Adding WriteZipper::remove_subtrie

[luketpeterson]

Adding WriteZipper::remove_subtrie as this is a more fundamental operation than remove_branches, but remove_branches is useful and correct in its own right.

This new method balances the API, helping define and explain the consistency fixes proposed in #104 Also, in the future this will be the lowest-cost way to just chop a subtrie, although the current implementation doesn't offer much perf benefit vs. remove_branches.

I also updated the lean model to include the new method. However, this shifted the number of opcodes and threatened to invalidate the existing .bin files. So I made a change to the .bin header to include the opcode count, rather than hard-coding the %56 that was in there.

I pondered going further and actually redoing the .bin format entirely, storing opcodes + args instead of the raw fuzz signal, so decoding only happens once before saving a reproduction - but I didn't go that far because it seemed unnecessary and like I'd be over-stepping.

Thoughts, @imlvts?

[github-actions]

Bench A/B vs base: done

job log · 2026-10-13 01:59:26 UTC

base cb4e1e4 → head d8597a2, 3 round(s), median of each run averaged; negative is faster

bench	cases	geomean	largest gain	largest loss	>5% faster	>5% slower
shakespeare	7	+4.5%	-1.3% shakespeare/shakespeare_words_insert	+23.1% shakespeare/shakespeare_sentences_val_count	0	2
cities	5	-0.1%	-2.0% cities/cities_val_count	+2.6% cities/cities_get	0	0
sparse_keys	96	-0.5%	-4.0% binary_drop_head/50	+4.8% sparse_iter/10	0	0
binary_keys	77	+0.7%	-4.3% binary_meet/400	+31.1% binary_zipper_iter/50	0	6
superdense_keys	104	-0.9%	-13.0% superdense_drop_head/1000	+7.6% superdense_meet_after_join/1000	4	1
act_paths	46	-1.5%	-11.5% shape_map_to_act_cata/random_long	+2.8% size_map_to_paths/100000	5	0
zipper_head_owned	16	-0.3%	-2.7% zipper_head_owned/owned_head_read_creation	+1.3% zipper_head_owned/map_reader_is_val	0	0
product_zipper	4	-0.8%	-1.5% product_zipper/generic_act_act	+1.1% product_zipper/generic_pathmap_act	0	0

18 case(s) moved more than 5%

bench	case	base	head	change
binary_keys	binary_zipper_iter/50	1.19 µs	1.56 µs	+31.1%
binary_keys	binary_zipper_iter/100	2.32 µs	2.96 µs	+27.8%
binary_keys	binary_zipper_iter/200	4.58 µs	5.73 µs	+25.2%
binary_keys	binary_zipper_iter/400	8.89 µs	11.05 µs	+24.3%
binary_keys	binary_zipper_iter/1600	34.02 µs	42.21 µs	+24.1%
binary_keys	binary_zipper_iter/800	17.29 µs	21.37 µs	+23.6%
shakespeare	shakespeare/shakespeare_sentences_val_count	2.63 ms	3.24 ms	+23.1%
superdense_keys	superdense_drop_head/1000	17.23 µs	15.00 µs	-13.0%
act_paths	shape_map_to_act_cata/random_long	44.29 ms	39.20 ms	-11.5%
superdense_keys	superdense_drop_head/2000	23.17 µs	20.85 µs	-10.0%
act_paths	shape_map_to_act_cata/clustered	9.14 ms	8.30 ms	-9.2%
shakespeare	shakespeare/shakespeare_words_val_count	304.30 µs	331.37 µs	+8.9%
act_paths	shape_map_to_act_cata/dense_wide	9.44 ms	8.61 ms	-8.9%
act_paths	shape_map_to_act_cata/dense_narrow	10.92 ms	9.96 ms	-8.8%
superdense_keys	superdense_meet_after_join/1000	9.14 µs	9.84 µs	+7.6%
act_paths	shape_map_to_act_cata/small_alphabet	10.94 ms	10.19 ms	-6.8%
superdense_keys	superdense_drop_head/4000	29.36 µs	27.47 µs	-6.4%
superdense_keys	superdense_drop_head/8000	42.05 µs	39.63 µs	-5.8%

Full tables per bench are in the job log and the bench-out artifact.

[Adam-Vandervorst]

I believe this is good iff it's upheld by a lot of laws.
Such as being equivalent to remove_branches ; remove val or restrict(PathMap::new())

[luketpeterson]

I believe this is good iff it's upheld by a lot of laws.
Such as being equivalent to remove_branches ; remove val or restrict(PathMap::new())

Yup. The former is even codified in the default impl as well as the lean model.

[imlvts]

mostly good, but I think: ditch the --input-header CLI argument.
when writing, always write the most recent header applicable.
when reading, auto-detect the format by the presence of the magic string.