[mshunshin]

Why

#167 drops the classic MQ encoder's terminal 0xFF, as T.800 B.10.7 requires ("The sequence of bytes actually included for any given code-block must not end in a 0xFF"). Its tests check each code-block's bytes: finish_omits_pending_terminal_ff, short_sequences_never_end_in_ff_and_round_trip and classic_code_block_contributions_never_end_in_ff.

Nothing checks what the rule protects on a whole codestream. Code-block data and packet headers follow one another directly, so a trailing 0xFF pairs with the next byte, and a byte-aligned pair in 0xFF90–0xFFFF is a marker value (D.4, note). A decoder resynchronising on SOP/EPH, or looking for the EOC after a tile-part with Psot 0, can stop at one.

Change

One integration test file, crates/j2k-native/tests/codeblock_termination.rs, and no library changes. It encodes through the public encode and finds each tile-part's data from its SOT's Psot and its SOD.

• tile_data_holds_no_marker_values: no byte-aligned 0xFF90–0xFFFF pair and no final 0xFF in any tile-part's data. It covers 16×9, 64×64 and 256×256 grey images, reversible and irreversible, classic and HTJ2K. The reversible codestreams must also decode back to their samples.
• packet_bodies_never_end_in_0xff: with SOP and EPH on, no packet body ends in 0xFF, which would otherwise pair with the next packet's SOP.

The samples are noise over a gradient, so code-blocks have enough coding passes that most MQ terminations would end in 0xFF without the fix.

Results

Both tests pass on main at 968dbb4. With the self.data.pop() in ArithmeticEncoder::finish_checked removed, both fail:

16x9 reversible=true ht=false: byte-aligned 0xFFCF in tile data
reversible=true: packet 2's body ends in 0xFF

No changelog entry, since the change is test-only (like #177 and #318).

Run locally (aarch64 macOS):

• cargo fmt --all --check
• the strict Clippy gate on this test target, with xtask's pedantic and nursery lints and baseline allows
• cargo xtask repo-lint
• cargo test --release -p j2k-native -p j2k-codec-math -p j2k: 1,661 passed, 0 failed

I wrote these tests with the same encoder fix on my fork; #167 landed first, so only the tests are left to offer.

AI-Tooling: Claude Opus 5.5 noreply@anthropic.com

[jcwal1516]

Thanks for catching this! I ran both new tests on my current checkout and they passed.

LGTM, I'll merge once CI is green.