[ThomasWaldmann]

Fixes for what a review of the chunker code (src/borg/chunkers/, C and Cython) turned up. None of them changes a cut point. One commit per issue:

1. chunking_time was a C float. It adds up one small increment per chunk, and a 32-bit float stops growing once the increments are below half an ulp of the sum (with 50 us per chunk at 1024 s at the latest). "Time spent in chunking" came out too low for long backups of small files. Now a double.
2. Parameter checks were asserts, but the scan code relies on them to stay inside its buffer and python -O drops asserts. They raise ValueError now. A negative window_size was not checked at all, neither by the buzhash chunkers nor by ChunkerParams (--chunker-params buzhash64,10,16,14,-5,0 crashed); both reject it now.
3. The normalized-chunking transition was computed as a position (last + normal_size) and compared as a pointer, which wraps for a huge normal_size and sends the scan out of the buffer. It is a length comparison now. API only: no borg command passes normal_size.
4. gcc -Warray-parameter warnings for ra_new / gl_new / tp_new: header and definition declared the tables parameter differently.
5. Stale comments and dead code: phte_core.h claimed everything in it is static; the avx512 "compiler too old" branches in fc_kernel_select() / bz64_kernel_select() were unreachable; "both" UHF-then-PRF chunkers are three.

Testing

• New tests: constructors reject invalid params with ValueError (all CDC chunkers), ChunkerParams rejects a negative window size, a huge normal_size gives the cuts of the strict mask alone.
• src/borg/testsuite/chunkers and helpers/parseformat_test.py pass, also the archive / create / recreate / transfer / extract tests (macOS arm64).
• Cut points before and after are the same: a pure-Python model of the cut rules was compared with buzhash, buzhash64, fastcdc and toeplitz-aes on several hundred random parameter/data cases, each via 6 input paths (BytesIO, regular file with and without fh, short-reading file object, pipe, sparse).
• The C sources build without warnings with gcc 14 -Wall -Wextra, and a standalone fuzz of all scan kernels against reference loops still passes on aarch64 (neon, blockwise, scalar, aes-arm64, evp) and on a Zen 4 (avx2, avx512, aes-ni, vaes, evp).

There is no test for 1: chunking_time is read-only and a meaningful check would need minutes of accumulated time.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.39%. Comparing base (0a319a2) to head (94ac75b).
⚠️ Report is 3 commits behind head on master.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #10603   +/-   ##
=======================================
  Coverage   89.38%   89.39%           
=======================================
  Files         105      105           
  Lines       20336    20340    +4     
  Branches     3192     3194    +2     
=======================================
+ Hits        18177    18182    +5     
+ Misses       1503     1500    -3     
- Partials      656      658    +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.