fix(ocr): create TMPDIR on startup and clear day-old orphans

On a fresh ocr_cache volume /app/cache/.tmp does not exist yet. The mkdir
ensures the first Surya model download can proceed without ENOSPC on the
512 MB /tmp tmpfs. The find cleanup removes fragments left by docker-kill
mid-download, preventing cross-job ground-truth leakage.

Fixes #614. See ADR-021.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ocr-service/test_tmpdir.py

@@ -0,0 +1,97 @@
+"""Tests for TMPDIR configuration and entrypoint mkdir behavior — ADR-021."""
+
+import os
+import subprocess
+import tempfile
+
+import pytest
+
+try:
+    from fastapi import HTTPException
+    from main import _validate_zip_entry
+    HAS_MAIN = True
+except ImportError:
+    HAS_MAIN = False
+
+_ENTRYPOINT = os.path.join(os.path.dirname(os.path.abspath(__file__)), "entrypoint.sh")
+
+
+def test_tempfile_uses_tmpdir_when_set(monkeypatch, tmp_path):
+    """Python honours the TMPDIR env var when creating temporary directories.
+
+    Documents the mechanism that routes Surya model staging to the persistent
+    cache volume instead of the 512 MB RAM tmpfs. See ADR-021.
+    """
+    custom_tmp = tmp_path / "model_staging"
+    custom_tmp.mkdir()
+    monkeypatch.setenv("TMPDIR", str(custom_tmp))
+    monkeypatch.setattr(tempfile, "tempdir", None)
+    with tempfile.TemporaryDirectory() as td:
+        assert td.startswith(str(custom_tmp))
+
+
+def test_entrypoint_creates_tmpdir(tmp_path):
+    """entrypoint.sh creates the TMPDIR directory when it does not exist.
+
+    On a fresh ocr_cache volume, /app/cache/.tmp is absent. The entrypoint
+    must create it before uvicorn starts so the first Surya model download
+    does not exhaust the 512 MB /tmp tmpfs (ENOSPC). See ADR-021.
+    """
+    custom_tmp = tmp_path / "model-staging"
+    assert not custom_tmp.exists(), "pre-condition: directory must not exist yet"
+
+    stub_bin = tmp_path / "stub_bin"
+    stub_bin.mkdir()
+    for name in ("python3", "uvicorn"):
+        stub = stub_bin / name
+        stub.write_text("#!/bin/sh\nexit 0\n")
+        stub.chmod(0o755)
+
+    env = {
+        **os.environ,
+        "TMPDIR": str(custom_tmp),
+        "PATH": f"{stub_bin}:{os.environ.get('PATH', '/usr/bin:/bin')}",
+    }
+    result = subprocess.run(
+        ["bash", _ENTRYPOINT],
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+    assert custom_tmp.exists(), (
+        f"entrypoint.sh did not create TMPDIR={custom_tmp}\n"
+        f"stdout: {result.stdout}\nstderr: {result.stderr}"
+    )
+
+
+@pytest.mark.skipif(
+    not os.environ.get("TMPDIR", "").startswith("/app/cache"),
+    reason="TMPDIR contract only enforced inside the OCR Docker container",
+)
+def test_tmpdir_is_inside_persistent_cache_volume():
+    """TMPDIR must point to the persistent cache volume, not a RAM tmpfs.
+
+    Catches accidental reversion to /tmp or any tmpfs-backed path.
+    Runs only inside the OCR Docker container where TMPDIR=/app/cache/.tmp.
+    See ADR-021.
+    """
+    tmpdir = os.environ["TMPDIR"]
+    assert tmpdir.startswith("/app/cache"), (
+        f"TMPDIR={tmpdir!r} must be under /app/cache to route model downloads "
+        "to the SSD-backed cache volume — see ADR-021"
+    )
+
+
+@pytest.mark.skipif(not HAS_MAIN, reason="requires full ML stack (not available in CI)")
+def test_zipslip_still_anchors_under_custom_tmpdir(tmp_path):
+    """_validate_zip_entry rejects path-traversal when extract_dir is under a custom TMPDIR.
+
+    When TMPDIR=/app/cache/.tmp, extraction dirs live under that path.
+    Verifies os.path.realpath() still anchors correctly against the non-default base.
+    """
+    extract_dir = tmp_path / "model-staging" / "tmpXXX"
+    extract_dir.mkdir(parents=True)
+
+    with pytest.raises(HTTPException) as exc_info:
+        _validate_zip_entry("../evil.py", str(extract_dir))
+    assert exc_info.value.status_code == 400