feat(cli): --show-default-cparams, --max, -B# and legacy training - #529

Merged

polaz merged 36 commits into

mainfrom

feat/#128-cli-remainder

Sep 27, 2026

Commits on Sep 25, 2026

feat(cli): --show-default-cparams, with a public LevelParameters
polaz
committed
770829b
fix(encoding): gate the post-split pass on the strategy
polaz
committed
7804b25
feat(cli): --max, every knob at its hardest end
polaz
committed
cdd2152
feat(cli): -B# cuts benchmark inputs into independent frames
polaz
committed
4ccdb7a
feat(dictionary): the legacy trainer, and --train-legacy
polaz
committed
be020b2
perf(dictionary): word compares and unchecked induction in legacy
polaz
committed
5a4ccef
perf(dictionary): SA-IS induction without the type array
polaz
committed
126128d
perf(dictionary): name LMS substrings inside the suffix array
polaz
committed
fde314f
perf(dictionary): drop the padded suffix-array copy in legacy
polaz
committed
fa57789

Commits on Sep 26, 2026

perf(encoding): derive split-probe length codes once per block
polaz
committed
f09e0d0
perf(encoding): borrow the split probes' Huffman table
polaz
committed
d2a3570
perf(encoding): build FSE tables in place without full-width clears
polaz
committed
0a15716
build(fse): gate the owning table builders on their consumers
polaz
committed
3644bf0
perf(encoding): make the optimal parser's price mode a const
polaz
committed
43ac634
perf(encoding): compute integer-weight prices in place
polaz
committed
32120ae
perf(encoding): walk the insertion tree on the stored index
polaz
committed
f3ae187
perf(encoding): run the optimal parser as one pass per block
polaz
committed
342c49d
perf(encoding): search on the repeat history in place
polaz
committed
21282a8
perf(encoding): insert a tree catch-up run in one call
polaz
committed
e20d1bb
test(miri): let the aarch64 prefetch hints fall back to no-ops
polaz
committed
ea02b1e
perf(encoding): let a match cell carry offset and length only
polaz
committed
d9107b8
perf(encoding): take the tree's coordinates once per block
polaz
committed
4513256
fix(dictionary): stop the legacy neighbour walks at the noise slots
polaz
committed
4305e19
docs: -M on legacy training, dictionary parity tests in CI, private doc links
polaz
committed
ce2eeaf
perf(dictionary): borrow the legacy training corpus
polaz
committed
15a106d
perf(dictionary): build the suffix array in place
polaz
committed
923f0c1
perf(dictionary): prefetch ahead of the induction sweeps
polaz
committed
bdb10fd
perf(dictionary): prefetch ahead of the legacy analysis
polaz
committed
31438d5
perf(dictionary): keep the recursion's bucket tables in the array
polaz
committed
1e5ee30
perf(dictionary): compare legacy prefixes with the vector kernel
polaz
committed
0992599
perf(encoding): compare 32 bytes a step in the NEON prefix kernel
polaz
committed
6d09f03
fix: benchmark input checks, training sizes, probe table handles
polaz
committed
5e5c24c
fix: bound benchmark frames and training sizes by their budget
polaz
committed
b5d9c9d
fix(cli): build the 32-bit --max test without a Debug bound on Options
polaz
committed
1fda06a
fix(cli): split the training budget, sum sample sizes exactly
polaz
committed
8255e81
fix(cli): read --block-size into the setting -B sets
polaz
committed
50ccfc4