[Kshot3000]

Fixes #946. Sibling of #945 (which fixed the same bug class in ergotree-ir's Sigma paths for #944); this covers the Scorex serialization layer and the crates built on it, which that audit didn't reach.

The gap

Four parse paths sized a buffer by an untrusted u32 prefix before reading a single payload byte, so a few input bytes aborted the process with an allocation failure instead of returning a parse error (all reproduced on develop under a 2 GB address-space limit):

• Vec<T>::scorex_parse — VLQ u32::MAX count → 17 GB allocation (for u32 items), SIGABRT
• NipopowProof::scorex_parse prefix count u32::MAX → ~1.58 TB allocation, SIGABRT
• PoPowHeader::scorex_parse header size u32::MAX → 4 GB allocation, SIGABRT
• ReducedTransaction::sigma_parse tx length u32::MAX → 4 GB allocation, SIGABRT

NiPoPoW proofs arrive from peers, and a reduced transaction (EIP-19) exists to be transferred and parsed by another party (e.g. a cold wallet) — both should fail with a parse error, not abort.

The fix (same pattern as #945)

• sigma-ser: Vec<T>::scorex_parse caps the pre-allocation at 1024 elements (the buffer still grows to the real count as items parse). New pub helper scorex_parse_bytes_exact reads exactly the claimed length, growing the buffer only as bytes arrive and erroring exactly like read_exact (Io("failed to fill whole buffer")) on a short stream.
• ergo-nipopow: prefix and suffix-tail pre-allocations capped the same way; the PoPowHeader header and interlinks-proof buffers now go through scorex_parse_bytes_exact.
• ergo-lib: ReducedTransaction's embedded-transaction buffer goes through the same helper; the helper's Io error is mapped back to SigmaParsingError::Io, so a short stream fails with exactly the error read_exact produced before (pinned by the new test).

Deliberate non-changes

The rest of the workspace's allocation sites were audited and are not affected: u8/u16 count prefixes (context extension, box tokens, tuples, collection/sigma-boolean items, ProofBytes, peer features, header fields), the transaction token-id count (bound-checked against MAX_OUTPUTS_COUNT * MAX_TOKENS_COUNT before allocating — the codebase's own precedent), and BigInt256/UnsignedBigInt (length checked ≤ 32 first). BatchMerkleProof collects its u32-length indices/proofs without pre-allocating by the claimed length, so a huge length already fails on the first read — verified empirically, left alone. The PoPowHeader interlinks collection is the same safe pattern.

Tests

• New regression tests for every fixed path (huge prefix → parse error, under the same allocation cap that aborted the old code), plus exact/short/empty pins for scorex_parse_bytes_exact.
• cargo test: sigma-ser 56, ergo-nipopow 5, ergo-lib (--all-features) 122+1, ergo-merkle-tree and ergo-chain-types green, ergotree-ir + ergotree-interpreter green. Clippy -D warnings clean on the touched crates; cargo fmt --check clean.

---

Built by @Kshot9000 — donations (Ergo): 9fcM5RWnAjmP4vx5bnW6yohB6H9bLq8sJbaPLHtwZLtQPB32Pvy

Note: cargo check -p ergo-nipopow --no-default-features fails inside ergotree-ir on pristine develop as well (verified with these changes stashed) — pre-existing and unrelated.