improve scaling of new clvm compression implementation

[arvidn]

at the expense of not always finding the best compression.

If the number of parents aren't capped, that list can become very long and very expensive to maintain and traverse. Not having access to every parent when finding a path is not optimal, but I believe this will give a good result most of the time, in the common cases. It definitely makes the algorithm scale a lot better.

The incremental serializer compares against node_to_bytes_backrefs():

Benchmarking many_spends/node_to_bytes_backrefs: Warming up for 1.0000 ns
Warning: Unable to complete 10 samples in 5.0s. You may wish to increase target time to 775.1s.
many_spends/node_to_bytes_backrefs
                        time:   [64.284 s 67.147 s 70.449 s]
Benchmarking many_spends/incremental_serializer: Warming up for 1.0000 ns
Warning: Unable to complete 10 samples in 5.0s. You may wish to increase target time to 10.0s.
many_spends/incremental_serializer
                        time:   [942.38 ms 952.31 ms 962.57 ms]
                        No change in performance detected.

---

Note

Medium Risk
Changes core compression path selection by dropping some parent edges, which can subtly affect compression ratios and edge-case behavior even though the logic remains bounded and well-scoped.

Overview
Improves scalability of the new incremental CLVM compression by capping TreeCache parent links per node (MAX_PARENTS = 8) and evicting older entries, trading optimal back-reference selection for bounded memory/CPU during find_path().

Adds a Criterion benchmark (many_spends) to stress incremental serialization on large spend-lists and removes the now-unneeded serializer-cmp fuzz target. CI wheel builds are also stabilized by pinning maturin to 1.12.2 (avoids --sdist artifact changes in 1.12.3).

^{Written by Cursor Bugbot for commit dd3679c. This will update automatically on new commits. Configure here.}

[coveralls-official]

Pull Request Test Coverage Report for Build 22324685918

Details

• 13 of 15 (86.67%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.01%) to 88.222%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/serde/tree_cache.rs	13	15	86.67%

Totals
Change from base Build 22176657628:	-0.01%
Covered Lines:	6831
Relevant Lines:	7743

---

💛 - Coveralls

[arvidn]

@cursor review

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[richardkiss]

Okay, I think I see more or less how it works. You're evicting a random parent so there are fewer things to look at. There might be a way to smartly evict potential parents that could never be reached. But also, I think this serialization format is a dead-end: serde 2026 is way better, both faster and better compression, and dead-simple now that interning is implemented. I have a draft PR here #708