claude review on linux build

NeptuneHub · NeptuneHub · commit f8ebaa98ba1a · 2026-06-05T23:03:44.000+02:00
diff --git a/.github/workflows/build-linux.yml b/.github/workflows/build-linux.yml
@@ -215,7 +215,12 @@ jobs:
           )
           missing=0
           for f in "${required[@]}"; do
-            if [ ! -e "$f" ]; then echo "::error::Missing or empty: $f"; missing=1; fi
+            # Flag a file that is missing OR a zero-byte/truncated download. The
+            # size test only applies to regular files (directory entries in the
+            # list pass on existence alone).
+            if [ ! -e "$f" ] || { [ -f "$f" ] && [ ! -s "$f" ]; }; then
+              echo "::error::Missing or empty: $f"; missing=1
+            fi
           done
           if [ -z "$(find model/huggingface/hub/models--roberta-base -name tokenizer.json -print -quit)" ]; then
             echo "::error::roberta-base tokenizer.json missing after HF-cache prune"; missing=1
diff --git a/.github/workflows/pr-test-link.yml b/.github/workflows/pr-test-link.yml
@@ -149,6 +149,18 @@ jobs:
             // Upsert the managed block, leaving the author's text untouched.
             const { data: pr } = await github.rest.pulls.get({ owner, repo, pull_number });
             let body = pr.body || '';
+
+            // Strip any legacy block left by the old in-build step (it used a
+            // different marker), so renaming the markers does not orphan a stale,
+            // never-updated block on in-flight PRs.
+            const LEGACY_START = '<!-- macos-test-build:start -->';
+            const LEGACY_END = '<!-- macos-test-build:end -->';
+            const ls = body.indexOf(LEGACY_START);
+            const le = body.indexOf(LEGACY_END);
+            if (ls !== -1 && le !== -1 && le > ls) {
+              body = (body.slice(0, ls) + body.slice(le + LEGACY_END.length)).replace(/\s+$/, '');
+            }
+
             const s = body.indexOf(START);
             const e = body.indexOf(END);
             if (s !== -1 && e !== -1 && e > s) {
diff --git a/linux/build.sh b/linux/build.sh
@@ -95,31 +95,29 @@ if [ "$UNAME_ARCH" != "x86_64" ]; then
     echo "::error::Expected bundled PostgreSQL (pgsql/bin) in $BUNDLE (aarch64 build)"; exit 1
   fi
 else
-  # x86_64: repair the bundled pgserver (pgserver wheel) PostgreSQL tree, which
-  # PyInstaller damages in two independent ways -- both confirmed to break
-  # startup, and both fixed by overlaying the pristine wheel tree below:
+  # x86_64: repair the bundled pgserver (pgserver wheel) PostgreSQL tree.
   #
-  #   1. STRIPPED EXECUTABLES. The spec sets strip=True to shrink the unstripped
-  #      scipy/numpy/onnx/PyAV wheels (hundreds of MB). But pgserver's binaries
-  #      were already processed by auditwheel/patchelf (mangled libpq soname +
-  #      injected RPATH); running GNU `strip` over a patchelf-modified ELF
-  #      corrupts it, so the libpq-linked executables (initdb, psql, pg_dump,
-  #      pg_restore) SIGSEGV at load time -- initdb dies before it can create
-  #      the cluster and the whole supervisor startup fails.
+  # The spec pulls the tree in via collect_data_files('pgserver'), but that
+  # helper EXCLUDES shared libraries, so every loadable module under
+  # pginstall/lib/postgresql is dropped: plpgsql.so, vector.so (pgvector),
+  # dict_snowball.so (initdb's post-bootstrap text-search setup loads it ->
+  # initdb fails without it), pgoutput, the encoding converters, etc. Without
+  # them initdb cannot create the cluster and supervisor startup fails.
   #
-  #   2. MISSING LOADABLE MODULES. The spec pulls the tree in via
-  #      collect_data_files('pgserver'), but that helper excludes shared
-  #      libraries, so every loadable module under pginstall/lib/postgresql is
-  #      dropped: plpgsql.so, vector.so (pgvector), dict_snowball.so (initdb's
-  #      post-bootstrap text-search setup loads it -> initdb fails even once it
-  #      no longer segfaults), pgoutput, the encoding converters, etc.
+  # NOTE: the executables themselves are fine -- the spec sets strip=False
+  # (stripping is disabled on Linux precisely because it corrupts pgserver's
+  # patchelf-modified ELFs; see AudioMuse-AI.spec), and collect_data_files
+  # copies binaries verbatim, so the only thing actually missing is the .so
+  # modules above. Do NOT re-enable strip=True to "shrink" the bundle: that is
+  # what would corrupt initdb/psql/pg_dump and make them SIGSEGV at load.
   #
   # Overlay the COMPLETE, pristine pginstall tree from the installed wheel onto
-  # the bundle. cp merges: it overwrites the corrupted executables and adds the
-  # missing .so modules, while leaving the vendored unaccent/pg_trgm contrib the
-  # spec grafted in (those are not in the wheel). Also refresh the external libs
-  # (pgserver.libs/). Must run inside the build venv where pgserver is
-  # importable (the CI step and build.sh's own usage both activate it).
+  # the bundle. cp merges: it adds the missing .so modules (and harmlessly
+  # re-copies the already-correct executables), while leaving the vendored
+  # unaccent/pg_trgm contrib the spec grafted in (those are not in the wheel).
+  # Also refresh the external libs (pgserver.libs/). Must run inside the build
+  # venv where pgserver is importable (the CI step and build.sh's own usage both
+  # activate it).
   PG_PKG="$(python -c 'import os, pgserver; print(os.path.dirname(os.path.abspath(pgserver.__file__)))')"
   PG_SITE="$(dirname "$PG_PKG")"
   DST_PGINSTALL="$BUNDLE/_internal/pgserver/pginstall"
diff --git a/linux/embedded_pg.py b/linux/embedded_pg.py
@@ -18,6 +18,7 @@
 
 import logging
 import os
+import shutil
 import subprocess
 import threading
 
@@ -28,6 +29,12 @@
 _lock = threading.RLock()
 _data_dir = None  # remembered so stop() can target the right cluster
 
+# Written only after initdb AND our config edits both succeed, so its presence
+# proves a *complete* cluster. PG_VERSION alone is not enough: initdb writes it
+# partway through, so a crash/SIGKILL mid-initdb leaves a half-built dir that
+# looks initialized forever and wedges every later start.
+_READY_MARKER = "audiomuse_initialized"
+
 
 def _pg_env():
     """Environment for the bundled Postgres tools.
@@ -56,8 +63,41 @@ def _bin(name):
 
 
 def _initialized(data_dir):
-    # PG_VERSION exists only after a successful initdb.
-    return os.path.exists(os.path.join(data_dir, "PG_VERSION"))
+    if os.path.exists(os.path.join(data_dir, _READY_MARKER)):
+        return True
+    # Legacy clusters created before the completion marker existed: adopt them if
+    # initdb actually finished. global/pg_control is written by initdb and is
+    # required for the server to start, so treat it as the completeness signal,
+    # then stamp the marker so future starts take the fast path.
+    if os.path.exists(os.path.join(data_dir, "global", "pg_control")):
+        try:
+            with open(os.path.join(data_dir, _READY_MARKER), "w", encoding="utf-8") as fh:
+                fh.write("ok\n")
+        except OSError:
+            pass
+        return True
+    return False
+
+
+def _reset_data_dir(data_dir):
+    """Empty a partially-initialized data dir so initdb can retry.
+
+    initdb refuses a non-empty target, so a half-built cluster (interrupted
+    initdb, no completion marker) would otherwise wedge every start forever.
+    Only ever called when :func:`_initialized` is False, so a complete cluster's
+    data is never touched."""
+    if not (os.path.isdir(data_dir) and os.listdir(data_dir)):
+        return
+    logger.warning("Clearing incomplete PostgreSQL data dir %s before re-init", data_dir)
+    for entry in os.listdir(data_dir):
+        target = os.path.join(data_dir, entry)
+        try:
+            if os.path.isdir(target) and not os.path.islink(target):
+                shutil.rmtree(target)
+            else:
+                os.unlink(target)
+        except OSError:
+            logger.exception("Could not remove %s", target)
 
 
 def _dsn(data_dir):
@@ -100,6 +140,11 @@ def _init_cluster(data_dir, env):
         fh.write("\n# --- AudioMuse-AI embedded overrides ---\n")
         fh.write(f"unix_socket_directories = '{data_dir}'\n")
         fh.write("listen_addresses = ''\n")
+    # Stamp the completion marker last: only now is the cluster fully usable, so
+    # an interruption before this point leaves _initialized() False and triggers
+    # a clean re-init on the next start.
+    with open(os.path.join(data_dir, _READY_MARKER), "w", encoding="utf-8") as fh:
+        fh.write("ok\n")
 
 
 def start(data_dir):
@@ -109,6 +154,7 @@ def start(data_dir):
         env = _pg_env()
         if not _initialized(data_dir):
             logger.info("Initializing embedded PostgreSQL cluster at %s", data_dir)
+            _reset_data_dir(data_dir)  # clear any half-built cluster first
             _init_cluster(data_dir, env)
         if not _is_running(data_dir, env):
             _run_checked([_bin("pg_ctl"), "-D", data_dir, "-w", "start"], env)
diff --git a/linux/launcher.py b/linux/launcher.py
@@ -173,17 +173,19 @@ def _handle_signal(signum, _frame):
     signal.signal(signal.SIGINT, _handle_signal)
     signal.signal(signal.SIGTERM, _handle_signal)
 
-    def _boot():
-        try:
-            supervisor.start_all()
-            print("AudioMuse-AI is running at %s" % WEB_URL)
-            if open_browser:
-                _open_browser()
-        except Exception as exc:  # startup already logged; surface a hint too
-            print("AudioMuse-AI failed to start: %s" % exc, file=sys.stderr)
-            stop_event.set()
-
-    threading.Thread(target=_boot, name="boot", daemon=True).start()
+    def _on_ready():
+        print("AudioMuse-AI is running at %s" % WEB_URL)
+        if open_browser:
+            _open_browser()
+
+    def _on_error(exc):  # startup already logged; surface a hint too
+        print("AudioMuse-AI failed to start: %s" % exc, file=sys.stderr)
+        stop_event.set()
+
+    # The supervisor owns the boot thread so stop_all() can join it before
+    # tearing down -- a SIGTERM racing an in-progress boot must not leave the
+    # boot thread spawning children after the teardown sweep.
+    supervisor.start_in_background(on_ready=_on_ready, on_error=_on_error)
 
     try:
         while not stop_event.wait(0.5):
@@ -230,8 +232,11 @@ def _cmd_open():
     from linux import paths
     if _running_supervisor_pid(paths) is None:
         # Not running yet: start it in the foreground (this call becomes the
-        # supervisor and opens the browser once it is up).
-        return _run_supervisor(open_browser=True)
+        # supervisor and opens the browser once it is up). Honor
+        # AUDIOMUSE_OPEN_BROWSER like `start` does, so a headless/service
+        # invocation does not try to spawn a browser.
+        open_browser = os.environ.get("AUDIOMUSE_OPEN_BROWSER", "1") != "0"
+        return _run_supervisor(open_browser=open_browser)
     _open_browser()
     return 0
 
diff --git a/linux/packaging/audiomuse-ai.service b/linux/packaging/audiomuse-ai.service
@@ -1,7 +1,6 @@
 [Unit]
 Description=AudioMuse-AI (standalone music analysis + smart playlists)
 Documentation=https://github.com/NeptuneHub/AudioMuse-AI
-After=default.target
 
 [Service]
 Type=simple
diff --git a/linux/supervisor.py b/linux/supervisor.py
