state.remainder -= US_PER_MS;
-    }
-  }
-  uint32_t result = state.cache;
-  xt_wsr_ps(ps);
-  return result;
-}
-// Poll-based delay that avoids ::delay() — Arduino's __delay has an intra-object
-// call to the original millis() that --wrap can't intercept, so calling ::delay()
-// would keep the slow Arduino millis body alive in IRAM. optimistic_yield still
-// enters esp_schedule()/esp_suspend_within_cont() via yield(), so SDK tasks and
-// WiFi run correctly. Theoretically less power-efficient than Arduino's
-// os_timer-based delay() for long waits, but nearly all ESPHome delays are short
-// (sensor/I²C/SPI settling in the 1–100 ms range) where the difference is
-// negligible.
-void HOT delay(uint32_t ms) {
-  if (ms == 0) {
-    optimistic_yield(1000);
-    return;
-  }
-  uint32_t start = millis();
-  while (millis() - start < ms) {
-    optimistic_yield(1000);
-  }
-}
-// delayMicroseconds(), arch_feed_wdt(), and progmem_read_*() are inlined in hal/hal_esp8266.h.
-void arch_restart() {
-  system_restart();
-  // restart() doesn't always end execution
-  while (true) {  // NOLINT(clang-diagnostic-unreachable-code)
-    yield();
-  }
-}
-void arch_init() {}
-uint32_t IRAM_ATTR HOT arch_get_cpu_cycle_count() { return esp_get_cycle_count(); }
-uint32_t arch_get_cpu_freq_hz() { return F_CPU; }
+// HAL functions live in hal.cpp. This file keeps only the ESP8266-specific
+// firmware bootstrap (Tasmota OTA magic bytes, optional GPIO pre-init).
 
 void force_link_symbols() {
   // Tasmota uses magic bytes in the binary to check if an OTA firmware is compatible
@@ -131,12 +45,4 @@ extern "C" void resetPins() {  // NOLINT
 
 }  // namespace esphome
 
-// Linker wrap: redirect all ::millis() calls (Arduino libs, ISRs) to our accumulator.
-// Requires -Wl,--wrap=millis in build flags (added by __init__.py).
-// NOLINTNEXTLINE(bugprone-reserved-identifier,cert-dcl37-c,cert-dcl51-cpp,readability-identifier-naming)
-extern "C" uint32_t IRAM_ATTR __wrap_millis() { return esphome::millis(); }
-// Note: Arduino's init() registers a 60-second overflow timer for micros64().
-// We leave it running — wrapping init() as a no-op would break micros64()'s
-// overflow tracking, and the timer's cost is negligible (~3 μs per 60 s).
-
 #endif  // USE_ESP8266
diff --git a/esphome/components/esp8266/hal.cpp b/esphome/components/esp8266/hal.cpp
new file mode 100644
index 00000000000..ef93a95273f
--- /dev/null
+++ b/esphome/components/esp8266/hal.cpp
@@ -0,0 +1,110 @@
+#ifdef USE_ESP8266
+
+#include "esphome/core/hal.h"
+
+#include <Arduino.h>
+#include <core_esp8266_features.h>
+
+extern "C" {
+#include <user_interface.h>
+}
+
+// Empty esp8266 namespace block to satisfy ci-custom's lint_namespace check.
+// HAL functions live in namespace esphome (root) — they are not part of the
+// esp8266 component's API.
+namespace esphome::esp8266 {}  // namespace esphome::esp8266
+
+namespace esphome {
+
+// yield(), micros(), millis_64(), delayMicroseconds(), arch_feed_wdt(),
+// progmem_read_*() are inlined in core/hal/hal_esp8266.h.
+//
+// Fast accumulator replacement for Arduino's millis() (~3.3 μs via 4× 64-bit
+// multiplies on the LX106). Tracks a running ms counter from 32-bit
+// system_get_time() deltas using pure 32-bit ops. Installed as __wrap_millis
+// (via -Wl,--wrap=millis) so Arduino libs and IRAM_ATTR ISR handlers (e.g.
+// Wiegand, ZyAura) also get the fast version. xt_rsil(15) guards the static
+// state against ISR re-entry; the critical section is bounded (≤10 while-loop
+// iterations, ~100 ns on the common path, or a constant-time /1000 ~2.5 μs on
+// the rare path — well under WiFi's ~10 μs ISR latency budget). NMIs (level
+// >15) are not masked, but the ESP8266 SDK's NMI handlers don't call millis().
+//
+// system_get_time() wraps every ~71.6 min; unsigned (now_us - last_us) handles
+// one wrap. The main loop calls millis() at 60+ Hz, so delta stays tiny — a
+// >71 min block would trip the watchdog long before it could matter here.
+static constexpr uint32_t MILLIS_RARE_PATH_THRESHOLD_US = 10000;
+static constexpr uint32_t US_PER_MS = 1000;
+
+uint32_t IRAM_ATTR HOT millis() {
+  // Struct packs the three statics so the compiler loads one base address
+  // instead of three separate literal pool entries (saves ~8 bytes IRAM).
+  static struct {
+    uint32_t cache;
+    uint32_t remainder;
+    uint32_t last_us;
+  } state = {0, 0, 0};
+  uint32_t ps = xt_rsil(15);
+  uint32_t now_us = system_get_time();
+  uint32_t delta = now_us - state.last_us;
+  state.last_us = now_us;
+  state.remainder += delta;
+  if (state.remainder >= MILLIS_RARE_PATH_THRESHOLD_US) {
+    // Rare path: large gap (WiFi scan, boot, long block). Constant-time
+    // conversion keeps the critical section bounded.
+    uint32_t ms = state.remainder / US_PER_MS;
+    state.cache += ms;
+    // Reuse ms instead of `remainder %= US_PER_MS` — `%` would compile to a
+    // second __umodsi3 call on the LX106 (no hardware divide).
+    state.remainder -= ms * US_PER_MS;
+  } else {
+    // Common path: small gap. At most ~10 iterations since remainder was
+    // < threshold (10 ms) on entry and delta adds at most one more threshold
+    // before exiting this branch.
+    while (state.remainder >= US_PER_MS) {
+      state.cache++;
+      state.remainder -= US_PER_MS;
+    }
+  }
+  uint32_t result = state.cache;
+  xt_wsr_ps(ps);
+  return result;
+}
+
+// Poll-based delay that avoids ::delay() — Arduino's __delay has an intra-object
+// call to the original millis() that --wrap can't intercept, so calling ::delay()
+// would keep the slow Arduino millis body alive in IRAM. optimistic_yield still
+// enters esp_schedule()/esp_suspend_within_cont() via yield(), so SDK tasks and
+// WiFi run correctly. Theoretically less power-efficient than Arduino's
+// os_timer-based delay() for long waits, but nearly all ESPHome delays are short
+// (sensor/I²C/SPI settling in the 1–100 ms range) where the difference is
+// negligible.
+void HOT delay(uint32_t ms) {
+  if (ms == 0) {
+    optimistic_yield(1000);
+    return;
+  }
+  uint32_t start = millis();
+  while (millis() - start < ms) {
+    optimistic_yield(1000);
+  }
+}
+
+void arch_restart() {
+  system_restart();
+  // restart() doesn't always end execution
+  while (true) {  // NOLINT(clang-diagnostic-unreachable-code)
+    yield();
+  }
+}
+
+}  // namespace esphome
+
+// Linker wrap: redirect all ::millis() calls (Arduino libs, ISRs) to our accumulator.
+// Requires -Wl,--wrap=millis in build flags (added by __init__.py).
+// NOLINTNEXTLINE(bugprone-reserved-identifier,cert-dcl37-c,cert-dcl51-cpp,readability-identifier-naming)
+extern "C" uint32_t IRAM_ATTR __wrap_millis() { return esphome::millis(); }
+// Note: Arduino's init() registers a 60-second overflow timer for micros64().
+// We leave it running — wrapping init() as a no-op would break micros64()'s
+// overflow tracking, and the timer's cost is negligible (~3 μs per 60 s).
+
+#endif  // USE_ESP8266
diff --git a/esphome/core/hal.h b/esphome/core/hal.h
index 312effa7b0e..a53296979c6 100644
--- a/esphome/core/hal.h
+++ b/esphome/core/hal.h
@@ -31,11 +31,10 @@
 namespace esphome {
 
 // 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.
+// arch_get_cpu_cycle_count(), arch_init(), arch_get_cpu_freq_hz() 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_freq_hz();
 
 #ifndef USE_ESP8266
 // All non-ESP8266 platforms: PROGMEM is a no-op, so these are direct dereferences.
diff --git a/esphome/core/hal/hal_esp32.h b/esphome/core/hal/hal_esp32.h
index 2bff4244419..d5d7752bf6a 100644
--- a/esphome/core/hal/hal_esp32.h
+++ b/esphome/core/hal/hal_esp32.h
@@ -42,6 +42,9 @@ __attribute__((always_inline)) inline void delayMicroseconds(uint32_t us) { dela
 __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(); }
 
+void arch_init();
+uint32_t arch_get_cpu_freq_hz();
+
 }  // namespace esphome
 
 #endif  // USE_ESP32
diff --git a/esphome/core/hal/hal_esp8266.h b/esphome/core/hal/hal_esp8266.h
index 484118f1f50..b6e3b1ee3cb 100644
--- a/esphome/core/hal/hal_esp8266.h
+++ b/esphome/core/hal/hal_esp8266.h
@@ -3,6 +3,7 @@
 #ifdef USE_ESP8266
 
 #include <c_types.h>
+#include <core_esp8266_features.h>
 #include <cstdint>
 #include <pgmspace.h>
 
@@ -59,8 +60,11 @@ __attribute__((always_inline)) inline uint16_t progmem_read_uint16(const uint16_
 // 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() { system_soft_wdt_feed(); }
-
-uint32_t arch_get_cpu_cycle_count();
+__attribute__((always_inline)) inline void arch_init() {}
+// esp_get_cycle_count() declared in <core_esp8266_features.h>; F_CPU is a
+// compiler-driven macro from the ESP8266 Arduino board defs (-DF_CPU=...).
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_cycle_count() { return esp_get_cycle_count(); }
+__attribute__((always_inline)) inline uint32_t arch_get_cpu_freq_hz() { return F_CPU; }
 
 }  // namespace esphome
 
diff --git a/esphome/core/hal/hal_host.h b/esphome/core/hal/hal_host.h
index 682a1a422b4..a8896fdf638 100644
--- a/esphome/core/hal/hal_host.h
+++ b/esphome/core/hal/hal_host.h
@@ -22,6 +22,8 @@ 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();
 
 }  // namespace esphome
 
diff --git a/esphome/core/hal/hal_libretiny.h b/esphome/core/hal/hal_libretiny.h
index e9d33b7753b..ecfe830fe35 100644
--- a/esphome/core/hal/hal_libretiny.h
+++ b/esphome/core/hal/hal_libretiny.h
@@ -91,6 +91,8 @@ __attribute__((always_inline)) inline uint64_t millis_64() { return Millis64Impl
 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
 
diff --git a/esphome/core/hal/hal_rp2040.h b/esphome/core/hal/hal_rp2040.h
index 2a1d67b4a32..46f6e421cd4 100644
--- a/esphome/core/hal/hal_rp2040.h
+++ b/esphome/core/hal/hal_rp2040.h
@@ -38,6 +38,8 @@ __attribute__((always_inline)) inline uint64_t millis_64() { return micros_to_mi
 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
 
diff --git a/esphome/core/hal/hal_zephyr.h b/esphome/core/hal/hal_zephyr.h
index 6707c85b2c2..d4b37b5eb6c 100644
--- a/esphome/core/hal/hal_zephyr.h
+++ b/esphome/core/hal/hal_zephyr.h
@@ -22,6 +22,8 @@ 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();
 
 }  // namespace esphome