Add an io::fsync module with a per-OS split (posix/windows) mirroring the
writeback_file convention, replacing two duplicated dir-fsync copies:
- dir(): POSIX directory fsync; a no-op on Windows, where std cannot open
a directory as a File and the failed open logged a spurious warning on
every mapfile write.
- file_durable(): opens the target read+write before sync_all so the flush
succeeds on Windows, where FlushFileBuffers rejects a read-only handle
with ERROR_ACCESS_DENIED.
Point the mapfile writer at the shared dir() helper.
Audited every #[repr(C)] struct and IOCTL/flag constant in scsi/windows.rs
against the authoritative Windows SDK headers (ntddscsi.h, winioctl.h,
devioctl.h, winnt.h, fileapi.h). All correct except the already-reverted
ScsiPassThroughDirect packing. Add the missing regression guards:
- StoragePropertyQuery layout (STORAGE_PROPERTY_QUERY: 0/4/8, size 12).
- IOCTL/flag constants, with IOCTLs asserted against an independent CTL_CODE
re-derivation (not a tautological literal) so a mistyped code is caught.
Validated compiling via cargo xwin check --target x86_64-pc-windows-msvc.
rc.4 added #[repr(C, packed(4))] to ScsiPassThroughDirect on the false premise
that ntddscsi.h wraps SCSI_PASS_THROUGH_DIRECT in #pragma pack(push, 4). It does
NOT — verified against the Windows SDK ntddscsi.h: the struct has no pragma pack
and uses natural alignment. On 64-bit Windows (LLP64) that puts DataBuffer at
offset 24 and the struct at 56 bytes, which bare #[repr(C)] produces and which
DeviceIoControl expects.
packed(4) instead imposed offset 20 / 48 bytes — the layout of the SDK's
SEPARATE 32-bit thunk struct SCSI_PASS_THROUGH_DIRECT32 (VOID* POINTER_32). Using
that on a 64-bit host malformed every IOCTL_SCSI_PASS_THROUGH_DIRECT, so the
INQUIRY in drive enumeration failed and autorip/CLI reported zero drives
('RC4 no longer detects my drive'). rc.3.1 (bare repr(C)) worked for the same
users; this restores that layout.
Replace the tautological packed-layout test (which asserted the same wrong
offsets the struct produced) with one cross-checked against the SDK header:
DataBuffer@24, SenseInfoOffset@32, Cdb@36, size 56. Verified compiling via
cargo xwin check for x86_64-pc-windows-msvc.
Expose the MKB Type field (record 0x10) as a typed MkbType enum with mkb_type()
/ mkb_type_raw() / mkb_is_uhd() helpers, so callers can distinguish AACS 1.0
(Blu-ray) from AACS 2.0/2.1 (UHD) discs without poking raw bytes.
When a scrambled AACS unit fails to decrypt under every available key
(a missing/wrong CPS sub-key, or a marginal unit that fails the TS-sync
verify), decrypt_sectors restored the original encrypted bytes and
returned Ok with no signal. Those still-encrypted bytes flowed to the TS
assembler, which silently dropped the non-syncing packets with no loss
counter. The only loss accounting was DiscStream's read-error zero-fill
path, so mux reported lost_video_secs=0 for decrypt-dropped content and
the abort gate accepted the rip even under abort_on_lost_secs=0. A rip
missing real video/audio segments was published as a perfect success.
decrypt_sectors now returns the number of bytes in scrambled units that
no key could decrypt. DecryptingSectorSource accumulates that into a
shared counter exposed via decrypt_loss(); both mux pipelines fold it
into lost_bytes() — the inline DiscStream path directly, and the
file-backed highway via PipelinedPesStream sharing the producer's
counter. Restore-to-original is unchanged, so clear nav-files are never
corrupted; metadata-probe callers that don't read the counter are
unaffected. Adds regression tests at the decrypt and decorator layers.
The doc claimed Error::KeydbInvalid for non-UTF-8 input, but the code
returns Error::KeydbParse (KeydbInvalid is reserved for the size-cap
violation). Correct the doc to match behavior and add a regression test
asserting non-UTF-8 yields KeydbParse.
In http_get, when the server closes the TCP connection before the HTTP
header block completes (n == 0 on the byte-by-byte header read), or sends
a header block exceeding 64 KiB, the code returned KeydbParse (E8004).
Both are connection/protocol-level faults from the server, not parse
failures of keydb content — the keydb bytes were never received. Return
KeydbConnect (E8000) instead, which already covers TCP-level exchange
failures. A CLI user hitting a transient drop or a redirect target that
immediately closes now sees the correct 'server hung up' diagnostic
rather than 'the downloaded file was malformed'.
Add a regression test that stands up a loopback listener which accepts
then drops the connection before headers, asserting KeydbConnect.
ntddscsi.h wraps SCSI_PASS_THROUGH_DIRECT in #pragma pack(push, 4),
forcing the PVOID DataBuffer field to 4-byte alignment even on 64-bit
hosts. The Rust struct used bare #[repr(C)], so the compiler applied
natural 8-byte pointer alignment and inserted 4 padding bytes after
TimeOutValue. That shifted DataBuffer to offset 24 (SDK: 20),
SenseInfoOffset to 32 (28), and Cdb to 36 (32), and grew the struct to
56 bytes (48). DeviceIoControl reads at the SDK offsets, so every SPTI
ioctl on 64-bit Windows either got rejected or interpreted garbage as
the CDB and DataBuffer pointer.
Add #[repr(C, packed(4))] to ScsiPassThroughDirect and the companion
SptwbDirect (so offset_of!(SptwbDirect, sense) stays correct for
SenseInfoOffset), plus a layout regression test asserting DataBuffer at
offset 20 and a 48-byte struct size.
The prefetch producer thread fires a BytesRead event after every
batch it reads, carrying a cumulative byte count. Nothing asserted
this callback actually fired, so a consumer that passed None for the
event_fn would silently get no progress events. Add a fixture-based
regression test that drives a finite extent through new_with_events,
captures the events, and asserts the cumulative count is
non-decreasing and reaches the full extent size at EOF.
This locks the contract autorip's mux progress bar and soft-stall
watchdog depend on.
A SEND KEY / REPORT KEY step in the bus-auth handshake mapped every
SCSI error to a cert/key-specific code (AacsCertRejected, AacsCertRead,
AacsKeyRead, AacsKeyRejected, etc.) via map_err(|_| ...). That discarded
the underlying SCSI error, so a transport-layer wedge (bridge crash / USB
disconnect) mid-handshake was reported as 'drive rejected your host cert',
sending operators down a keydb/host-cert dead end for what is really a
replug/power-cycle situation.
Add a handshake_err() helper that keeps the original error when it is a
transport failure (is_scsi_transport_failure) and only substitutes the
handshake-specific code for genuine SCSI rejections. Apply it at every
SEND KEY / REPORT KEY / REPORT DISC STRUCTURE step in both the AACS 1.0
and AACS 2.0 paths. Add a regression test covering both branches.
read_udf treated a READ CAPACITY SCSI failure as a 0-sector disc via
unwrap_or(0) with no diagnostic. capacity=0 then skews the layer
heuristic (always reports 1 layer, even for dual-layer discs) and the
canonical title-ordering sort, with nothing in /api/state or info to
indicate the command actually failed. Emit a tracing::warn carrying the
original error at the fallback site so a transient capacity failure is
visible. Recovery behavior is unchanged: 0 is still used as the
fallback.
default_path returned Error::KeydbParse (E8004, rendered as 'failed to
parse the keydb file') when HOME/USERPROFILE was unset. That misreports
an environment failure — a process with no home directory, typically a
stripped container or CI config — as a corrupt keydb file the code never
read. Map it to an IoError(NotFound) in the I/O category instead, so no
display path blames the keydb. Add a regression test.
Disc::read_aacs_inputs opened the ISO via FileSectorSource::open and
mapped any failure to Error::AacsNoKeys (E7000), discarding the real
Error::IoError (E5000) and its OS errno. A missing or unreadable ISO
(ENOENT/EPERM) is an I/O fault, not a key-resolution failure; callers
that dispatch on the error code would wrongly tell the user to check
their keys when the ISO simply does not exist.
Propagate the open error unchanged and add a regression test asserting
a nonexistent ISO yields E_IO_ERROR, not E_AACS_NO_KEYS.
detect_max_batch_sectors() is a Linux-sysfs probe with no platform
gate. It derived the device name with rsplit('/'), which never splits a
Windows \.\CdRom0 / \.\D: path, so the whole path became the device
name, no /sys node matched, is_optical fell to false, and the function
returned the 8192-sector block default (16 MiB/request) instead of the
60-sector optical default. That value then took the Some(b) arm in
Disc::copy and bypassed the 510-sector optical clamp that lives only in
the sysfs branch, leaving every Windows rip/verify running ~16x over the
optical cap (coarser bad-sector recovery, 16 MiB UDF reads).
Gate the sysfs probe behind a new sysfs_batch_probe_supported() helper
(Linux-only, requires a '/'-delimited path) and return the optical
default for any path the probe can't handle. Add regression tests for
the \.\ device-path forms.
STORAGE_ADAPTER_DESCRIPTOR.BusType is STORAGE_BUS_TYPE, an int-sized
(4-byte) enum, but the repr(C) struct declared it as u8. The total
size stayed 32 bytes by coincidence of alignment padding, and the two
fields actually read today (MaximumTransferLength at offset 8 and
AlignmentMask at offset 16) sit ahead of BusType, so there was no
runtime impact. But BusMajorVersion and BusMinorVersion landed at
offsets 26 and 28 instead of the SDK's 28 and 30, so any future reader
of those fields would have gotten wrong values.
Widen BusType to u32 so BusMajorVersion/BusMinorVersion fall at the
correct offsets, and add a layout regression test asserting every
field offset and the 32-byte total against the winioctl.h layout.
The doc comment claimed the version was a BE u32 at offset 8 of the
record body (offset 12 from pos), but the code correctly reads pos+8
(body offset 4): a 4-byte record header at pos, the Type field at body
offset 0, then the version at body offset 4. Rewrite the comment to
match the actual read so a maintainer does not 'correct' the offset and
break MKB version parsing. Clarify the matching test comment too.
aligned_unit_keys_validate accepted a unit-key set as soon as ONE
scrambled sample decrypted. On a multi-CPS-unit disc a set covering
CPS unit 0 but not CPS unit 1 therefore passed: decrypt_with committed
it, the sweep proceeded, and CPS-unit-1 sectors passed through as raw
encrypted bytes into the ISO/MKV with no error surfaced anywhere.
Require every scrambled sample to be descrambled by some unit key.
A sample no key covers now fails the gate, so an incomplete set is
rejected (AacsKeyRejected) and the caller falls through to the next
candidate, ultimately surfacing a key error instead of silently
writing ciphertext. Wholly-wrong-key rejection is unchanged.
Add a regression test for the partial-coverage case.
The comment on the bridge-degradation branch in handle_read_error
claimed it matched the NOT_READY 04/3E sense signature, but
is_bridge_degradation() keys solely on a non-standard SCSI status byte
(anything that is not GOOD/CHECK CONDITION/TRANSPORT FAILURE) and
ignores sense_key/ASC/ASCQ. A real 04/3E bad-sector error arrives as
CHECK CONDITION (0x02), so it never took this branch — it falls through
to the generic NOT_READY retry. Rewrite the comment to describe the
status-byte condition the predicate actually detects, and drop the
parallel misleading note in the not_ready_err test helper.
Add a regression test asserting a NOT_READY 04/3E error is not
classified as bridge degradation and routes to the NOT_READY retry
(3 s pause) rather than the bridge cooldown (15 s pause).
SptiTransport::reset() unconditionally slept 2 seconds after sending
IOCTL_STORAGE_RESET_DEVICE, even when the IOCTL failed (e.g.
ERROR_INVALID_FUNCTION on a driver that does not support the reset).
On failure no reset occurred, so there is nothing to settle and the
2-second penalty was pure waste. Gate the settle sleep on the IOCTL
return so it only fires when the drive was actually reset.
keydb::save() overwrote the live keydb.cfg with a bare in-place
std::fs::write. keydb.cfg is the single source of AACS truth and this
path runs unattended (first-boot download + daily-refresh thread, with
a container restart on every release), so a SIGKILL, OOM-kill, power
loss, or ENOSPC mid-write could leave the file truncated with the prior
good copy already gone. A truncated keydb does not error at write time;
it surfaces later as failed key resolution on every AACS rip.
Factor the write into write_atomic(): create the parent dir, write a
unique sibling temp file, fsync, then rename (atomic within a
filesystem). On any write/fsync/rename failure the temp is removed and
the existing keydb is left untouched. Same pattern already used by the
settings and mover write paths. Add regression tests covering in-place
replacement (no stray temp) and prior-copy preservation on failure.
DiscStream skips a whole AACS unit (3 sectors = 6144 bytes) per
read-error event, but only the skip-event count was exposed. Loss
estimates built from errors*2048 therefore undercounted AACS loss ~3x.
Add a lost_bytes field that accumulates the actual zero-filled byte
count at each skip, expose it via a new Stream::lost_bytes() accessor
(default 0; DiscStream and CountingStream override), so consumers can
scale lost-video time by real bytes lost rather than the event count.
Regression tests assert the AACS path records 6144 B/event (and
exceeds the errors*2048 undercount) while the align=1 path records
2048 B/event.
When finish_with_halt's grace period expires it detaches from the
consumer thread and returns an error to the caller, but the leaked
consumer kept running to completion: once its wedged write syscall
returned it would fall through to sink.close(). For the mux writer
close() finalises the MKV (Cues block + segment-header patch), so a
leaked consumer could finalise — and keep writing to — an output file
the caller had already reported as failed, racing a fresh rip for the
same device over the same path.
Add a shared abandonment flag the consumer polls in its drain loop and
again before close(). finish_with_grace sets it before dropping the
JoinHandle, so the moment the wedged syscall returns the consumer skips
any further apply and skips close() entirely, then exits. This does not
interrupt the in-flight syscall (only its return or process exit can),
but it bounds the damage to the write already in flight instead of a
full finalise of an abandoned file.
Regression tests cover both sides: a consumer leaked past the grace
period skips close(), while one that finishes inside the grace window
still calls close() and finalises normally.
SptiTransport::reset() discarded the DeviceIoControl return value, so a
wrong or unsupported reset IOCTL would fail with ERROR_INVALID_FUNCTION
and silently no-op while the unconditional 2s settle sleep made it look
like a reset happened. That is exactly the regression class the doc block
records for the two earlier (incorrect) code values.
Bind the result and warn (with GetLastError) when the reset fails, debug
on success. Lift IOCTL_STORAGE_RESET_DEVICE to module scope and add a
test recomputing it from the CTL_CODE formula so a wrong value can't slip
back in unnoticed.
Mapfile::flush() wrote the new state to a .tmp sibling, sync_all()'d the
temp file, then rename(2)'d it over the final mapfile path — but never
fsynced the parent directory. After the rename the new dirent lives only
in the directory's page cache, so a crash or power loss in the
rename-commit window (the wide window on NFS, the very case the temp
fsync guards) can lose it: resume then reads a stale or absent mapfile
even though the data bytes were durable, silently discarding multi-pass
recovery progress.
Add a best-effort fsync_dir() on the path's parent after the rename,
mirroring the established dirent-durability pattern in autorip's mover.
A directory that can't be opened or synced is logged and ignored rather
than failing the write, since the file bytes are already durable.
Adds a regression test exercising the parent-fsync branch against a real
subdirectory and asserting the helper is a no-op on a missing directory.
Tests, crates.io publish, and the GitHub Release no longer serialize
behind each other: each job only needs the version-check (verify) gate.
Binary consumers git-tag-pin libfreemkv, so they start building the
instant the tag exists rather than waiting on the crates.io publish.
crates.io publish runs as an independent job for external consumers and
uses --no-verify (CI already compiled this commit in the test job).
Drive::read fires hundreds of thousands of times per rip. Logging its
entry at DEBUG floods a diagnostic log and buries the events that
actually matter. Move it to TRACE so a level-3 (debug) bug-report log
stays readable; level-4 (trace) still captures it for deep dives.
The docs still documented the old in-tree firmware unlocker: the MediaTek
MT1959 variant table, the READ BUFFER unlock CDB bytes, the profiles.json
schema (unlock_mode/unlock_buf_id/unlock_cdb), the platform/mt1959 driver
listings, and the 'why unlock is needed' handshake mechanism. None of that
lives in libfreemkv anymore — the core is firmware-clean and ships only the
pluggable Unlocker trait + registry (src/unlock.rs).
Rewrite drive-access, architecture, api-design, disc-to-rip, and the README
to describe only the generic Unlocker seam: the trait, register_unlocker, the
registry routing, and the host-cert fallback when no unlocker matches. Point
readers to the freemkv-unlock repo for concrete unlockers. No source change.
The dual-model Windows audit (Sonnet) caught that the prior 'fix' (0x002DD000)
was also wrong: that decodes to the OBSOLETE RESET_BUS code (function 0x400,
R|W access) which class drivers reject. Canonical ntddstor.h:
IOCTL_STORAGE_RESET_DEVICE = CTL_CODE(0x2D, 0x0401, METHOD_BUFFERED, FILE_READ_ACCESS)
= (0x2D<<16)|(1<<14)|(0x401<<2)|0 = 0x002D5004.
Only affects the best-effort drive-reset recovery path (return value is
discarded); the normal read/unlock/rip path never calls it, so this does not
change normal Windows operation. Verified correct-by-construction; the
AlignmentMask finding is deferred to rc4 (benign on USB optical bridges,
needs a real Windows SCSI-HBA rig to validate).
- patch (Pass N) now aborts immediately on transport failure (status=0xFF),
symmetric with the sweep and single-pass mux. Previously a USB-bridge crash
was treated as an ordinary bad sector and the pass hammered the crashed
device sector-by-sector until the per-range watchdog expired. (medium)
- patch AACS recovery reads are now unit-aligned: a mid-unit single-sector
read on an AACS disc was rejected by the decrypting reader (DecryptFailed)
and the sector abandoned without asking the drive. The read is now widened
to the enclosing whole 3-sector unit and the requested window copied out,
leaving all recovery accounting (pos/block_bytes/cursor) untouched so it
cannot desync. Only affected CLI decrypt-to-ISO --multipass re-runs. (low)
- IOCTL_STORAGE_RESET_DEVICE corrected 0x002D1004 -> 0x002DD000 (the old value
decoded to function 0x401 with the access bits cleared, so DeviceIoControl
would fail ERROR_INVALID_FUNCTION instead of resetting). Windows-only. (low)
Adds a transport-failure classification regression test.
A direct disc://→mkv:// single-pass rip drives fill_extents in
skip_errors mode. On a read failure it shrank the batch, retried, and
once bottomed out zero-filled + skipped the unit and continued. A SCSI
transport failure (status=0xFF) is a USB-bridge crash, NOT a skippable
bad sector: the bridge is wedged and every subsequent read fails the
same way. So the loop marched the entire disc at one ~15s bridge-
recovery per probe, producing no MKV — the user-reported 'hundreds of
0x28/0xff warnings, runs forever, Movie.mkv never created'.
Fix: short-circuit to an error on transport failure before any
shrink/skip, even under skip_errors — mirroring the multipass sweep's
transport-failure rule in read_error::handle_read_error. The CLI
surfaces it so the user power-cycles the drive or switches to multipass
recovery. Regression test asserts exactly one read is issued and no skip
is counted (no infinite march).
Three fixes for a Windows ASUS Blu-ray drive that failed/spammed errors:
- resolve.rs: accept disk:// as an alias for disc:// (identical behavior;
empty = auto-detect, path = device). Windows users commonly type
disk://i: after the drive-letter convention.
- drive::find_drive: prefer a drive that reports media present. Enumerate
all optical drives, query Drive::drive_status() (GET EVENT STATUS, works
regardless of firmware), and return the first reporting DiscPresent;
fall back to the first enumerated drive when none report a disc so
single-drive / quirky setups don't regress. Selection policy split into
select_drive_with_media() for unit testing.
- READ chunking: add ScsiTransport::max_transfer_bytes() (default 1 MiB).
Windows SPTI overrides it with the adapter MaximumTransferLength queried
via IOCTL_STORAGE_QUERY_PROPERTY / StorageAdapterProperty, clamped to a
64 KiB floor (fallback on query failure). Drive::read now caps each
READ(10) to that limit: small reads take the unchanged single-CDB path,
larger reads loop over read_one() chunks, reporting the failing chunk's
LBA on error. This stops the 16 MiB single read that exceeded the
adapter limit, made DeviceIoControl fail, and spammed transport-failure
warnings with slow tiny-read fallbacks.
Tests added for the disk:// alias, media-preference selection, and READ
chunk decomposition / per-chunk error LBA.
Complete the OEM/AACS cert baseline so host certs are a KeySource output,
never compiled in. With an unlocker present the OEM route is unused
(unlocker_read_volume_id short-circuits); without one, the cert handshake
runs when a keysource supplies a host cert and fails gracefully when none
does.
- KeySource trait gains host_certs() (default empty), reusing the existing
aacs::HostCert type. A source holds certs as its second kind of AACS
material alongside decryption keys.
- ScanOptions gains key_sources so the handshake can collect certs across
the app's keysource layer, unioned with DriveCredentials.
- do_handshake_cert collects certs via collect_host_certs (credentials +
every key source). Zero certs from any source now returns the new
graceful Error::AacsNoHostCert (code 7024, sentinel <no host cert>)
instead of silently skipping; resolution still falls back to the
path-1 disc-hash -> VUK lookup, which drops the error on a hit.
- error.rs: add E_AACS_NO_HOST_CERT / Error::AacsNoHostCert, wired into
code(), Display, and the round-trip + sentinel tests.
HandshakeResult { volume_id, read_data_key } unchanged: the cert path
still yields both the VID and the bus key.
Rename the trait to a generic, drive-neutral capability contract so future
unlockers don't conform to LibreDrive specifics:
- unlock(...) -> unlock_drive(...) (the one required capability)
- read_vid(...) -> read_volume_id(...) (no-op default)
- add set_max_read_speed(...) (no-op default)
The trait doc now states the contract in one place: unlockers are optional
drive-capability providers; the AACS layer is the always-present baseline and
falls back to the full cert handshake when no unlocker matches. Implement only
the capabilities your drive supports.
Registry: route_unlock now calls unlock_drive; unlocker_read_vid renamed to
unlocker_read_volume_id; add unlocker_set_max_read_speed (mirrors route_unlock
resolution, first matching unlocker, no-op if none match). drive::init calls
it on a matched drive in the post-unlock path; a speed-set failure is logged
and does not fail the rip. encrypt.rs handshake updated to the new VID helper.
Tests updated for the renames; added a set_max_read_speed routing test
(match invokes, no-match is a safe no-op).
An Unlocker unlocks drive functionality, not just the disc: unlock() is
one capability, OEM VID retrieval is another. Widen the Unlocker trait
with a default-no-op read_vid(), add an unlocker_read_vid registry helper
that mirrors route_unlock resolution, and consult it in do_handshake_cert
before the cert-based VID read. A matching unlocker that serves a VID via
its OEM path short-circuits the cert handshake — VID is obtained without
the host certificate + HRL (restoring the pre-refactor decoupled OEM VID
path, now living inside the unlocker). Non-matching drives, and unlockers
without an OEM VID path, fall through to cert auth unchanged.
is_unlocked() now reports the honest signal (a registered unlocker matched
this drive) instead of const false.
libfreemkv must stay firmware-clean for crates.io. Move ALL drive-unlock
knowledge — firmware blobs, WRITE_BUFFER/MODE SELECT upload, unlock CDBs,
the MT1959 variant-A/B handshake, the 800 KB profiles.json database, and
the DriveProfile parsing — out into the freemkv-unlock-ld crate.
libfreemkv now keeps only the seam:
- Unlocker trait (name/matches/unlock) + a process-wide ordered registry
(register_unlocker / route_unlock) in src/unlock.rs
- Drive::init() walks the registry; the first unlocker whose matches(id)
is true runs unlock(scsi, id); if none match the drive is left in
stock mode and the host-cert AACS handshake (the OEM route) carries
the disc.
The unlocker issues its own CDBs through the public ScsiTransport::execute,
so libfreemkv knows nothing about how unlocking happens.
Removed:
- profiles.json
- src/platform/mt1959/{mod,variant_a,variant_b}.rs
- src/profile.rs (DriveProfile, ProfilesFile, find_by_drive_id, ...)
- the PlatformDriver trait
Because the Unlocker seam reports only success/failure (no extended-access
marker), VID acquisition is now always via the cert-based handshake; the
per-drive OEM-VID-CDB shortcut and Drive::is_unlocked() (now const false)
are removed/neutralized. Disc-speed calibration moved into the unlocker's
unlock(); Drive::probe_disc() is a no-op.
git grep over src/ is firmware-blob/profiles/WRITE_BUFFER/mt1959-free.
All tests pass on Rust 1.86 (precommit green).
Fixes corrupt MKV seek index on single-clip titles with many
interleaved tracks (Top Gun UHD: 2 video, 11 audio, 32 PGS).
TimelineContinuity previously shared one high_ns frontier + offset_ns
across ALL tracks. A sparse, lagging non-video frame (subtitle/audio)
ratcheted the frontier up; the next normal video frame then sat >3s
below it and was misread as a clip-boundary discontinuity, permanently
bumping offset_ns. On a one-clip title this fired thousands of times
and inflated Cue/cluster timestamps into the billions of ms, destroying
the seek index (ffmpeg then seeked to wrong positions and emitted
spurious 'Could not find ref with POC N' errors).
Now only the VIDEO track drives epoch decisions: video alone advances
the frontier and opens a new epoch on a real backward PTS jump.
Non-video tracks are remapped under the current offset and never touch
the frontier or offset. A lagging non-video tail straggler at a genuine
multi-clip boundary (old-epoch raw PTS under the new offset) is
recognised via the previous offset and remapped to the seam, so it
neither flies forward nor forces a back/forward-dated split cluster.
Genuine multi-clip seamless rebasing is preserved.
Also drop TimestampScale from 1ms to 0.1ms (100_000 ns/tick) so
23.976fps frames and 0.833ms TrueHD AUs stop colliding on a single
tick (the source of the non-monotonic-DTS warnings and the audio
cadence flattening). The finer scale shrinks the i16 block-relative
span to ~3.27s, so: cluster duration is set to 2s nominal (keeps
keyframe-driven clusters within the i16 range for typical GOPs), and
the i16-overflow cluster-split path now emits a Cue for the split
cluster so the seek index has no gaps.
Regression tests: single-clip late-subtitle must not inflate offset_ns;
non-video must not advance the frontier; non-video straggler remapped
to seam at a real boundary; every cluster (incl. i16-split) carries a
Cue. Existing tick/duration assertions updated for the new scale.
Matthew Jackson