diff --git a/esphome/build_gen/arduino8266.py b/esphome/build_gen/arduino8266.py
index 645df0131c..c6a6ca7855 100644
--- a/esphome/build_gen/arduino8266.py
+++ b/esphome/build_gen/arduino8266.py
@@ -131,6 +131,9 @@ _MMU_VARIANTS = (
 )
 _MMU_DEFAULT = ("MMU_IRAM_SIZE=0x8000", "MMU_ICACHE_SIZE=0x8000")
 
+# Upstream's CXXFLAGS (-fno-rtti, the -std level, -f(no-)exceptions) and the
+# trailing stdc++/m/c/gcc system libs are composed at emission
+# (write_project) from CORE.cpp_standard and _BuildConfig.exceptions.
 _ASFLAGS = ["-mlongcalls", "-mtext-section-literals"]
 _CFLAGS = [
     "-std=gnu17",
@@ -208,7 +211,10 @@ class _BuildConfig:
 def _flag_defines() -> dict[str, str]:
     """Map define name -> full ``NAME[=VALUE]`` for every -D build flag."""
     defines: dict[str, str] = {}
-    for flag in CORE.build_flags:
+    # Sorted so duplicate defines resolve the same way every run: the
+    # winner feeds the linker-script preprocessor line, which is also the
+    # cache stamp
+    for flag in sorted(CORE.build_flags):
         # Shell-lex every entry the way PlatformIO's ParseFlags does, so a
         # knob in "-DKNOB -DOTHER", a spaced "-D KNOB", and quoted bodies all
         # read identically to _project_flags (and the compile line).
@@ -414,9 +420,15 @@ def generate_ld_scripts(
         and stamp.is_file()
         and stamp.read_text(encoding="utf-8") == stamp_content
     ):
-        result = subprocess.run(
-            cmd, capture_output=True, text=True, check=False, close_fds=False
-        )
+        try:
+            result = subprocess.run(
+                cmd, capture_output=True, text=True, check=False, close_fds=False
+            )
+        except OSError as err:
+            # A half-extracted or half-deleted toolchain cache reaches here
+            raise EsphomeError(
+                f"Could not run {gcc}: {err}; run 'esphome clean-all' and retry"
+            ) from err
         if result.returncode != 0:
             raise EsphomeError(f"Generating the linker script failed:\n{result.stderr}")
         content = build_surgery.relocate_ratetable(result.stdout)
diff --git a/tests/unit_tests/build_gen/test_arduino8266.py b/tests/unit_tests/build_gen/test_arduino8266.py
index 244f00cbda..8d9894e3ce 100644
--- a/tests/unit_tests/build_gen/test_arduino8266.py
+++ b/tests/unit_tests/build_gen/test_arduino8266.py
@@ -727,3 +727,48 @@ def test_flag_defines_lexes_quoted_single_tokens() -> None:
     """A quoted single-token define reads the same as on the compile line."""
     _set_flags('-DMMU_SEC_HEAP="0x40108000"')
     assert _flag_defines()["MMU_SEC_HEAP"] == "MMU_SEC_HEAP=0x40108000"
+
+
+def test_flag_defines_duplicate_defines_resolve_deterministically() -> None:
+    """Duplicate conflicting defines pick the same winner every run (sorted
+    iteration, last writer wins), independent of the set's hash seed."""
+    _set_flags("-DMMU_IRAM_SIZE=0x8000", "-DMMU_IRAM_SIZE=0xC000")
+    assert _flag_defines()["MMU_IRAM_SIZE"] == "MMU_IRAM_SIZE=0xC000"
+
+
+def test_flag_tables_match_platformio_builder() -> None:
+    """The transliterated flag lists pinned verbatim, like the define set:
+    a drift lands as a test failure, not a binary-size regression."""
+    assert arduino8266._ASFLAGS == ["-mlongcalls", "-mtext-section-literals"]
+    assert arduino8266._CFLAGS == [
+        "-std=gnu17",
+        "-Wpointer-arith",
+        "-Wno-implicit-function-declaration",
+        "-Wl,-EL",
+        "-fno-inline-functions",
+        "-nostdlib",
+    ]
+    assert arduino8266._CCFLAGS[:6] == [
+        "-Os",
+        "-mlongcalls",
+        "-mtext-section-literals",
+        "-falign-functions=4",
+        "-U__STRICT_ANSI__",
+        "-ffunction-sections",
+    ]
+    assert arduino8266._LINKFLAGS[:5] == [
+        "-Os",
+        "-nostdlib",
+        "-Wl,--no-check-sections",
+        "-Wl,-static",
+        "-Wl,--gc-sections",
+    ]
+
+
+def test_generate_ld_scripts_missing_compiler_is_clean(tmp_path: Path) -> None:
+    """A half-deleted toolchain cache fails with an ESPHome error naming the
+    binary, not a FileNotFoundError traceback."""
+    paths = _make_framework(tmp_path)
+    _set_flags()
+    with pytest.raises(EsphomeError, match="Could not run"):
+        _run_generate_ld_scripts(paths)