[andrewchen5678]

Two questions, and only one has an exact answer.

The exact one

OPUS_GET_FINAL_RANGE. If the decoder's entropy coder ends in the same state as the encoder's, the decoder walked an identical path through that packet's range coder — libopus's own test suite uses it the same way. Unlike a digest of the bitstream it is valid across platforms, since the AVX2 and NEON archives take different float paths and would produce different bytes for reasons that are not bugs. Asserted for all 800 packets.

The approximate one

Four signal classes — tone, chord, band-limited noise, speech-like — across 48 kHz stereo and mono, 16 kHz and 8 kHz.

measure	asserted	why
per-band energy error	≤ 6 dB	phase-blind, survives what the other two trip over. The gate.
correlation	per signal class	catches waveform damage band energy would miss
SNR	reported only	collapses on noise by design — a working encoder scores 2.9 dB

The correlation floor cannot be one number: noise round-trips at 0.70 at 8 kHz on a healthy encoder, so anything strict enough to mean something for a chord fails every clean build. 0.99 tone, 0.95 chord and speech, 0.60 noise.

Two bugs found by measuring, both in the test

• Alignment cannot be a fixed offset. get_lookahead() is exact — the aperiodic noise signal locks onto it off-by-zero at every rate, now asserted — but periodic content aligns equally well a whole period late, and the speech-like signal genuinely lands 16 samples early at 16 kHz. Forcing the reported value dropped its correlation to 0.7180 and failed six checks against an encoder that was working perfectly.
• A search window must leave room for the largest lag it tries. At 8 kHz a one-second signal is 8000 samples, so a fixed 8192-sample window left room for exactly one candidate — delay 0. Every metric downstream then compared signals 52 samples apart and reported a broken codec.

Margins, and cross-platform stability

Worst band error is 4.1 dB against the 6 dB gate. Tightest correlation margin is 0.977 vs 0.95.

	macOS arm64	Linux aarch64
48k stereo noise corr	0.9429	0.9430
16k noise corr	0.8150	0.8166
16k noise band	4.1 dB	3.8 dB
8k noise corr	0.7029	0.7027

Agreement to within 0.002 correlation and 0.3 dB across two architectures is the property that makes an approximate check work where a digest could not.

Cost

74 checks, up from 37. About 0.4 s added to the binary, and quick enough under valgrind that test-docker.sh needed no change — verified: --- linux-aarch64 OK with valgrind clean and its control still detected.

🤖 Generated with Claude Code