…ling

Follow-up to #123. That fix covered deflate/deflate64/zlib; reviewing all
53 `raw_decode` implementations found the same defect in seven more.

A decoder parked in its terminal state that returns `done: false` with
nothing consumed and nothing written, while the caller still holds input,
is mapped by the RawDecoder->Decoder bridge to `Status::OutputFull` — "call
me again" — with nothing to progress on. Any loop waiting for StreamEnd
spins: CPU-bound, no allocation, so no output or memory cap catches it.

Fixed, each verified by reading its state machine and confirmed with a
probe that drives the real encoder output plus trailing bytes:

- zstd, xz          Done arms returned false; both are terminal (neither
                    supports concatenated frames/streams).
- lz4 block, lzo    Done after the zero-length terminator block.
- lz4 frame         only reachable by calling past StreamEnd, but the state
                    is terminal either way.
- lzx, amiga_lzx    Done arms returned false.
- rar5              returns on `State::Done` *before* the block that accepts
                    caller bytes, so a container passing the next header
                    after the payload could never make progress.
- gzip              subtler: the `BetweenMembers` arm sets `phase = Done`
                    intending to reach the Done arm that swallows trailing
                    bytes, but match arms do not fall through, so the loop's
                    no-progress check returned first and the swallow never
                    ran. Needs an explicit `continue`, like the 0x1F path.

Two call shapes matter and only the first was covered before: trailing bytes
in the same call, and trailing bytes arriving *after* the payload completes.
The second is what a container does when it hands over the next header, and
it is what xz and gzip failed — both looked safe under the first shape.

Codecs that only ever report `InputEmpty` and produce output on `finish`
(brotli, lzma, and the other buffering decoders) are correct as-is: the
trait permits that, and it terminates. The invariant asserted is not "must
report StreamEnd" but "must never report OutputFull with zero progress".

A sweep over all 50 round-trippable algorithms now reports no stalls.
tests/terminal_state.rs covers each fixed codec plus controls; rar5 is
decoder-only so it gets equivalent tests against its own fixture.

…ling (#124)

Follow-up to #123. That fix covered deflate/deflate64/zlib; reviewing all
53 `raw_decode` implementations found the same defect in seven more.

A decoder parked in its terminal state that returns `done: false` with
nothing consumed and nothing written, while the caller still holds input,
is mapped by the RawDecoder->Decoder bridge to `Status::OutputFull` — "call
me again" — with nothing to progress on. Any loop waiting for StreamEnd
spins: CPU-bound, no allocation, so no output or memory cap catches it.

Fixed, each verified by reading its state machine and confirmed with a
probe that drives the real encoder output plus trailing bytes:

- zstd, xz          Done arms returned false; both are terminal (neither
                    supports concatenated frames/streams).
- lz4 block, lzo    Done after the zero-length terminator block.
- lz4 frame         only reachable by calling past StreamEnd, but the state
                    is terminal either way.
- lzx, amiga_lzx    Done arms returned false.
- rar5              returns on `State::Done` *before* the block that accepts
                    caller bytes, so a container passing the next header
                    after the payload could never make progress.
- gzip              subtler: the `BetweenMembers` arm sets `phase = Done`
                    intending to reach the Done arm that swallows trailing
                    bytes, but match arms do not fall through, so the loop's
                    no-progress check returned first and the swallow never
                    ran. Needs an explicit `continue`, like the 0x1F path.

Two call shapes matter and only the first was covered before: trailing bytes
in the same call, and trailing bytes arriving *after* the payload completes.
The second is what a container does when it hands over the next header, and
it is what xz and gzip failed — both looked safe under the first shape.

Codecs that only ever report `InputEmpty` and produce output on `finish`
(brotli, lzma, and the other buffering decoders) are correct as-is: the
trait permits that, and it terminates. The invariant asserted is not "must
report StreamEnd" but "must never report OutputFull with zero progress".

A sweep over all 50 round-trippable algorithms now reports no stalls.
tests/terminal_state.rs covers each fixed codec plus controls; rar5 is
decoder-only so it gets equivalent tests against its own fixture.