Retry a network source whose session dropped, not just failing media

CHANGELOG.md

@@ -27,6 +27,17 @@ project uses [semantic versioning][semver].
 
 ### Fixed
 
+- **A source on a network share survives its session dropping.** The retry
+  machinery reopened a dead handle and resumed from the last delivered chunk
+  already; the classifier just did not recognise a dropped SMB session as
+  transient, so the retry never fired and the file failed outright. The
+  network-mount codes now qualify — including `ERROR_UNEXP_NET_ERR`, which
+  Explorer reports as **0x8007003B** before abandoning the whole transfer, and
+  `ECONNRESET`, `ENETRESET`, `EPIPE` and `ESTALE` on a POSIX mount. Observed
+  offloading from a NAS over a VPN link: the path pings clean either side of
+  the drop, which is exactly why the retry is worth making. Recovery costs one
+  re-read of the chunk in flight.
+
 - **A decoder ffmpeg lacks is probed once per job, not per clip.** Extracting
   thumbnails from BRAW with a stock ffmpeg fails identically for every clip;
   each one still paid four doomed process spawns. The first clip of a suffix

ROADMAP.md

@@ -33,6 +33,10 @@ not only the file hashes — so a rename or a moved file, which every file hash
 agrees is fine, is reported as the structure-hash mismatch it is. See
 [`docs/ascmhl.md`](docs/ascmhl.md#directory-hashes).
 
+A source on a network mount is handled like marginal media, because it fails
+like it: the dropped-session error codes are retried, the handle reopened and
+the read resumed from the last delivered chunk.
+
 `--paranoid` reads every source file a second time and compares, which is the
 only thing that catches a read returning wrong bytes without reporting an error.
 Retry works at the chunk that failed rather than restarting the file. Sidecars

docs/data-safety.md

@@ -202,7 +202,7 @@ known.
 - **Concurrent instances.** One app instance serialises its queue. Two instances
   pointed at the same destination are not coordinated.
 
-## Marginal media
+## Marginal media and dropped links
 
 Cards and readers fail intermittently long before they fail for good. A read
 that fails for a transient-looking reason is retried — three attempts by
@@ -216,6 +216,20 @@ delay, so only errors with a plausible transient cause qualify: `EIO`, `EBUSY`,
 `ERROR_SHARING_VIOLATION` (usually antivirus, usually brief) and
 `ERROR_IO_DEVICE`. `ENOENT` and `ENOSPC` fail immediately.
 
+A source on a network share qualifies for the same treatment, because it fails
+the same way: an SMB session over a VPN or a WAN link dies mid-transfer and
+takes the open handle with it, while the path is perfectly reachable either side
+of the drop. So the network-mount codes are retried too — `ERROR_BAD_NETPATH`,
+`ERROR_NETWORK_BUSY`, `ERROR_BAD_NET_RESP`, `ERROR_UNEXP_NET_ERR`,
+`ERROR_NETNAME_DELETED`, `ERROR_REQ_NOT_ACCEP`, the two unreachables and
+`ERROR_NO_SYSTEM_RESOURCES`, plus `ECONNRESET`, `ENETRESET`, `EPIPE` and
+`ESTALE` on a POSIX mount. A read proves the path resolved a moment earlier, so
+mid-copy these are a session that died rather than a name that was never right.
+
+`ERROR_UNEXP_NET_ERR` is the one Explorer reports as **0x8007003B**, "an
+unexpected network error occurred", before abandoning the entire transfer. Here
+it costs one re-read of the chunk in flight.
+
 A failed *read* is retried at the chunk it failed on, not by restarting the
 file. A chunk is only hashed once it has arrived whole, so a read that failed
 produced no checksum state to unwind — recovering a bad sector near the end of a

src/offloader/retry.py

@@ -5,6 +5,12 @@
 antivirus holds a handle open for a beat. robocopy has retried this way for
 decades (`/R`, `/W`) and it is the main thing it does that this engine did not.
 
+A network source fails the same way for different reasons. An SMB session over a
+VPN or a WAN link dies mid-transfer and the handle with it, while the path stays
+perfectly reachable either side of the drop — the case Explorer's copy engine
+gives up on and robocopy rides out. Those codes are here for the same reason the
+media ones are: the fault is in the link, not in the data.
+
 The discrimination matters more than the retrying. Retrying a missing file, a
 permission denial or a full disk wastes time and hides the real problem, so only
 errors with a plausible transient cause are retried, and a file that needed one
@@ -31,6 +37,16 @@
     errno.ETIMEDOUT,
     errno.ENODEV,     # device dropped off the bus and may come back
     errno.ENXIO,
+    # A network mount that dropped its session. Only ever consulted on POSIX:
+    # `is_transient` short-circuits on winerror, which Windows always sets.
+    errno.ECONNRESET,
+    errno.ECONNABORTED,
+    errno.ENETRESET,
+    errno.ENETUNREACH,
+    errno.EHOSTUNREACH,
+    errno.ENOTCONN,
+    errno.EPIPE,
+    errno.ESTALE,     # the NFS classic: handle outlived the thing it named
 }
 
 #: Windows error codes worth a second attempt. Python surfaces these on
@@ -40,11 +56,22 @@
     23,    # ERROR_CRC — "Data error (cyclic redundancy check)"
     32,    # ERROR_SHARING_VIOLATION — usually antivirus, and usually brief
     33,    # ERROR_LOCK_VIOLATION
+    # The network-mount family. A read mid-copy proves the path resolved a
+    # moment ago, so these are a session that died under us rather than a name
+    # that was never right — the distinction that makes them worth retrying.
+    53,    # ERROR_BAD_NETPATH
+    54,    # ERROR_NETWORK_BUSY
+    58,    # ERROR_BAD_NET_RESP
+    59,    # ERROR_UNEXP_NET_ERR — what Explorer surfaces as 0x8007003B
     64,    # ERROR_NETNAME_DELETED — network destination blipped
+    71,    # ERROR_REQ_NOT_ACCEP — server at its connection limit
     121,   # ERROR_SEM_TIMEOUT
     170,   # ERROR_BUSY
     1117,  # ERROR_IO_DEVICE
     1167,  # ERROR_DEVICE_NOT_CONNECTED
+    1231,  # ERROR_NETWORK_UNREACHABLE
+    1232,  # ERROR_HOST_UNREACHABLE
+    1450,  # ERROR_NO_SYSTEM_RESOURCES — seen on sustained large SMB reads
 }
 
 class UnstableRead(OSError):

tests/test_retry.py

@@ -54,6 +54,56 @@ def test_windows_error_codes(winerror: int, expected: bool):
     assert retry.is_transient(_os_error(errno.EIO, winerror=winerror)) is expected
 
 
+@pytest.mark.parametrize("winerror", [53, 54, 58, 59, 64, 71, 1231, 1232, 1450])
+def test_dropped_network_session_is_retried(winerror: int):
+    """A source on an SMB share over a VPN loses its session mid-copy and the
+    handle with it. 59 (ERROR_UNEXP_NET_ERR) is the one Explorer reports as
+    0x8007003B before abandoning the whole transfer; the engine reopens and
+    resumes from the last delivered chunk instead."""
+    assert retry.is_transient(_os_error(errno.EIO, winerror=winerror))
+
+
+@pytest.mark.parametrize("code", [errno.ECONNRESET, errno.ENETRESET,
+                                  errno.EHOSTUNREACH, errno.ENOTCONN,
+                                  errno.EPIPE, errno.ESTALE])
+def test_posix_network_mount_errors_are_retried(code: int):
+    """The same drop seen through a POSIX mount, where there is no winerror to
+    short-circuit on."""
+    assert retry.is_transient(_os_error(code))
+
+
+def test_a_windows_error_code_decides_over_the_mapped_errno():
+    """Python maps a winerror onto whichever errno it thinks fits, and for a
+    dropped session that mapping can land on something in the permanent set.
+    The winerror is the specific fact and has to win, or the errno silently
+    vetoes a code that was deliberately added."""
+    dropped = _os_error(errno.ENOENT, winerror=59)
+    assert retry.is_transient(dropped)
+
+
+def test_a_permanent_windows_code_is_not_rescued_by_a_transient_errno():
+    """The precedence has to cut both ways, or it is not precedence — it is
+    just a second chance for anything with the right errno."""
+    denied = _os_error(errno.EIO, winerror=5)
+    assert not retry.is_transient(denied)
+
+
+def test_an_unknown_windows_code_is_not_retried():
+    """The default is to fail. A code nobody has reasoned about is not given
+    the benefit of the doubt, because the cost of guessing wrong is a delay
+    that hides the real fault."""
+    assert not retry.is_transient(_os_error(errno.EIO, winerror=999999))
+
+
+def test_exhausted_wrapping_a_network_drop_is_not_retried_again():
+    """A chunk-level loop raises Exhausted once it has spent its attempts.
+    Retrying it at a coarser level would repeat the same attempts against the
+    same dead session and re-read everything that already succeeded."""
+    spent = retry.Exhausted("read failed at offset 0 after 3 attempts")
+    spent.winerror = 59
+    assert not retry.is_transient(spent)
+
+
 def test_non_os_errors_are_never_retried():
     assert not retry.is_transient(ValueError("nope"))
     assert not retry.is_transient(KeyboardInterrupt())