Rounding out specification of prune behavior inside lean model

[luketpeterson]

No description provided.

[github-actions]

Bench A/B vs base: done

job log · 2026-10-04 01:52:43 UTC

base b4a6abd → head 85220b4, 3 round(s), median of each run averaged; negative is faster

bench	cases	geomean	largest gain	largest loss	>5% faster	>5% slower
shakespeare	7	-2.1%	-8.0% shakespeare/shakespeare_words_val_count	+7.3% shakespeare/shakespeare_sentences_get	1	1
cities	5	-0.3%	-3.2% cities/cities_val_count_act	+1.0% cities/cities_val_count	0	0
sparse_keys	96	-0.7%	-4.8% sparse_drop_bench/400	+3.4% join_sparse/50	0	0
binary_keys	77	-1.0%	-7.2% binary_zipper_iter/50	+13.2% binary_set_val_at_long_replace/160	7	6
superdense_keys	104	+0.2%	-5.7% superdense_drop_head/2000	+5.2% superdense_insert/200	1	1
act_paths	46	-1.0%	-5.6% shape_map_to_act_cata/random_long	+4.3% size_paths_to_act/25000	1	0
zipper_head_owned	6	+1.4%	-2.9% zipper_head_owned/owned_head_write_creation_cleanup_unchecked	+7.8% zipper_head_owned/borrowed_head_read_creation	0	2
product_zipper	4	-1.3%	-2.2% product_zipper/introspecting_pathmap_pathmap	+0.9% product_zipper/generic_act_act	0	0

20 case(s) moved more than 5%

bench	case	base	head	change
binary_keys	binary_set_val_at_long_replace/160	50 ns	56 ns	+13.2%
binary_keys	binary_set_val_at_long_replace/256	70 ns	78 ns	+10.7%
binary_keys	binary_insert/1600	157.17 µs	172.47 µs	+9.7%
binary_keys	binary_set_val_at_short_replace/3	24 ns	26 ns	+8.3%
shakespeare	shakespeare/shakespeare_words_val_count	337.93 µs	310.93 µs	-8.0%
zipper_head_owned	zipper_head_owned/borrowed_head_read_creation	11.01 µs	11.86 µs	+7.8%
shakespeare	shakespeare/shakespeare_sentences_get	47.28 ms	50.74 ms	+7.3%
binary_keys	binary_zipper_iter/50	1.26 µs	1.17 µs	-7.2%
binary_keys	binary_zipper_iter/100	2.45 µs	2.29 µs	-6.3%
binary_keys	binary_drop_head/50	4.73 µs	4.44 µs	-6.2%
binary_keys	binary_zipper_iter/400	9.30 µs	8.76 µs	-5.9%
binary_keys	binary_zipper_iter/800	18.15 µs	17.10 µs	-5.8%
superdense_keys	superdense_drop_head/2000	22.57 µs	21.27 µs	-5.7%
binary_keys	binary_zipper_iter/1600	35.81 µs	33.78 µs	-5.7%
act_paths	shape_map_to_act_cata/random_long	50.41 ms	47.58 ms	-5.6%
binary_keys	binary_descend_until/2000	190.67 µs	180.20 µs	-5.5%
superdense_keys	superdense_insert/200	2.17 µs	2.28 µs	+5.2%
zipper_head_owned	zipper_head_owned/owned_head_read_creation	11.81 µs	12.43 µs	+5.2%
binary_keys	binary_insert/100	7.30 µs	7.67 µs	+5.2%
binary_keys	binary_insert/200	15.49 µs	16.27 µs	+5.1%

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

[imlvts]

My criteria are the following:

• The model covers more
• The model is simpler (fewer special cases)

As far as I can tell this PR covers that. Simpler model, covers more.
Please upload reproducer corpus files
corpus/prune{-flag-remove-val.bin,-remove-val-valued.bin,-unmasked.bin,-subtract.bin}
and we can merge