hashsig-glue: replace Box allocs with Zig-owned placement-init pattern

[zclawz]

Addresses the review comment #918 (comment) (ref PR #918) requesting that key-generation and signing be refactored to avoid additional Rust heap allocations, consistent with the caller-supplies-buffer pattern already used in multisig-glue.

Problem

hashsig-glue was Box-allocating every KeyPair, Signature, and PublicKey on the Rust heap and handing raw pointers to Zig. Zig then had to call back into Rust (hashsig_keypair_free etc.) to free them. This created a split-ownership model where allocation and deallocation crossed the FFI boundary in opposite directions, complicating lifetime tracking and adding unnecessary alloc/free pairs.

Solution: placement-init pattern

Rust exports layout queries so Zig can pre-allocate correctly-sized storage, then passes that storage to Rust for in-place initialisation. Cleanup runs Rust Drop in-place, then Zig frees its own buffer. The Rust heap is never involved for the outer struct.

rust/hashsig-glue/src/lib.rs

New layout queries:

• hashsig_sizeof_keypair / _signature / _public_key — Zig queries at runtime to know how much to allocate
• hashsig_alignof_keypair / _signature / _public_key — required alignment (C malloc always satisfies this for #[repr(C)] types)

New placement-init (replaces Box-returning functions):

• hashsig_keypair_generate_into(out: *mut KeyPair, ...) — uses std::ptr::write, returns i32 (0 / -1)
• hashsig_keypair_from_ssz_into, hashsig_sign_into, hashsig_signature_from_ssz_into, hashsig_public_key_from_ssz_into — same pattern

New drop-in-place (replaces Box::from_raw frees):

• hashsig_keypair_deinit / hashsig_signature_deinit / hashsig_public_key_deinit — call std::ptr::drop_in_place, do not free the buffer (caller owns it)

Removed (all were Box-allocating):
hashsig_keypair_generate, hashsig_keypair_from_ssz, hashsig_keypair_free, hashsig_sign, hashsig_signature_free, hashsig_signature_from_ssz, hashsig_public_key_from_ssz, hashsig_public_key_free

pkgs/xmss/src/hashsig.zig

• KeyPair, Signature, PublicKey wrappers now own their storage via cAlloc/cFree (thin wrappers over std.c.malloc/std.c.free)
• C malloc always returns memory aligned to max_align_t (>=8 bytes on LP64), satisfying the alignment of any Rust #[repr(C)] type used here
• Public API is unchanged — same generate, fromSsz, sign, verify, toBytes, deinit methods; no changes needed in aggregation.zig or any other caller
• PublicKeyCache slot semantics unchanged; deinit correctly reconstructs _buf + handle from the stored integer pointer

What this does NOT eliminate

• Inner leansig Vec<u8> data inside the XMSS key/signature types — these are owned by the leansig library and freed correctly when drop_in_place runs Drop
• The intermediate Vec<u8> inside the to_bytes() serialisation helpers — a leansig API constraint; short-lived temporaries that never escape the FFI call

Verification

• cargo check -p hashsig-glue clean
• zig ast-check on hashsig.zig, aggregation.zig, lib.zig all pass
• No remaining callers of any removed Box-based FFI symbol in the Zig tree

[zclawz]

CI fix pushed — commit 4aced8e

Root cause: pkgs/xmss/src/lib.zig has its own copy of PublicKeyCache (separate from the one in hashsig.zig). When PublicKey gained the _buf field in this PR, hashsig.zig's deinit was correctly updated to supply both fields:

const raw: [*]u8 = @ptrFromInt(ptr_int);
var pk = PublicKey{ ._buf = raw, .handle = @ptrCast(raw) };

But lib.zig's deinit was left with the old single-field init:

var pk = PublicKey{ .handle = @ptrFromInt(ptr_int) }; // missing _buf

This caused error: missing struct field: _buf on all three macOS checks (build / test / Dummy prove).

Fix: port the same corrected pattern to lib.zig's deinit. Both _buf and handle point to the same C-heap allocation start, so the stored atomic usize (which is @intFromPtr(handle)) is sufficient to reconstruct both fields.

🤖 zclawz

[zclawz]

CI fix: Rebased onto main to resolve build failure.

The CI was failing because main renamed initTestThreadPool → setupTestPrimitives in forkchoice.zig and testing.zig. The PR branch was forked before that rename, so the merge commit had an undeclared identifier at line 3967.

Fix: clean rebase onto main (no conflicts, PR changes to xmss/hashsig-glue preserved). Force-pushed to re-trigger CI.