Define a single source of truth for the codec format contract

[audexdev]

Summary

The lossless invariant requires the encoder, decoder, and SIMD paths to agree bit-for-bit on every format-defining formula. Today those formulas are reimplemented independently in several files instead of living in one shared module, so a one-sided edit can silently break round-trip losslessness with no compiler help.

Affected areas

• src/codec/block/encoder.cpp
• src/codec/block/decoder.cpp
• src/codec/simd/neon.cpp
• src/codec/lpc/lpc.cpp
• src/codec/block/constants.hpp

Problem details

The same wire-format / arithmetic contract is written two or three times:

• Fixed predictor taps {1} / {2,-1} / {3,-3,1} / {4,-6,4,-1}: encoder compute_fixed_residual (encoder.cpp:265-292) vs decoder restore_fixed_in_place (decoder.cpp:320-341).
• FIR predictor taps/shift {3,-1}, >> 2: encoder uses named constants; decoder hardcodes 3LL*pcm[i-1] - pcm[i-2] >> 2 literally and does not reference the encoder constants (decoder.cpp:347-349).
• Zigzag mapping: unsigned_from_residual / zigzag_encode in the encoder vs to_unsigned / zigzag_decode lambdas in the decoder (decoder.cpp:85-88,120-124). A correct shared copy also exists in Rice::signed_to_unsigned (rice.cpp:7-15) — three implementations of one mapping.
• Stateless adapt_k: encoder adapt_k_stateless vs decoder inline lambda (decoder.cpp:90-96) — same (sum + count/2)/count / bit_width(mean-1) math twice.
• Partition-size math: encoder partition_sizes_for_block / max_partition_order_for_block (encoder.cpp:93-119) vs decoder partition_size_at (decoder.cpp:98-102).
• Q15 scale/shift (15 / 32768): the binding quantization contract is a bare literal on the decode side (decoder.cpp:43,50,388,398) and implicit on the encode side (via lpc.compute_residual_q15); it is not a named shared constant.
• Wire-format tag/mode vocabulary: kBinTag* / kResidualMode* / kPredictor* (encoder.cpp:46-56) and zero-run sub-tags kTagNormal/Run/Escape (encoder.cpp:670-672) are redefined in the decoder (decoder.cpp:68-72,139-141). None live in constants.hpp.

constants.hpp currently centralizes only block/partition limits and the control-byte bit layout — exactly the values that are not the lock-step hazard. The values that must stay identical are the ones duplicated.

Note: this is distinct from #8 (which targets UB/narrowing in those paths) and from the closed #19 (which documented the spec). This issue is about a single in-code source of truth.

Acceptance criteria

• Predictor formulas (fixed/FIR), zigzag mapping, stateless adapt_k, partition-size math, the Q15 shift/scale, and the tag/mode/predictor constants are each defined once and consumed by encoder, decoder, and SIMD paths.
• Removing or changing a formula in one place forces a build break or test failure if a consumer drifts.
• A round-trip test asserts encoder output and decoder input use the shared definitions (e.g. via a small shared header consumed by both).
• No behavior change to the emitted bitstream; existing fixtures still round-trip bit-perfectly.