The per-range walk was depth-first: each bad range ran the full handler
chain (fast + slow deep-recovery + bisect) before the next range was
touched. So a handful of tiny dead fragments at one end of the disc
burned the whole pass and the big NonTrimmed ranges elsewhere — usually
sweep-jump over-marks that read straight back — were never attempted.
Now recovery runs in two breadth-first tiers over ALL sections:
- Tier 0 gives every section one fast full-batch attempt (fast reads
only), largest ranges first, so the recoverable bulk of the disc comes
back in the first minutes.
- Tier 1 deep-recovers only the residue tier 0 could not pull.
Per-section still-bad sets persist across tiers. Largest-first ordering
means a big readable region is reached before time is spent on tiny dead
fragments.
Linear no longer collapses a failed batch to count=1 single-sector reads
(live probing: a marginal sector recovers in a large read, not a lone
one) — a failed 32-batch stays 32 and is re-attempted at full size by the
next handler/pass; Bisect salvages readable islands.
Adds a handler-start trace line so the debug log shows which handler is
running and the hand-off to the next.
Pass-N recovery is now a chain of time-bounded recovery handlers instead
of one monolithic per-range loop that could grind the front of a bad
range for up to 30 min, wedge the drive, and abort the whole pass.
A bad range is a SubRanges set; recovery is an ordered list of
SectionHandlers (Linear{reverse,fast} covering back/forward x fast/slow,
and Bisect). A coordinator runs each handler with a hard per-handler
deadline: a handler recovers what it can (removing it from the still-bad
set) and hands the rest to the next handler; whatever is still bad after
the chain becomes NonTrimmed residue and we move on to the next range.
Guarantees, now structural rather than bolted-on:
- never hangs: every handler is deadline-bounded; the loop always drains
to recovered-or-residue.
- always moves on: a range that cannot be finished leaves residue and
advances; only a genuine transport fault or user halt ends the pass.
- extensible: a new recovery idea is one SectionHandler impl added to the
chain; a proven-ineffective one is removed. The engine never changes.
Removes ~1.9k lines of the old inner loop (watchdogs, skip escalation,
NOT_READY grind, wedge counters) and their tests. fast_capture is now
inert (the chain supersedes it); breadth-first ordering becomes a future
scheduler concern. New module: disc/section_recover.rs (8 fixture tests,
injectable clock — bounded/never-hang proven without touching a drive).
Two A/B tests updated to the chain's strictly-better recovery counts.
Every Drive::read now forces the read from physical media instead of
letting the drive satisfy it from cache. A recovery tool must not trust
the cache: on a re-read of a marginal sector the BU40N can return a
cached miss (~4 ms) instead of giving the surface a fresh physical
attempt (~40-250 ms), masking a sector that would otherwise read on a
real media hit. FUA (READ(10) byte 1 bit 3) is validated on the
BU40N/Initio bridge.
Spin the disc down (START=0, LOEJ=0) then back up (START=1) to clear the
BU40N/Initio fast-fail wedge state a run of HARDWARE_ERROR reads leaves the drive
in — the non-eject power-cycle our notes say the wedge needs. The disc stays
loaded; we never eject (slot-loading — a human eject is a product failure for an
unattended service). Validated live 2026-07-01: took the drive from
failing-every-read back to reading at MB/s.
A fast_capture pass defers every failed block to a granular pass anyway, so
spending the drive's 60s deep-recovery timeout on it here only freezes the
breadth-first sweep on a pure-bad cluster (~25s per 32-block, incl. NOT_READY
grind). Read with recovery=false in fast_capture so a bad block fails on the
short timeout (~10s) and the sweep steps on; the granular passes
(fast_capture=false) do the deep recovery on what's left. No recovery lost —
the block stays NonTrimmed for the granular retry.
A PatchOptions.fast_capture pass reads each bad range ONCE at the full batch and
leaves every FAILED block NonTrimmed for a later pass — no bisect, no per-sector
grind, no retry. This lets a first retry pass grab the readable blocks (the
sweep's good skip-ahead overshoot) of EVERY section quickly, before any single
section's slow per-sector recovery — instead of grinding section 1 to exhaustion
before even touching section 2. A later pass (fast_capture = false) does the
granular bisect/retry on what's left.
Load-bearing invariant (fixture test): NO data is dropped. A failed block becomes
NonTrimmed (pending, retried by a granular pass), NEVER Unreadable. The test pins
that the readable half of a range recovers, the bad half stays NonTrimmed (not
Unreadable), and the bad block is marked in ONE batch read with zero bisection.
Disc::copy's internal patch keeps fast_capture=false (single-call full recovery).
The library now offers a one-shot PassProgress snapshot built from a mapfile on
disk + title, so a client gets the fully-rendered drilldown without parsing the
mapfile itself — used for autorip's pass-boundary and mux-entry paints (no live
callback yet at those points). Adds a locate_ranges test covering in-feature vs
out-of-feature at-risk time (ported from autorip's removed from_map tests).
Add LocatedRange + LocatedProgress to the progress contract and a 'located'
field on PassProgress, populated by the sweep + patch emitters from the
in-memory bad-range set + title. Move the range->chapter/time annotation
(locate_ranges, range_chapter, byte_offset_in_title) into the library so a
client renders the disc map + at-risk movie time straight from PassProgress
and never reads the mapfile itself — if the mapfile becomes a mapdb, this type
and its producer change, clients don't.
PassProgress is no longer Copy (located carries a Vec); it's built once per
throttled emission and passed by reference. Non-locating phases (verify,
extract) emit LocatedProgress::default(). Adds consts::MILLIS_PER_SEC.
Consumer-side wiring (autorip drops Mapfile::load) follows.
Live drive probing (cold-single vs in-stream batch vs centered window,
one clean uncached read per sector) showed recovery RATE is identical
across approach: 3/5 every mode. The drive's per-sector ECC is
media-bound and partly stochastic, NOT approach-bound. So scatter
(seek-away recalibration + far-anchor re-read) does not improve recovery
and is dead weight — removed (function, read_good_sectors, SCATTER_*
consts, and its tests).
Add SubRanges: the still-bad (pos,len) sub-ranges of a section with
remove()/split, the work-list foundation for the upcoming per-section
recovery decomposition (recover_section -> stream/bisect/retry phase
helpers). Unit-tested; unused until the decomposition lands.
25 patch + 8 AB profile tests green; precommit clean on Rust 1.86.
Live probing on the BU40N/UHD testbed (rprobe/dprobe) settled the speed
question empirically: a marginal sector reads ~12x FASTER at MAX speed
than at slow, and slow NEVER recovered one MAX didn't. Cleanly-tested
(one uncached read per sector) the recovery RATE is identical across
speed/direction/window/batch — the drive's per-sector ECC is media-bound,
not approach-bound. The only host lever that matters is read SPEED.
So the old 'drop to 0x0000 on first failure and hold it for the whole
range' only slowed the GOOD sectors of a bad range — measured ~3x slower
overall. Now the range stays at MAX; the first failure just re-attempts
once (stochastic media) and falls through. Live A/B: 11 KB/s -> 33 KB/s,
bytes_lost unchanged.
Also: batch adaptation is now halve-on-failure / double-on-success
(geometric), so a bisected batch climbs back through clean runs from any
size (the old count==1-only upscale left it stuck at mid sizes). Removed
the dead consecutive_singles_ok counter and the inert batch-retry probe.
25 patch tests green; precommit clean on Rust 1.86.
Two patch-recovery speed fixes, grounded in live rip1 telemetry on the
Dune UHD bad regions.
1. Pull prime_cache. In the reverse walk it issued throwaway reads of
lba-3..lba-1 — i.e. straight into the not-yet-visited dead L-EC core —
grinding each ~7-9s at the fast timeout, then discarding the result.
The live trace showed a good target read (read_duration_ms=1) taking
9s wall-clock purely from the prime grind ahead of it. The dead-zone
skip (Tier 1) + scatter's recalibration read (read_good_sectors)
cover the cache-warm intent without grinding the dead zone, so prime
is redundant harm — removed (function + CACHE_PRIME_SECTORS + call).
2. Bisect on batch failure. handle_read_failure used to collapse a
failed count>1 batch straight to count=1 and single-walk the entire
batch. Now it halves (count/2) and retries the same start: a good
half recovers in BULK, the bad sector is isolated in O(log n) reads
instead of O(n). Invariant preserved — only a count==1 failure marks
NonTrimmed, so no good sector is lumped into a bad mark.
These are Tier 2 (bisect) of the progressive recovery model
(fast-mass-skip -> bisect -> scatter); composable so order can be tuned.
25 lib + 8 AB fixture tests green; precommit clean on Rust 1.86.
The 'reset, read good data, come back for one sector' technique. A
genuinely-damaged sector makes the drive grind its full recovery timeout
and still fail; re-reading consecutive bad LBAs at identical conditions
re-fails AND is the rapid-failure cadence that drops the BU40N into a
firmware fast-fail wedge (CLAUDE.md hard-rule #2).
scatter_recover: on a stuck single sector, make up to SCATTER_MAX_ATTEMPTS
fresh tries, each preceded by read_good_sectors() — a seek to a far
known-good anchor + a sustained read that re-seats the head/servo and
breaks the failure cadence. The fresh re-read uses the FAST timeout
(recovery=false) so a recalibrated marginal sector reads quick and a
truly-dead one fails fast instead of burning 60s per attempt. The seek +
good read IS the settle (matches ddrescue/MakeMKV) — no idle sleep.
Gated to genuine single-sector MEDIUM_ERROR (0x03): transport faults
still abort the pass, NOT_READY keeps its own retry path, wedge-family
senses are left for cooldown/eject. A scatter recovery is recorded
exactly like a normal read success (write-then-mark-Finished); nothing
marks a range good without a successful read + write.
Logs anchor_ms/reread_ms per attempt for live tuning. 5 fixture tests
(recover-marginal, give-up-dead, skip-non-medium, skip-batch, AACS
widen); 2 AB read-count guards raised for scatter's bounded extra reads.
Precommit green on Rust 1.86.
Pure structural refactor of the Pass-N recovery path, no behavior change.
- PatchCtx<R> holds the one-pass context (reader, pipe, shared, opts,
reused read buffer, cross-region state) and owns orchestration.
- PatchCtx::run() walks the ordered bad ranges, applies the inter-range
cooldown (armed on a grind, consumed at the next range's entry), and
stops the pass the moment a range reports halt/wedge/transport-fault.
- PatchCtx::patch_region() is the focused single-range loop: read ->
success/failure -> damage-skip -> per-range watchdog. Every former
break / break 'outer is now a typed RegionOutcome (Completed,
SkipLimit, BudgetExceeded, Wedged, Halted, TransportFault) that run()
maps back to next-range vs stop-pass exactly as before.
- recovery_read() extracts the AACS mid-unit-window widening + plain
read into one reusable helper (the building block the upcoming
scatter re-read will share).
- Logging unified onto a single dotted patch.* taxonomy (patch.region.*,
patch.read.*, patch.probe.*, patch.wedge.*, patch.batch.*,
patch.mapfile.*, patch.skip.*, patch.stall, patch.speed); the
freemkv::scan pass-boundary markers are left intact.
20/20 disc::patch tests green; precommit (fmt + clippy -D warnings +
tests) clean on Rust 1.86.
is_scsi_transport_failure recognized only the 0xFF SCSI sentinel, so the two
NON-SCSI dead-bus faults the Linux SG_IO transport returns — Error::IoError
(ioctl(SG_IO) == -1: ENODEV/EIO on an unplugged bridge) and Error::DeviceNotFound
(fd gone) — were classified as recoverable bad sectors. Every caller
(sweep read_error, patch, single-pass fill_extents) would then zero-fill /
skip-ahead across a wedged bus instead of aborting the pass and re-enumerating
the bridge — potentially zero-filling a huge span. Add both to the classifier so
a dead bus aborts the pass everywhere. (Companion to the unlock_bridge fix, which
maps the same two faults to 0xFF across the freemkv-unlock seam.)
Round-2 audit gap in the prior bridge fix: extract_scsi_context only recognizes
Error::ScsiError / DiscRead and collapses every other variant to (0, None). But
the Linux SG_IO transport returns genuine dead-bus faults as non-SCSI variants —
Error::IoError (ioctl(SG_IO) == -1: ENODEV/EIO on an unplugged bridge) and
Error::DeviceNotFound (fd gone) — which were crossing the seam as status 0x00, so
the unlock crate classified neither as a transport fault (needs 0xFF) nor as a
sense-bearing rejection, and kept hammering a dead bus.
Match the error variant in the adapter: a SCSI status (CHECK CONDITION or a
drive-tagged 0xFF) keeps its real status + sense; any other variant is a
non-SCSI transport/IO fault → SCSI_STATUS_TRANSPORT_FAILURE / None. Keying off
`sense.is_none()` would be wrong (a senseless CHECK CONDITION is a rejection).
Add bridge tests covering CHECK-CONDITION sense preservation, 0xFF passthrough,
and IoError/DeviceNotFound → 0xFF.
DiscStream (the live-drive single-pass path) enables AACS decrypt-loss
concealment (NULL-TS fill on an undecryptable unit) but, unlike the file-backed
PipelinedPesStream, had no B1 resync gate — so after a concealed gap it forwarded
inter-coded video frames referencing the now-missing data, producing
decode-broken MKV output (dangling-reference frames).
Mirror PipelinedPesStream: add per-stream ResyncGate + is_video, and route every
codec-frame emit on the TS path (in-stream parse, demuxer EOF flush, parser EOF
flush) through gate.admit(is_video, frame.discontinuity, frame.keyframe) so a
video track drops forward to the next keyframe after a concealment event. Warn
once if a gate is still armed at EOF (a concealed gap in the final GOP with no
trailing keyframe). Audio/subtitle always admit.
Audit fixes (v1.1.0..HEAD regressions in the unlock migration + adaptive patch
speed):
- unlock_bridge ScsiAdapter: libfreemkv's transport returns Err on ANY non-zero
SCSI status (a normal CHECK CONDITION), not only transport faults. The adapter
was collapsing every such Err to { status: 0xFF, sense: None }, which discarded
the parsed sense and defeated the AACS handshake's ILLEGAL_REQUEST wedge guard
(so it kept hammering the drive — hard-rule #2) and inverted its
transport-vs-rejection diagnosis. Now reconstruct status + the 32-byte sense
buffer (sense_key@2, asc@12, ascq@13) and only emit 0xFF/None for a genuine
transport fault.
- Drive::init: a genuine transport fault during the drive-prep unlock means the
bus is dead — propagate it (the v1.1.0 invariant) instead of silently
swallowing it via `if let Ok`. Other errors (no matching unlocker) still fall
through to stock mode. SET CD SPEED max now runs only when the bus is alive.
- disc::patch: on the first read failure in a range, drop to slow recovery speed
and RE-ATTEMPT the same position at slow speed before marking it. A
single-sector range's first failing sector was being marked from a MAX-speed
read it never got to recover.
- docs: lib.rs architecture diagram (handshake → host_certs) and README (stale
pluggable-unlock-seam / register-unlocker / crates.io / docs.rs references).
Two recovery-path fixes:
- patch: log the per-range speed transitions (INFO, phase="patch_speed") — each
range enters at 0xFFFF (max) and drops to 0x0000 (slow recovery) on its first
read failure. Previously the adaptive-speed behavior was invisible in the logs.
- drive init: issue the generic SET CD SPEED max UNCONDITIONALLY at drive-open,
not only when a firmware unlocker matched. A stock-mode BD/UHD drive (no
firmware unlock) was left riplocked because the call sat inside the
unlocker-matched branch.
libfreemkv now git-deps freemkv-unlock (which carries drive firmware and can
never be on crates.io), so libfreemkv itself can only be consumed by git tag.
The freemkv-unlock dep stays an interim path dep for local cross-repo dev; the
release script swaps it to a git tag in the tagged commit.
Rewire the three unlock dispatch points through the freemkv-unlock crate via a
private `unlock_bridge`: drive-prep (kind=Unknown) at `Drive::init`, AACS cert
(kind=Aacs) at `do_handshake_cert`, CSS bus-auth (kind=Css) at scan. The bridge
news up `all_unlockers()` and runs the first matching one, mapping its
`Unlocked` result to the bus-key gate. After a successful drive unlock,
libfreemkv issues a generic SET CD SPEED (max) itself — the old per-unlocker
trait method is gone.
Delete the in-tree unlock code now owned by freemkv-unlock: the AACS cert
handshake (`aacs/handshake.rs`), the CSS bus-auth (`css/auth.rs`), and the
unlock registry (`unlock.rs`). Host-cert collection (a keysource concern) stays
in a small `aacs/host_certs.rs`. No public unlock surface remains — clients
touch libfreemkv only, oblivious to unlockers (as they are to SCSI). 2277 tests
pass.
Add the freemkv-unlock dependency (path dep for local dev; re-pinned to a git
tag at release) and the consumption bridge: one generic ScsiAdapter mapping
libfreemkv's transport to the unlock contract, host-cert + DriveId mapping, and
run_unlockers() which news up all_unlockers() and runs the first matching one.
Additive — validates the dep resolves and the adapter compiles; the in-tree
registry/handshake/css path is still active and gets replaced in 4b.
Pull the wedge-guarded cert loop and host-cert collection out of the in-tree
AacsCertUnlocker into public aacs::handshake primitives (run_cert_handshake +
CertHandshake, collect_host_certs). The in-tree path now delegates to them, so
the external freemkv-unlock-aacs plugin runs the IDENTICAL cert handshake — one
implementation, two callers. Pure refactor of the live AACS path; the existing
handshake + collect_host_certs unit tests validate it unchanged.
Replace route_unlock's Option<(name, Vid)> with a structured UnlockRoute
{ Unlocked(name, Unlocked) | Failed(UnlockError) | NoMatch } so a single
dispatch serves every caller: drive-prep wants "did anything unlock", and the
AACS cert route (next) needs the FAILURE REASON to render "missing keys" vs
"host cert rejected" instead of collapsing it to a bare None. Only a genuine
SCSI transport fault still returns Err (abort). UnlockCtx gains an optional
ScanOptions (the cert route's host-cert source), and read_mkb_from_drive now
takes &mut dyn ScsiTransport — both prerequisites for the cert handshake to
become an external freemkv-unlock-aacs unlocker. Drive-prep + CSS callers fold
the new outcome; no behavior change.
Extend Unlocked with drive_unlocked + read_data_key_err so the registry result
carries everything the downstream bus-key gate needs — a firmware unlocker sets
drive_unlocked=true (bus encryption removed at the drive, no bus key), the cert
route will set read_data_key (+ read_data_key_err on a failed bus-key read).
route_unlock now returns the full Unlocked rather than just a Vid, so a single
dispatch can fold the firmware, cert, and CSS results uniformly. Prep for the
AACS cert handshake becoming a registry unlocker.
Convert the CSS read-unlock into a first-class registry Unlocker (CssUnlocker)
dispatched through route_unlock like every other barrier removal, instead of
a direct call in scan. libfreemkv appends the built-in CSS unlocker (and, next,
the AACS cert handshake) exactly once via ensure_builtins(), AFTER any
client-registered firmware unlocker — so the registry order is firmware → cert
→ css, owned by the lib, not the client.
Defense in depth: the unlocker does NOT trust the caller-declared DiscKind.
matches() filters on the declared kind (Css), but unlock() self-verifies
against the drive's GET CONFIGURATION profile and refuses (UnlockError::
NotApplicable, a new shared "this unlocker doesn't apply" variant) WITHOUT
issuing a single CSS CDB if the drive reports a non-DVD profile — so a
mis-routed Blu-ray is never sent CSS bus-auth. Guard the firmware unlocker the
same structural way (it matches only the drive-prep phase, kind == Unknown).
Tests: CssUnlocker matches only DiscKind::Css; a BD-profile drive yields
NotApplicable with zero CSS CDBs issued.
Reshape the Unlocker seam so every unlocker is dispatched at ONE place from
ONE ordered registry — the firmware, cert, and CSS routes are all "remove the
bus-encryption barrier", differing only in what they key off. matches() and
unlock() now take an UnlockCtx { drive_id, kind: DiscKind } instead of a bare
DriveId: a firmware unlocker keys off drive_id (kind irrelevant), the cert
unlocker will match DiscKind::Aacs, the CSS unlocker DiscKind::Css. UnlockCtx
is #[non_exhaustive] so a host-cert source can be added without breaking
external unlockers. Drive-prep dispatch passes DiscKind::Unknown (no disc
probed yet); the cert/CSS registry impls + the single post-probe dispatch
point follow in subsequent commits.
Thread &mut dyn ScsiTransport through unlock_css_reads and its bus-auth /
disc-key primitives instead of &mut Drive — they only ever issued SCSI via
drive.scsi_mut(), so this is mechanical and lossless, and it makes the CSS
unlock speak the same transport interface as the registry Unlocker trait
(prerequisite for CSS becoming a uniform unlocker).
Also clean up the badly-named, half-dead bus_auth:
- rename bus_auth -> establish_authenticated_session: it is run for its
side effect (sets the drive's ASF=1, unlocking scrambled-sector reads),
which the name now states.
- drop the derived CSS bus key from the return + computation: it had no
consumer (descrambling is keyless via the Stevenson attack), so it was
dead crypto computed on every DVD unlock. Return just the negotiated AGID,
which the caller genuinely needs for the best-effort disc-key REPORT KEY.
The AACS cert-auth primitives (aacs_authenticate, the AACS 2.0 P-256
variants, read_volume_id, read_data_keys) and their scsi_read/scsi_write
helpers touched the drive ONLY through Drive::scsi_execute — a pure
pass-through to the transport. Thread &mut dyn ScsiTransport instead of
&mut Drive so these primitives are transport-level, matching the firmware
Unlocker seam (which hands out &mut dyn ScsiTransport for testability).
Pure mechanical signature change, no logic change; the cert orchestrator
(do_handshake_cert) keeps &mut Drive for the OEM-VID shortcut and passes
session.scsi_mut() into the primitives. Step toward making the cert
handshake a uniform registry unlocker.
Generalize the unlock seam toward uniform unlockers (firmware / cert / CSS).
`unlock` now reports what it LEARNED — a Volume ID and/or an AACS bus key —
instead of just a Vid; libfreemkv files those onto the disc/drive in one
place, so a plugin never touches Disc. The hardware side-effect (extended
mode) still happens inside unlock. route_unlock walks the registry in
registration order (the single ordered place that decides run order); no
priority field. Firmware route reports {vid: Some, read_data_key: None}.
The bus-key gate only credited the cert handshake's read_data_key as proof
bus encryption was removed. A firmware unlocker removes it AT THE DRIVE
(serves clear content) and yields no read_data_key — so a SUCCESSFUL
firmware unlock (VID present, read_data_key None) tripped the gate and
blocked ALL key resolution, including the online source. That was the
root cause of live UHD discs reporting "missing keys" after an unlock.
Now a single predicate answers "is bus encryption gone?": never-had-it ||
file/ISO || firmware-unlocked || cert-bus-key. The gate is just
`if !bus_encryption_removed { error }` — no enumerated cases. HandshakeResult
gains `drive_unlocked`, and the read_data_key failure reason is captured so
the warn says WHY the bus key is missing.
Also: reword the first hardware-sense escalation as "fast-fail escalation"
(it is often transient — the drive recovers), reserving "wedge" for a
persistent run; and scrub the product name from core comments (it belongs
only in the unlocker crate).
When a bus-encrypted disc's handshake yields no read_data_key, the gate
logged a bare "bus_key_unavailable" with no indication of WHY — turning
every occurrence into archaeology (is the bus key not-attempted, or did
the read fail?).
Now:
- read_data_keys failure is captured (error code) instead of swallowed by
`.ok()`, and logged at the handshake with its consequence.
- HandshakeResult carries `read_data_key_err: Option<u16>` so the gate
distinguishes "never attempted" (None — VID-only/OEM path) from "read
FAILED" (a code), and the bus_key_unavailable warn now reports the code
plus whether a Volume ID was present.
No behavior change — purely diagnostic. The handshake_ok debug also now
reports has_volume_id.
Two defensive hardenings from the post-fix audit (vs FFmpeg/GStreamer):
1. Move the `pes.discontinuity` partial-drop ABOVE the empty-data guard in
all three audio parsers (ac3/dts/truehd), so a discontinuity signal can
never be stranded by an empty post-gap PES. The demuxer only emits
non-empty PES today; this is defense-in-depth for any future caller.
2. A PES with no PTS must not reset the timeline to 0. ac3 now carries
`flush_pts_ns`, dts continues from the most recent known base; truehd
already kept its running cadence on a None PTS. Matches OSS behavior
(PTS rebases off the next PES that actually carries a PTS).
Adds an ac3 regression test (empty-payload discontinuity PES still drops
the stranded partial). Loss accounting was reviewed: TS-demux CC-gaps are
NOT counted toward lost_video_secs / abort (that is sector-based via
DiscStream::errors / mapfile bytes_unreadable), so a source splice never
inflates loss — no gating needed there.
The AC-3, DTS and TrueHD parsers buffer access units across PES
boundaries. At a concealed-loss gap the buffered unit is truncated:
splicing post-gap bytes onto it manufactures a corrupt frame on top of
the real loss (FFmpeg "Failed to decode block code(s)" / "Invalid data
found" at the gap) and, for TrueHD, strands the PTS cadence into the
non-monotonic audio-DTS band seen on multi-clip titles.
The video parsers already handle this via the ResyncGate, but the
discontinuity signal was only wired into video — audio parsers ignored
pes.discontinuity and spliced across the gap.
Now, when pes.discontinuity is set, each audio parser drops the partial
(clears buf, and for DTS its PTS marks / pending base) so the post-gap
PES re-bases a fresh unit. A lost gap degrades to a clean single-frame
drop instead of a corrupt spliced frame. No effect on perfect rips: the
branch only runs when concealment inserted a discontinuity marker.
Adds a per-parser test feeding a partial frame then a discontinuity PES,
asserting the truncated partial is dropped (not spliced) and the post-gap
PTS is adopted.
Pass N (patch) pinned the slow recovery speed (0x0000) for the whole pass.
But Pass 1's damage-jump overshoots, so most of a jumped range is clean data
the reverse-walk reads first — grinding it at slow speed wastes minutes per
gap. Now each range mirrors Pass 1's model:
- Enter at max speed (0xFFFF) + reset to the initial batch (current_batch
carries across ranges, so a prior single-sector grind would otherwise start
the next range slow). Reads the clean overshoot fast.
- First read failure in the range → drop to 0x0000 (once; idempotent SET CD
SPEED) for the rest of that range — grind only the genuine damage.
- 10 s halt-responsive cooldown between ranges, gated on "this range actually
grinded" so a many-small-range pass (100+ ranges) doesn't stall ~20 min on
unconditional pauses.
No change to good/bad classification, the mapfile, or the abort math — purely
read speed + an inter-range cooldown. Synthetic speed-transition test to follow.
Non-blocking follow-ups from the 1.2.0 audit:
- conceal loop: if decrypt reported loss but the padding-aware predicate
matched nothing to conceal (a ~256^-31 contradiction), fall back to the
strict predicate and conceal whatever it flags, loudly — belt-and-
suspenders so ciphertext can never reach the mux.
- decrypt_dropped doc: reflect 1.2.0 (mux-path loss is concealed + tallied,
not silently dropped).
- direct unit test for aacs_unit_still_ciphertext (the padding-aware
conceal predicate): clear/all-zero/full-decrypted/full-ciphertext/
decrypted-short-tail.
- fix three stale "v1.1.1" comment refs (the fragment-tail fix ships in
1.2.0; there is no v1.1.1 release).
Closes the three residual holes where a concealed/lost gap could still let
a dangling-reference frame reach the muxer (degraded/undecryptable-disc
path only; clean rips are byte-identical and untouched). Root cause: the
discontinuity signal was reconstructed from the 4-bit continuity counter
and applied per-PES, both of which are lossy.
Three coordinated changes:
1. CC-INDEPENDENT marker. fill_null_ts_unit now tags its NULL packets with
an adaptation-field discontinuity_indicator; the demuxer recognises a
0x1FFF packet carrying it as a concealed gap and forces a discontinuity
on every tracked PID (the lost unit's PID is unknowable). This survives
a loss that is an exact multiple of 16 packets (CC aliases to in-sequence
— hole 3) and a loss at a PID's very start (no prior CC — hole 4); it
also drops any open, potentially-truncated partial PES.
2. PUSI ATTRIBUTION. A gap landing on a PES boundary now flags the PES
STARTING after it, not the one flushed at the boundary (hole 1) —
stamping the pre-gap frame could arm-then-disarm the gate on a keyframe
and admit the real post-gap inter frame.
3. PER-FRAME signal. codec::Frame gains `discontinuity`; each parser
propagates it onto the first post-gap frame. MPEG-2 buffers whole GOPs
asynchronously, so it associates the gap by ES OFFSET (like PTS/source),
landing it on the exact post-gap picture mid-GOP (hole 2) — a per-PES
flag stamped the previous picture. consume_ts (and the EOF flush drain)
gate on frame.discontinuity.
Tests: CC-independent marker with in-sequence CC + leading-loss; PUSI
attribution flags the post-gap PES; MPEG-2 offset-mark stamps the post-gap
picture through GOP reorder, not the previous one. Existing B1 gate + EOF
tests still green (2270 lib tests).
Per the release decision: ship the AACS-version-stride API changes under
the 1.2.0 minor bump, with an explicit Breaking section (DiscInputs new
field, DiscInputsCtx::new arity, read_aacs_inputs* 3-tuple return).
The P3 concealment loop in DecryptingSectorSource::read_sectors keyed on
aacs_unit_needs_decrypt, whose sync check is the majority-vote
ts_sync_destroyed (<=16 of 32 syncs). A successfully padding-aware-
decrypted content-fragment TAIL unit (e.g. 11 content packets + 21 zero
padding) has only 11 syncs, so the majority vote called it "still
encrypted" — and when such a good unit shared a read buffer with a
genuinely-undecryptable one (dropped>0), the loop overwrote the GOOD
decrypted tail with NULL-TS, silently discarding correct video and
over-counting concealed units vs the tallied dropped bytes.
Add aacs_unit_still_ciphertext (padding-aware): encrypted AND at least
one non-zero (non-padding) 192-byte packet missing its 0x47 sync — the
same discriminator decrypt_unit uses to accept a fragment tail. The
conceal loop now uses it, so only genuinely-unrestored ciphertext is
concealed. Full and fully-ciphertext units are unchanged. Regression
test: a decrypted short-padding-tail co-resident with a failed unit is
left byte-for-byte intact while the failed unit is concealed.
Pairs with A2 (read-path NULL-TS concealment). When the demux assembler
sees a TS continuity gap it now stamps `discontinuity` on the next
completed PES; the codec-parse stage carries that onto a per-track
ResyncGate. After a gap on an inter-coded video track the gate drops
forward to the next IRAP/IDR keyframe so no frame with a dangling
reference reaches the muxer (an ffmpeg deep scan would otherwise report
a missing-reference / non-existing-PPS error). Audio and subtitle tracks
have no cross-frame references, so the gate is a no-op there.
- ts.rs: PesPacket gains `discontinuity`; PesAssembler tracks a sticky
pending_discontinuity flag set on CC gap / discontinuity_indicator and
carried to the next completed/flushed PES.
- resync.rs (new): ResyncGate — per-track arm-on-gap, drop non-keyframes
until the next keyframe disarms and resumes. Logs the resync + drop
count once at the keyframe.
- pipelined_stream.rs: precompute per-track is_video, apply the gate in
consume_ts. Out-of-range track index emits as-is (defensive).
Tests: ResyncGate unit tests; ts.rs gap-stamps-discontinuity; end-to-end
B1 video-drops-to-keyframe and audio-never-drops through PipelinedPesStream.
Decrypt-verify is a RIP gate, not a MUX gate. On the mux read path an
undecryptable content unit must never abort the mux:
- DecryptingSectorSource gains tolerate_decrypt_loss(): when set, an
undecryptable in-content unit is tallied, overwritten with valid NULL
TS packets (PID 0x1FFF) via aacs::fill_null_ts_unit, logged loud with
its LBA, and the read returns Ok — the stream keeps flowing. The rip
paths keep the fail-loud DECRYPT_VERIFY_READ decorator (re-read off the
disc); only the mux opts in.
- Wire it into both mux read paths: the file-backed highway
(build_iso_pipeline) and the inline DiscStream.
- NULL-TS fill keeps the demuxer byte-synced on the 192-byte stride; the
lost video/audio PID packets surface as a CC gap the TS assembler
already drops a partial PES on (the B1 foundation). Ciphertext is never
passed downstream either way.
- Fix stale resolve_vid_only no-cert test: default is UHD (audit #4).
Tests: conceal-as-NULL-TS, fill well-formedness, fail-loud still holds.
- One workspace hex parser (libfreemkv::hex): the keydb / online / mapfile
parsers had divergent prefix handling (0x vs 0X vs none) — a key written with
a prefix one parser didn't expect was silently dropped. All three now call it.
- read_encrypted_units probes several points spread across each extent instead
of only the midpoint-forward window, so a late-starting / sparse encrypted
body still yields samples (empty samples make decrypt_with skip wrong-key
validation). A read error at one probe no longer abandons the extent.
- read_aacs_inputs* now returns the AACS major version; DiscInputs carries it,
and DiscInputsCtx parses Unit_Key_RO.inf at the disc's own stride (fixes the
hardcoded-V20 read-time fetch for V10 discs). One source of truth, no version
argument to drift.
- Disc::inputs() is the single complete AACS-input source (inf/MKB/VID/hash/
version); the out-of-band duplicate readers go away.
- Named constants for AACS file paths (aacs::PATH_*) and the AACS majors
(aacs::AACS_MAJOR_*, AacsVersion::major/from_major) replace magic strings/ints.
- push_ranges saturating (corrupt-disc panic guard).
- decrypt_unit: padding-aware acceptance — recover real video at content-
fragment tails (the phantom mux-loss class) without weakening wrong-key
rejection (a full content unit still needs all 32 TS syncs).
- scan: read the MKB via the bounded read_mkb_content so Disc::inputs()
carries it. Online key resolution was shipping mkb=0 (a full read of the
~128 MiB MKB_RO allocation fails) → the decode service 404'd.
- resolve_vid_only: surface an MKB read error instead of silently emptying.
- fetch: a per-sample dry-set replaces the global fetch_spent latch, so a
second CPS unit's key can still be fetched after the first came back empty.
- verify::push_ranges: saturating arithmetic (corrupt-disc panic guard).
- Tests for all of the above.
ContainerKind {Ts,Ps} + ClipLayout.container thread the post-decrypt structural check per clip; decryptability() dispatches it (TS: unit_is_clean_ts, PS: unit_is_clean_ps). New decrypt_unit_checked(unit,key,accept) decouples the container-agnostic AACS crypto from the format-specific acceptance (decrypt_unit delegates with the TS check). unit_is_clean_ps is the MPEG-2 PS pack-start check, documented UNVALIDATED for HD-DVD (.evo unit/seed/pack alignment must be confirmed on real media). clip_layouts assigns Ts today; .evo->Ps is the one-line HD-DVD hook.
reverify_iso now takes an is_finished predicate and SKIPS any unit with a non-Finished backing sector: we can't verify what wasn't read (a non-Finished sector is zero-filled because the drive read failed there), and must never waste a key lookup on a block the read already knows is bad. observe() (sweep) was already safe (only fed Good bytes).
Post-read verify gate (new src/disc/verify.rs): UnitVerifier buffers/aligns the disc-absolute read stream into clip-file 6144-byte units, then makes one decryptability() decision per unit (CPI gate -> held keys -> key_fetch -> strict TS). POST_READ_VERIFY const kill-switch; fail-safe contract (only ever downgrades units it is confident are undecryptable; every doubt skips). Hooked into Disc::sweep (producer observes ciphertext -> WorkItem::MarkBad after the Good, FIFO-ordered) and Disc::patch (post-loop reverify_iso reads recovered units whole from the patched ISO). extract::clip_layouts enumerates AACS clips for the gate.
Standards-correct AACS verify: aacs::unit_is_clean_ts is a strict port of libaacs _verify_ts (all 32 TS syncs, not a majority vote); decrypt_unit accepts a key only on it; the majority verify_ts is removed. Deleted the Disc::verify_clips post-pass bolt-on (its primitive is absorbed by the read-path gate).
libaacs/DVD audit fixes: content-cert bus_encryption flag now read from bit 7 (was bit 0 - defeated the bus-key fail-loud gate); cc_id read from offset 14; title_cps_unit range-validated + 1->0 index-converted per libaacs. Corrected attack_crib ("functionally-equivalent" not "exact" port) and read_disc_key (READ DVD STRUCTURE 0xAD, not REPORT KEY) doc comments.
Also includes accumulated uncommitted work: key-fetch seam and TrueHD/DTS audio fix.
Use 'Semantics / reference' as the last-column header across all spec
tables, and break enum/option/note cells onto separate lines with <br>
for readability. Formatting only; no semantic changes.
The CLI --version, the MKV muxing/writing-application field, and the FVI generator
all derive from one libfreemkv const, so a binary reports the exact same label it
stamps into the files it produces — no split-brain where an MKV claims one version
and the binary another.
Reformat all three JSON code blocks (section 8 ext example, Appendix A
Header schema, Appendix A Record schema) with consistent 2-space
indentation and valid syntax. No field or semantic changes.
build.rs emits FREEMKV_VERSION = FREEMKV_BUILD_LABEL when that env is set
(non-empty), else the Cargo package version. The muxing/writing-application
field and the FVI generator tag use it, so a pre-release/test build can be
stamped without bumping Cargo.toml and disturbing the tag-pinned [patch]
version matching. The git short hash is still appended either way.
TrueHD: when the PES PTS lags the access-unit cadence, resync to the PTS
but never snap the running timestamp backward, so the emitted DTS stays
monotonic across the resync (next_pts_ns = max(next_pts_ns, pts)).
DTS-HD MA: size each EXSS extension substream exactly from its header
(exss_frame_size) and skip it as a unit, so a false 0x7FFE8001 core sync
inside the lossless extension payload can no longer split the access unit
and truncate the extension. Falls back to a bounded scan when the header
is unparseable.
Provenance: build.rs bakes the git short hash into GIT_SUFFIX; the muxing/
writing-application field and the FVI generator tag now record the exact
build (e.g. "freemkv 1.1.0-beta.1 (g835cc99)"), so any output file is
traceable to the revision that produced it.
keydb I/O moves out of libfreemkv into freemkv-keysources
(KeydbSource::save / ::update). Delete src/keydb.rs entirely (save,
http_get, default_path, write_atomic, UpdateResult) and drop `pub mod
keydb;` — http_get had no real callers. The shared Keydb* Error variants
stay in error.rs (keysources raises them; the every-error-has-a-code
contract depends on them). flate2 is no longer used here, so drop it
from Cargo.toml (zip stays for labels/jar.rs).
CHANGELOG: note the keydb-I/O move; reword the rc.5.2
DefaultDecodedFieldDuration entry to state only the action taken (the
revert) rather than an unverified Windows-fps outcome.
Surgical fixes (each with a regression test that fails without the change):
mux/mkv.rs, mux/demux_sink.rs: drive the clip-boundary timeline epoch
off the resolved PRIMARY VIDEO track, not the literal stream index 0.
An M2TS/PMT title can list an audio ES before video, so streams[0] may
be audio; a non-video epoch driver ratchets the frontier and inflates
the timeline. mkv cluster-opening falls back to track 0 for audio-only
titles so they still open clusters.
mux/codec/ac3.rs: correct ACMOD_CHANNELS — acmod=5 (3/1) is 4 channels,
not 3 (was undercounting a 3/1 stream); fix the A/52 Table 5.8 doc.
disc/mod.rs: HDMV coding_type 0x91 (Interactive Graphics / menus) no
longer maps to PGS subtitle — it falls through to Unknown so the PMT/STN
walker drops it instead of surfacing a bogus subtitle track.
mux/videomap.rs + mux/mkv.rs: FVI colour now mirrors the MKV muxer's CICP
precedence (measured CICP authoritative; HDR-driven PQ/HLG transfer
override) via a shared cicp_for_video helper, so the two sinks can't
disagree (HDR10 BT.2020 no longer emits SDR transfer 14).
mux/mkvstream.rs: saturating_add on cluster_ts + rel_ts so an adversarial
CLUSTER_TIMESTAMP near i64::MAX can't overflow/panic before the existing
saturating_mul.
mux/timeline.rs: tighten the tail-straggler clamp so a normal new-epoch
non-video frame leading the sparse video frontier by >3s is not demoted
into the previous clip's epoch.
mux/m2ts_mux/mod.rs: re-stamp PCR per video TS packet (mid-PES), not only
at PES boundaries, so a large UHD I-frame can't open a multi-second PCR
gap; modular 33-bit PTS rebasing so a real 90 kHz clock wrap is not
collapsed to PTS 0 (pre-base frames still floor to 0).
io/byte_prefetcher.rs, sector/prefetched.rs: wrap the producer feed loop
in catch_unwind and emit a typed error sentinel on panic, so a mid-stream
producer panic is not read as a clean EOF at the demux boundary (which
would silently truncate the mux).
mux/codec/h264.rs: extend HIGH_PROFILES to the full ISO/IEC 14496-15 set
that mandates the avcC chroma/bit-depth extension (adds 244 et al.).
Doc/comment accuracy: css/mod.rs (50000 sectors, not scrambled-sectors),
aacs/decrypt.rs (decrypt_unit already-clear path), ifo.rs (TT_SRPT at
0xC4), css/lfsr.rs (LFSR0 24-bit; TAB1-then-XOR cipher; real scramble-flag
predicate), disc/read_error.rs (for_sweep does bounded transient retries).
Skipped: keydb.rs SSRF guard (low/latent, no live caller) — a hard
loopback block breaks an existing behavioral test that exercises the
header-EOF path over a loopback server; a clean fix needs a resolver test
seam beyond this surgical pass. The sibling keydb_fetch.rs comment fix is
out of scope (freemkv crate).
Parse the two HDR10 HEVC SEI messages and emit the corresponding
Matroska Colour metadata, only when actually present in the bitstream
(SDR / no-SEI tracks omit it; nothing is fabricated).
Parse (Rec. ITU-T H.265 Annex D):
- Mastering Display Colour Volume SEI, payloadType 137 (D.2.28):
display_primaries_x/y[3] (SEI order G,B,R), white_point_x/y
(0.00002 units), max/min_display_mastering_luminance (0.0001 cd/m²).
- Content Light Level Info SEI, payloadType 144 (D.2.35):
MaxCLL / MaxFALL (cd/m² integers).
HevcParser::scan_sei walks the sei_rbsp ff-extension payloadType/
payloadSize coding and de-emulates (00 00 03) before reading, reusing
the existing strip_emulation_prevention helper. Both SEI are required
before any metadata is surfaced; SEI NALs still pass through unchanged.
Carry: the measured Hdr10Metadata rides PictureInfo (the same per-coded-
picture seam FieldOrder uses), flowing through from_codec_frame onto
PesFrame.coding to the deferred-muxer activate path, where
apply_coding_to_track stamps it on the video track before the header is
written. Set only when both SEI were seen.
Emit (RFC 9559 / Matroska): new Colour children in ebml.rs
(MasteringMetadata 0x55D0, Primary R/G/B + WhitePoint chromaticity
0x55D1..0x55D8, Luminance max/min 0x55D9/0x55DA, MaxCLL 0x55BC,
MaxFALL 0x55BD). write_hdr10 converts chromaticity SEI int × 0.00002 →
Matroska float, luminance SEI int × 0.0001 → cd/m² float; MaxCLL/MaxFALL
are uints verbatim. SEI primary index 0/1/2 (G/B/R) mapped to the
Matroska R/G/B element layout. Emitted only when hdr10 is present.
Tests: SEI parse with exact raw values, requires-both-SEI, SDR omission,
and emulation-prevention stripping (hevc.rs); muxer emit with exact unit
scaling + SDR omission of MasteringMetadata/MaxCLL/MaxFALL (mkv.rs);
apply_coding_to_track HDR10 plumbing (mkvstream.rs).
Owner-flagged invariant: freemkv -t 1 ALWAYS selects the main feature because the
CLI's title 1 maps to titles[0] and the list is ordered by canonical_title_order
(main feature first). Adds a DVD-shaped contract pin asserting titles[0] is the
movie after sorting (a regression there is a title-ordering bug, not a remux issue),
complementing the existing branching-UHD / normal-disc order tests.
Also clears two pre-existing test-only clippy nits surfaced under --all-targets:
unused `lba` in ClearStubReader::read_sectors, and an unneeded `mut` on the
h264 population test's closure.
mux/mkv: assert emitted CODEC_PRIVATE bytes verbatim for H.264/HEVC/VC-1/
MPEG-2 (direct TrackEntry child, not nested in Video) and DefaultDuration ns
for all eight frame rates, read back out of a real MkvMuxer.
disc/sweep: end-to-end Disc::sweep against a synthetic MockReader with an
injected bad region, asserting the resulting mapfile marks the clean lead
Finished and the failed batch + zero-filled skip-ahead gap NonTrimmed,
proving the Pass-1 damage-jump engaged.
disc/patch: introduce a minimal clock seam (fn() -> Instant on the internal
PatchLoopState, defaulting to Instant::now) so the per-range and whole-pass
watchdogs are deterministically testable; public API and callers unchanged,
production behavior identical. Add tests that advance a fake clock to trip the
range budget and whole-pass stall predicate.
aacs: add an AES-128-CBC known-answer test for aes_cbc_decrypt using the
published NIST SP 800-38A F.2.2 vector (blocks 1..3 exact; block 0 via the
documented fixed-AACS-IV substitution).
Port + adapt the freemkv native per-picture video index (FVI) from the
old feat/fvi-sink branch onto rc6's codec-agnostic PictureInfo model.
This is a surgical adaptation, not a merge.
Adaptations (fvi_sink.rs, videomap.rs, tests/fvi_pipeline.rs):
- Retarget from the removed crate::mux::codec::mpeg2::PictureInfo (raw
public fields) to rc6's authoritative crate::mux::codec::PictureInfo
in codec/coding.rs, via its accessors.
- type from coding_type() -> CodingType{I,P,B}; emitted for ANY frame
that carries coding (every video codec now fills it), with the
keyframe-flag I/P fallback only when coding is absent.
- Replace the mpeg2-only tff/rff/progressive members with codec-agnostic
members derived through the accessors: field_order (tff/bff/progressive)
and progressive, emitted ONLY when the codec measured the signal
(Option::Some) and omitted otherwise; plus nb_fields.
- Test fixtures rebuilt via PictureInfo::mpeg2(CodingType, Mpeg2Coding{..})
/ coding_type_only(..); added measured_cicp: None to VideoStream
literals for rc6's struct.
Honesty decision (key / random-access):
- The codec-agnostic PictureInfo carries NO GOP-closure (no closed_gop /
gop_start), so key is set from the frame's intra / decode-restart flag
(frame.keyframe == coding.keyframe() for video), NOT a fabricated
clean-RAP claim. The old gop member is honestly omitted. FVI_FORMAT.md
is updated to document this as a limitation: key is an intra picture /
parser-flagged decode-restart point; MPEG-2 open-GOP clean-RAP precision
(closed_gop) is not currently distinguished. §7.1 rewritten for the
new field_order/progressive/nb_fields members.
Wiring:
- mux/mod.rs: pub(crate) mod fvi_sink; pub(crate) mod videomap
(#[allow(dead_code)] on videomap — the VideoMap accumulator is staged
for side-channel reuse, sink builds records directly); pub use
fvi_sink::FviSink.
- mux/resolve.rs: add the fvi:// output scheme to StreamUrl, parse_url,
scheme(), path_str(), input() (write-only reject) and output()
(constructs FviSink), mirroring the mkv:///demux:// patterns.
Provenance fix surfaced by the end-to-end test:
- pipelined_stream::consume_ps was dropping the PS demuxer's byte-exact
source stamp (source: None) when rebuilding PesPacket, so PS/DVD-path
frames reached the mux/index with no provenance (FVI src null). Carry
ps.source through, matching the TS path; the real-pipeline fvi test now
sees the stamped src sectors.
Gate: cargo +1.86 fmt + clippy --lib -D warnings clean; cargo +1.86 test
--lib (2182 passed) and --test fvi_pipeline (2 passed); precommit.sh
libfreemkv green.
Honest PictureInfo population for VC-1 — completes the set (MPEG-2, H.264, HEVC,
VC-1 all populate now).
- Read the advanced-profile sequence header INTERLACE flag (SMPTE 421M §6.1.1,
bit 41), de-escaping emulation-prevention bytes as parse_vc1_resolution does.
- For a PROGRESSIVE sequence, decode the picture-layer PTYPE VLC (0=P, 10=B,
110=I, 1110=BI→I, 1111=Skipped→P) from the first bits after the frame start
code. For INTERLACED (FCM/FPTYPE precede PTYPE) or simple/main/unknown, decline
→ coding stays None: honestly absent, never read at a wrong bit offset.
- Set coding = coding_type_only(...) and source = pes.source; field order is not
decoded, so field_order() is honestly None.
- Reuses the shared startcode::BitReader (no new primitive).
- Tests: progressive I/P/B from real PTYPE VLCs + source carry + field-order
absence; interlaced honest-decline.
Honest PictureInfo population for HEVC, on the same principle as H.264.
- Consolidate the bit reader: hevc.rs had its own BitReader (used by the SPS
parser); h264 just gained one in startcode. Promote startcode's to the single
shared pub(crate) reader (adds read_bits) and delete hevc's copy — one proven
primitive, reused (SPS parse + both slice-type decoders).
- hevc: decode slice_type from the first coded slice's slice_segment_header
(H.265 §7.3.6.1) → I/P/B (§7.4.7.1). The offset to slice_type depends on
num_extra_slice_header_bits, which lives in the PPS — so we parse it from the
ACTIVE PPS (§7.3.2.3) and only measure slice_type when that PPS is known.
With no active PPS we decline rather than guess: coding stays None, honestly
absent. Set coding = coding_type_only(...) and source = pes.source; field
order (pic_struct SEI) is not decoded, so field_order() is honestly None.
- Tests: I/P/B from real slice headers, source carry, field-order absence, and
the no-PPS honest-omission case.
Honest PictureInfo population for H.264 — the foundation that lets the video
index claim a frame's type and be believed (a faked "P" on a B-frame would make
the index confidently wrong).
- startcode: shared minimal MSB-first BitReader with Exp-Golomb ue(v), reused by
H.264 (and next HEVC). Documents the emulation-prevention caveat: only the
leading slice-header fields are read, where 00 00 03 cannot intervene.
- h264: decode first_mb_in_slice + slice_type (H.264 §7.3.3) from the first
coded slice and map to I/P/B (§7.4.3 Table 7-6; SP→P, SI→I). Set
coding = PictureInfo::coding_type_only(...) and source = pes.source. Field
order is NOT decoded here, so field_order() stays None — honestly absent,
never guessed.
- Tests: I/P/B from real ue-encoded slice headers, source carry, field-order
absence; BitReader Exp-Golomb table + truncation.
Carry per-picture truth and byte-exact source provenance THROUGH the stream so
the muxer (and the upcoming video index) read MEASURED facts instead of
assuming them. Honest data in, honest data out.
- codec/coding.rs: codec-agnostic PictureInfo (CodingType / FieldOrder + the
accessors field_order/coding_type/nb_fields/progressive/keyframe). Each codec
folds its raw signals in; consumers use only accessors, never branch on codec.
- mpeg2: builds PictureInfo from the picture coding extension and carries it +
SourcePos (source_marks, parallel to pts_marks) on every emitted frame.
- pes / codec::Frame: additive `coding` + `source`, forwarded through the
highway; None for audio/subtitle and the network/stdio deserialize hop.
- mkvstream: DEFER muxer construction until the first coded picture, set the
video track's FieldOrder from the MEASURED value, THEN write the header —
right the first time, no guess, no seek-back. An interlaced track that arrives
with no measured order is LOGGED loudly and left UNDETERMINED, never faked.
- mkv: MkvTrack::video no longer guesses TFF (a bitstream property the scan
cannot know is UNDETERMINED at build). Removed VideoStream::top_field_first
(the dead scan-time guess) crate-wide.
- Tests: parser population (every PictureInfo facet + per-PES source carry) and
mux-stream consumption (measured -> correct; missing -> UNDETERMINED, not
faked). Two obsolete tests updated only after confirming (their own comments)
they existed to enforce the deleted hardcoded-TFF.
- FieldOrder now derives from the bitstream's measured top_field_first
(Some(true)→TFF, Some(false)→BFF) instead of hardcoding TFF for all
interlaced content; falls back to TFF when unmeasured. Adds
VideoStream::top_field_first; DVD/BD scan sets None with a precise
TODO(spec) for parser→title plumbing.
- CICP (matrix/transfer/primaries/range) now prefers measured CICP from
the bitstream (VideoStream::measured_cicp) over the coarse ColorSpace
enum, so the container stops assuming a colour space the stream may
contradict. Enum remains the fallback.
- VobSub S_VOBSUB CodecPrivate now emits a `size: WxH` line ahead of the
palette per the .idx format so players place/scale subs correctly.
- SeekHead: when zero cues are written, the CUES Seek entry is Voided
instead of leaving a dangling pointer to the Cues offset (now Tags/EOF).
- AC-3 Channels back-patch offset is captured from the writer instead of
the hardcoded chan_elem_pos+2 (decoupled from the VINT width choice).
- Hoisted inline CICP codes and the dvcC fourcc to named constants citing
ITU-T H.273 / RFC 9559; fixed the stale FieldOrder comment.
- DefaultDuration vs pulldown: precise TODO(spec) left (needs the same
parser→title channel as top_field_first).
Tests: BFF-from-measured-flag, measured-CICP-overrides-enum, VobSub size:
line present/omitted, zero-cue SeekHead Void. precommit (1.86) green.
Delete re-implementations in the demux:// sink and wire to proven helpers;
keep only genuinely-new functionality.
- AnnexB reframing: delete the sink's local length_prefixed_to_annexb (it
break'd on a zero-length NAL, dropping the rest of the access unit) and
call the canonical append_length_prefixed_as_annex_b in mux::hevc, which
skips just the empty NAL.
- HEVC param sets: delete hvcc_param_sets; reuse hvcc_to_annex_b.
- avcC param sets: hoist as the new canonical avcc_to_annex_b in mux::hevc,
next to hvcc_to_annex_b (the symmetry point); the sink calls it.
- PGS .sup: emit a synthetic clear display set (empty PCS + END) at
pts + duration_ns so subtitles time out instead of lingering to EOF.
- TimelineContinuity: move verbatim into the shared mux::timeline module
(with the prev_offset straggler-remap intact) and use it from both the
MKV muxer and the demux sink; delete the sink's drifted TimelineRebase
copy (which lacked the straggler branch).
- VobSub .idx: emit the conventional 'id: <lang2>, index: 0' line mkvmerge
reads to assign the subtitle language; palette reuse unchanged.
- output(): seed DemuxOptions.base from title.playlist when non-empty.
New constants for the PGS clear-segment framing and avcC header cite the
public HDMV PGS (BD-ROM Part 3) and ISO/IEC 14496-15 specs.
Tests: a zero-length NAL mid-frame no longer truncates the AU; a frame with
duration_ns produces a .sup clear segment; existing demux tests stay green.
The DVD seek-index concern is that the mux highway
(PsDemuxer -> PipelinedPesStream codec parse -> frame out -> muxer)
might drop the keyframe flag or per-frame duration that Mpeg2Parser sets
on each Frame, which would stop the muxer's cluster/cue open
(keyframe && track 0) from ever firing and leave a DVD MKV with many
clusters and zero cues.
Add dvd_highway_preserves_video_keyframe_and_duration: drives a real
Mpeg2Parser through PipelinedPesStream via DemuxBatch::Ps batches (one
PTS-stamped MPEG-2 PS video PES per GOP, decode-order I + P/B), reads the
frames back through the highway's read(), and asserts the GOP-opening
I-frames arrive as keyframes, every frame keeps its duration (BlockGroup
path), and video routes to track 0. Guards from_codec_frame's
keyframe/duration propagation across the demux-thread + recycled-buffer
highway, which the prior cue tests (codec parser straight into the muxer)
did not exercise.
Verified against the real current-VFR-build output
(Greenland-feature.mkv, today): 3365 clusters / 3365 CuePoints, all
BlockGroups, cues resolve to clusters and track the video — the highway
preserves the flags and the seek index is complete.
The existing cue tests (cue_count_equals_cluster_count,
cue_positions_resolve_to_clusters, cue_times_match_cluster_timestamps)
all feed frames with duration_ns=None, so they exercise only the
SimpleBlock write path (UHD/HEVC). DVD MPEG-2 video is now VFR: every
coded picture carries a per-frame duration_ns=Some(..), so it is written
as a BlockGroup, not a SimpleBlock. That cue path was untested.
Add cue_count_equals_cluster_count_blockgroup_vfr, which drives the real
Mpeg2Parser end-to-end (decode-order frames, non-monotonic B-frame
display PTS, telecine field durations) into the muxer and asserts the
output is wholly BlockGroup (no SimpleBlock) and that the Cues index has
exactly one cue per cluster, with every cue resolving to a real cluster.
Guards the DVD seek index (scrub/fast-forward) against regressing to the
chapter-seek-only, zero-cue state.
New write-only pes::Stream sink that taps the per-track PesFrame stream
(the seam right before MKV muxing) and writes each track to its own
elementary-stream file, plus chapters and per-audio-track delay metadata.
Purely additive — the MKV mux path is untouched.
- mux/demux_sink.rs: DemuxSink + EsWriter dispatch. Pass-through for
codecs whose Frame.data is already standalone ES (MPEG-2, VC-1, AC3/
E-AC3, DTS/DTS-HD, TrueHD, LPCM). Non-trivial writers:
- AnnexBWriter: reframes hvcC/avcC 4-byte-length-prefixed NALs to
Annex-B and prepends VPS/SPS/PPS parsed out of the codec_private
configuration record (HEVC .hevc / H.264 .h264).
- PgsSupWriter: rebuilds the HDMV 'PG' segment framing the parser
strips, with 90kHz PTS/DTS (.sup).
- VobSubWriter: writes raw SPUs to .sub and synthesizes the .idx
sidecar (palette + per-SPU timestamp/filepos).
- Delay-in-filename (mkvmerge-readable 'DELAY <n>ms') + chapter XML/OGM
export. TimelineRebase ports the MKV muxer's seamless-branch epoch
logic so per-track ES timestamps stay continuous across clip joins.
- mux/resolve.rs: StreamUrl::Demux variant + scheme/path_str/parse_url/
input(write-only)/output arms.
- mux/mod.rs: module + public type re-exports.
16 unit tests: Annex-B reframing, hvcC/avcC param extraction, delay
sign/rounding + mkvmerge-regex match, PGS .sup framing, VobSub .idx
synthesis, chapter XML/OGM, timeline rebase, and end-to-end file-keying
by track + track selection.
A genuinely-clear or uncrackable extra title (a tiny menu/nav stub) no
longer poisons a multi-title rip with a false CssKeyMissing (E7023).
- decrypt_keys_for_title_checked: re-crack a non-overlapping VTS via
crack_key_outcome and report title_is_clear when the title's own
extents show no scrambling. A genuinely-clear stub on an otherwise-CSS
disc needs no key.
- ensure_title_decryptable: pass a clear stub without a key; a scrambled-
but-uncrackable title still hard-fails with CssKeyMissing.
- is_scrambled_pack: hardened scramble-evidence gate for the crack scan —
requires the MPEG-PS pack-start signature before trusting the 0x14
scramble bits, so a clear stub with stray 0x14 bits can't flip
saw_scrambled. The descramble loop keeps the looser is_scrambled.
- mux/resolve: ISO per-title gate routes through the clear-aware check.
Add PassProgress::bytes_retryable_total (NonTrimmed/NonScraped — failed
and awaiting retry), distinct from bytes_pending_total which also folds
in not-yet-attempted (NonTried) bytes. Set it at every construction site
(Sweep from the snapshot, Patch from stats, 0 for sequential/placeholder
paths). The disc-level 'lost' display in the CLI can now use
unreadable+retryable instead of unreadable+pending, so a healthy
in-progress rip no longer reports its unread remainder as lost.
Disc::copy gated its entire mapfile-resume dispatch behind
`if opts.multipass`, so a plain (non-multipass) `disc:// iso://`
copy always called sweep_internal(resume=false) — wiping the
mapfile + ISO and re-sweeping from sector 0. The CLI help and
rip_iso examples promise "auto-resumes if interrupted", and the
sweep already flushes a crash-safe mapfile per block, so the only
missing piece was consulting it on re-run.
Lift the mapfile-resume dispatch out of the multipass guard:
clean->no-op, size-mismatch->fresh sweep, NonTried tail->resume
sweep now apply to both modes. The patch (Pass N) dispatch on
retryable bytes stays multipass-only; a plain copy with a
fully-attempted mapfile returns a terminal result instead of
restarting. autorip (always multipass=true) is unchanged.
Add a regression test that an interrupted plain copy resumes the
NonTried tail and does NOT re-read the Finished prefix.
Targeted tests for the rc.6 surfaces, strengthening (not duplicating) the
regression tests the rc.6 commits already shipped. No production code changes.
aacs/keys.rs — resolve_keys_with_reason / classify_resolve_failure:
The E7021/E7022 split is already proven end-to-end through the
ensure_decryptable gate (disc/mod.rs). These pin the classifier directly at
the keys.rs seam for the branches the gate test does not reach:
- processing-keys-only + zero VID -> VidUnavailable (the gate test only
exercises the device-keys arm of has_derivation_material).
- VID PRESENT + material -> NoMaterial: a non-zero VID must never be
reported as VidUnavailable however much material is on hand (the has_vid
short-circuit; the gate test only uses the zero-VID sentinel).
- VID present + no material -> NoMaterial.
- version dispatch: version 1 routes the V10 resolver (stamps V10), any
other value routes the V20->V21 chain; a resolved disc returns Ok, never
Err(ResolveFailure).
unlock.rs — Unlocker seam introspection + ordering:
- matching_name reports the first matching unlocker without running it, and
is None for an unsupported drive; registered_count grows after a
registration (monotonic check — the registry is process-wide and shared
across the unlock tests, so no exact-delta assertion).
- route_unlock first-registered-match-wins: two unlockers matching the same
identity, the earlier-registered one runs and the later is never consulted.
The dir:// extractor set the AACS unit-alignment base ONCE to the first
extent's start, then read every extent against that single base. For a
multi-extent (fragmented / Long-AD / continuation-ICB) file the second
and later extents start at arbitrary LBAs whose offset from the first
extent is generally not a multiple of 3 sectors, so the first read of
each later extent failed the decrypt-on-read gate
(is_unit_aligned(lba, unit_base)), returned DecryptFailed, and recorded
the whole extent as a zero-filled hole even though the data was readable.
Re-anchor the unit base PER extent (matching mux/disc.rs and
sector/prefetched.rs), so each clip's encrypted region gates on its own
unit grid. Same bug class as the rc.5.2 clip-anchor fix.
Also harden the extract + keydb write paths:
- finalize_file: fsync the .partial after set_len (the truncation runs
on a second handle the content fsync never touched) and fsync the
parent dir after rename so the new dirent is crash-durable.
- keydb write_atomic: fsync the parent dir after rename (POSIX dirent
durability), matching the finalize_file pattern.
- AACS tail batch: document that decrypt_sectors' trailing-partial
contract already handles the short final unit; no math change.
- decrypt-loss delta loads use Acquire (defensive happens-before if
file extraction is ever parallelised).
- is_windows_reserved: add CONIN$/CONOUT$/CLOCK$; reserved names are
now substituted (prefix _) instead of aborting the whole tree walk, so
a legal Linux-authored NUL.cfg extracts.
- http_get header cap: >= MAX_HEADER_BYTES (was > , one byte over).
Regression tests: multi-extent AACS file (Δ4-sector extents) extracts
both extents with zero loss; focused per-extent alignment-arithmetic
test; reserved-name substitution assertions.
When key resolution had derivation material (device or processing keys)
but no Volume ID was available to derive the unit key, surface
Error::AacsVidUnavailable instead of the generic NoDiscKey. When there
was no usable key material at all, keep NoDiscKey.
resolve_keys_classical / resolve_keys_v21 still return a bare
Option<ResolvedKeys> (all existing callers unchanged); a new
resolve_keys_with_reason wrapper threads the typed ResolveFailure
(VidUnavailable | NoMaterial) out. decrypt_with uses it; the
ensure_decryptable_keys gate maps a captured AacsVidUnavailable reason to
E7021, otherwise E7022. No decryption math, key derivation, or descramble
logic changed -- only the reason reported on a resolution failure.
Adds ensure_decryptable_aacs_vid_unavailable_vs_no_key proving both
branches (device-keys + zero VID -> E7021; no keys -> E7022).
Rewrite the AACS scan/VID trace lines so a reader understands them without
opening the source: name the real thing (AACS host certificate, Volume ID,
decryption key), say "key source" not "keydb", and describe what happened.
The VID flow is unchanged (unlocker OEM VID → cert handshake → continue); a
missing VID is logged, never fatal. All strings are in tracing macros (the
sanctioned debug-log channel) — no English added to any Error.
default_path() now resolves <dir of current exe>/keydb.cfg with no OS-config
fallback — freemkv is a portable standalone binary, so the keydb lives next to
it. --keydb still overrides.
Sibling of Disc::copy specialized to write per-file instead of a whole ISO
image, decrypting on the way out: walk the UDF tree, read each file's extents
through the shared DecryptingSectorSource (AACS unit-aligned, CSS per-VTS),
strip AACS/, sanitize host paths per component, .partial+rename, 1-shot with
per-file loss accounting (no mapfile; recovery stays the iso:// multipass
path). Reuses UdfFs + the decrypt seam; only the per-file orchestration is new.
Adds Disc::ensure_decryptable / ensure_decryptable_keys, the single decrypt
gate consulted before any copy or mux. When the source is encrypted and no
key resolved (and not --raw), abort with a typed error and write nothing,
instead of silently emitting ciphertext at exit 0. Unifies the prior ad-hoc
CSS/AACS checks.
Three Silence-of-the-Lambs (R2 PAL SD-DVD) follow-ups for rc.5.2.
SUB-TASK 1 — Windows Explorer showed 12.5 fps (half) for the 576i25 track.
Root cause: the DefaultDecodedFieldDuration (20 ms field) element rc.5.1
added to "fix" Windows fps did the opposite. With FlagInterlaced=1 +
DefaultDuration=40 ms + DefaultDecodedFieldDuration=20 ms, Explorer halved
to 12.5 fps and MediaInfo flipped to VFR. MakeMKV's correct rip omits the
field-duration element, keeps FlagInterlaced=1 + FieldOrder=TFF +
full-frame DefaultDuration (40 ms), and Explorer shows 25 fps / MediaInfo
CFR. Fix: MkvTrack::video now passes field_duration_ns == 0 so the element
is no longer written; the 1/DefaultDuration = 25 fps signal (the only one
tools trust) is the full-frame value. Interlace signalling (FlagInterlaced,
FieldOrder=TFF) is retained — MediaInfo reads scan type from the MPEG-2 ES
picture coding extension, so it still reports Interlaced / Top Field First.
Tests pin the new TrackEntry elements (element present/absent + values).
SUB-TASK 2 — opening "menu"/still-frame video. Traced the MPEG-2
opening-GOP path; the wrong/last seq header and PTS-floor-to-0 hypotheses
are RULED OUT with file:line evidence: codecPrivate is the FIRST sequence
header (read once at headers-ready, mkvstream.rs:115 + pipelined_stream.rs:289),
DVD VOBU structure guarantees each title opens on seq header + I-frame (no
mid-GOP open), the parser back-anchors leading still-frames to the disc's
real timeline (mpeg2.rs:296-303), and the muxer anchors base on the opening
keyframe's real PTS so the t=0 floor (mkv.rs:963) never corrupts it.
Regression tests pin all three (parser + muxer level).
SUB-TASK 3 — make --log-level 3 self-sufficient (diag.rs + minimal hooks).
(a) dump the ACTUAL MKV TrackEntry elements written per track
(tag=mkv.track: FlagInterlaced, FieldOrder, DefaultDuration, field duration,
Display dims, codecPrivate hex) so Windows-fps-class metadata is verifiable
from a log alone. (b) capture the first ~100 coded frames per track (raw)
to <output>.opening.bin with a per-frame summary line (tag=mkv.opening.frame:
track, key/delta, size, PTS) so opening-GOP/menu issues are diagnosable from
a future log without the disc. Both gated to log-level 3; normal runs open
no side file and record nothing.
CI gate (Rust 1.86): fmt --check, clippy -D warnings, and test --tests all
green.
The DVD AC-3 sub-stream probe recorded the FIRST decodable frame of each
physical 0x8x sub-stream as its channel count. A DVD feature opens with
logos/warnings whose audio is often a thin 2.0 bed on 0x80 before the
real 5.1 main mix begins a fraction of a second later. The probe locked
onto that opening 2.0 frame and reported 0x80=2, missing the 5.1
entirely (confirmed on Greenland: 0x80's head frames are acmod=2, then
acmod=7+lfe). With no 6-channel sub-stream found, channel-match routing
fell back to the ordinal map — harmless on Greenland, but on a disc where
the 5.1 lives on a non-ordinal sub-stream the wrong-substream bug stays
unfixed.
Fix: scan EVERY 0x0B77 frame of each sub-stream in the probe window and
keep the MAXIMUM channel count (the sub-stream's real main-mix
capability), advancing frame-by-frame via ac3_frame_size so a frame body
can't be mistaken for a new sync. Also bump PROBE_SECTORS 512->1024: the
1 MiB head window saw ONLY 0x80; 2 MiB reliably contains a frame of every
physical sub-stream.
Greenland tag=dvd.substream: before 0x80=2 (only); after 0x80=6, 0x81=2,
0x82=2 — matching the IFO and the decoded output.
Adds probe_reads_max_channels_no_cross_contamination regression test.
1. HEVC CRA->BLA false-trigger on 33-bit PTS wraparound
(src/mux/codec/hevc.rs): the clip-boundary auto-detect compared the
RAW 33-bit PES PTS against the high-water mark, so a single-clip title
crossing 2^33->0 (~26.5h) false-armed pending_clip_boundary and rewrote
a legitimate in-clip CRA(21)->BLA_W_LP(16), dropping valid RASL pictures
(visible corruption) and breaking the single-clip byte-identical
guarantee. Now unwrap the PTS onto a monotonic 64-bit timeline first
(a near-full-period backstep is a wrap: add 2^33, update the watermark,
do not arm). Regression test cra_after_33bit_pts_wrap_not_rewritten;
the genuine-clip-join test still passes.
2. Single-pass recovery read bypassed the transport-failure abort
(src/mux/disc.rs): the line-442 short-circuit only inspected the 10s
read res. A transport failure (status 0xFF, wedged USB bridge) on the
60s recovery read fell into the skip_errors branch and zero-filled/
advanced, marching the disc at one bridge-recovery per probe
(run-forever, hard rule #2). Re-check the recovery error for
is_scsi_transport_failure() before the skip block and abort with
Error::DiscRead. Test transport_failure_on_recovery_read_aborts_even_with_skip_errors.
3. Recovery-read SUCCESS branch had no coverage (src/mux/disc.rs tests):
added RecoverableReader (errors when recovery=false, succeeds when
recovery=true) and test recovery_read_success_muxes_recovered_data_no_skip
driving fill_extents to the size-1 bottom-out and asserting the recovered
data is muxed (counters advance, no skip).
4. TrueHD channel-correction probe omitted set_unit_base
(src/disc/mod.rs correct_truehd_channels): the probe read via a
DecryptingSectorSource without anchoring the AACS unit-alignment gate,
so it degraded to absolute start_lba % 3 and returned DecryptFailed on a
non-3-aligned extent, silently understating Atmos/7.1 as 5.1. Now call
set_unit_base(ext.start_lba) before the probe read (no-op for CSS/None).
5. is_unit_aligned lba<unit_base latent trap (src/aacs/decrypt.rs):
wrapping_sub mis-gated when lba < unit_base (2^32 == 1 mod 3). Switched
to saturating_sub (clamps offset to 0, a unit boundary) and pinned the
contract with is_unit_aligned_lba_below_base_is_well_defined plus
is_unit_aligned_relative_to_base.
cargo +1.86 fmt --check / clippy -D warnings / test --tests all green.
Fixes the "Silence of the Lambs" R2 PAL wrong-substream rip: the feature's
IFO declares one 5.1 AC-3 stream, but the scan assigned it the on-wire
sub-stream id 0x80 purely by per-codec ordinal (ifo::assign_audio_sub_stream_ids).
On this disc the physical 0x80 carries the 2.0 down-mix and the 5.1 main mix
lives at a different 0x8x sub-stream, so the rip muxed 2.0 while labelling it
"Dolby Digital 5.1" (the acmod fixup in mkv.rs then corrected only the Channels
element, surfacing the mismatch as the "IFO claimed 6 but acmod says 2" warning
— too late to re-route).
New src/disc/dvd_audio_probe.rs probes each physical AC-3 sub-stream's real
channel count from the head of the feature (the acmod/lfeon of its first frame
after the 0x0B77 sync) and re-routes each IFO-declared AC-3 stream onto the
physical sub-stream whose actual channel count matches the declared count,
instead of trusting the ordinal. Wired into both mux demux paths
(DiscStream::new and resolve::build_iso_pipeline) over the decrypting reader,
so it works on CSS discs and the autorip ISO-remux path alike. Bounded
512-sector best-effort read; an empty/unreadable probe degrades to the original
ordinal mapping (no regression on normal discs).
The cell selection is left unchanged: the feature's cell 0 (cat=0x02, 302.4s)
is chapter 1 of the movie (matches MakeMKV's chapter map and 1h53 duration
exactly), so it must NOT be dropped — the perceived "wrong video at the start"
was the wrong 2.0 audio over the opening, the same root cause.
Diagnostics (--log-level 3): new tag=dvd.substream rows dump the ACTUAL acmod
channel count of each physical 0x8x sub-stream read from the VOB, and the
per-cell tag=dvd.cell verdict now spells out the keep/skip reason. With the
existing tag=dvd.aattr (IFO declared sub_id + channels) a bug log alone now
shows whether the ordinal 0x80 really carries the declared layout — no disc
needed to diagnose this class.
expose ac3::find_ac3_sync as pub(crate) for the probe.
The "TopGun bug" (Top Gun 1986 UHD, DV Profile 7 dual-layer): the auditor
flags the rip `corrupt` with a flood of HEVC "Could not find ref with POC N"
decode errors (POC 114/210/228/234/240/246/252/318/336/354/372/399 — one
cluster per clip join).
Root cause: the title is multiple .m2ts clips joined at non-seamless
boundaries (mpls connection_condition 0x05/0x06), read as one concatenated
stream. Each next clip opens with a CRA whose RASL leading pictures reference
pre-join frames gone after concatenation. hevc.rs already has the spec remedy
(rewrite splice CRA_NUT 21 → BLA_W_LP 16 so a decoder sets NoRaslOutput and
drops the dangling RASL), behind mark_clip_boundary() — but nothing ever
called it: connection_condition is not plumbed through the threaded mux
pipeline, so the whole CRA→BLA mechanism was dead code.
Fix: detect the boundary inside the parser from the bitstream. Each clip
carries its own PES PTS base, so a non-seamless join is a large backward
PTS reset. HevcParser::parse now tracks a PTS high-water mark and, on a
backward step beyond 3 s (270000 ticks @ 90 kHz — mirroring the mux-side
DISCONTINUITY_BACKSTEP_NS), arms the existing CRA→BLA rewrite for the new
clip's first IRAP. Self-contained: works for every mux path, no cross-thread
plumbing. The 3 s threshold sits above any HEVC B-frame reorder dip and far
below any clip duration, so it never false-triggers in-clip; single-clip /
seamless titles never arm it and stay byte-identical.
Adds regression test cra_at_auto_detected_pts_backstep_rewritten_to_bla
(in-clip dip must not trigger; splice CRA after a backward reset must become
BLA; one-shot). hevc suite 55/55, mux suite 780/780.
Matthew Jackson