perf: use c allocator for shared io - #92

🤖 I have created a release *beep* *boop*
---


##
[4.4.0](zapi-v4.3.0...zapi-v4.4.0)
(2026-11-13)


### Features

* **js:** add exact BigInt conversions, toI128LowBits and fromWords
([#91](#91))
([b56bf2d](b56bf2d))


### Performance Improvements

* use c allocator for shared io
([#92](#92))
([a736d8a](a736d8a))

---
This PR was generated with [Release
Please](https://github.com/googleapis/release-please). See
[documentation](https://github.com/googleapis/release-please#release-please).

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

…10612)

**Motivation**

`std.heap.page_allocator` maps at least one page per allocation and
unmaps it on free, with no reuse. It fits large, long-lived buffers, but
a few places in the Zig code use it for small per-call allocations.

**Description**

- `bindings/napi/pubkeys.zig`: keep `page_allocator` for the long-lived
pubkey cache storage, including `pkix.load`, since `pkix.install` swaps
the loaded storage and its allocator into the live cache. Per-call
allocations (path strings, aggregate indices, `syncPubkeys` pointer
arrays) now use `c_allocator`.
- `PubkeyCache.syncPubkeys` takes a scratch allocator for its `prepared`
and `batch_errors` arrays instead of borrowing the cache allocator.
- `era.getShortEraRoot` takes an allocator from the caller instead of
hardcoding `page_allocator`.

The same issue in zapi's shared `std.Io.Threaded` (which allocates every
`async`/`concurrent`/group task closure through its gpa) is fixed in
ChainSafe/zapi#92. The zapi bump to 4.4.0 will land in a separate PR.

**Testing**

- `zig build test:state_transition test:era test:tidy`
- `pnpm --filter @lodestar/native prepare-mainnet`, then binding tests
`pubkeys`, `rewards`, `proposerRewards`, `attestationRewards`

> This PR was written primarily by Claude Code.

---------

Co-authored-by: Cayman <caymannava@gmail.com>

**Motivation**

zapi 4.4.0 includes ChainSafe/zapi#92: the shared `std.Io.Threaded`
behind `js.io()` now allocates its `async`/`concurrent`/group task
closures with `c_allocator` instead of `page_allocator`, so each task no
longer costs an mmap/munmap pair. Follow-up to #10612.

**Description**

- Pin zapi to the `zapi-v4.4.0` tag commit `0cc6b2c`.
- zapi 4.2.0 renamed `Number.toU64Exact` to `toSafeInteger`
(ChainSafe/zapi#81) with unchanged semantics. Update the binding call
sites.

Other changes since the previous pin (4.1.0) are additive
(`js.scratch()`, exact BigInt conversions). 4.2.0 also requires
`setEnv`/`restoreEnv` to be paired; lodestar does not call them
directly.

**Testing**

- `zig build test:tidy`, bindings build
- `pnpm --filter @lodestar/native prepare-mainnet` and the full binding
test suite: 26 files passed, 1 skipped (506 tests passed, 11 skipped)

> This PR was written primarily by Claude Code.