Merged elements to 29.4

[tomt1664]

Upstream merges from Bitcoin 29.x to 29.4.

Additional fixes for CI:

1. Merge setup & CI infrastructure

• Merge script sanity check — confirmed all 42 upstream backport PRs (e9e6825b8c..3fc0865963) were merge commits with no stray non-merge first-parent commits; merge-prs.sh pointed at elements-29.x.
• CI migration, Cirrus CI → GitHub Actions — resolved Docker/host environment conflicts, macOS path divergence (~/.elements vs ~/Library/Application Support/Elements), subtree lint failures (secp256k1, minisketch), test-each-commit FETCH_DEPTH issues, Windows MSVC job config.
• ci/test/03_test_script.sh — removed dead autotools-era code paths (./autogen.sh, ./configure), rewrote to match upstream's CMake-only flow.
• doc/man generation — resolved merge conflicts from upstream's doc/man/bitcoin-*.1 vs Elements' elements-*.1 naming; regenerate via contrib/devtools/gen-manpages.py at release time instead of merging.
• contrib/devtools/check-deps.sh — added ~30 missing SUPPRESS[...] entries recording Elements-specific cross-module dependencies (confidential_validation, pegins, mainchainrpc, pak, dynafed, block_proof, versionbits) that upstream's dependency-graph checker had no knowledge of.
• CentOS/depends/GUI job — "No space left on device" — infrastructure issue (disk/ccache exhaustion on the runner), not a code bug; not something fixable via source changes.
• --exclude feature_dbcrash — feature_dbcrash.py is commented out of Elements' own test_runner.py test list (too disk/time-intensive for GHA); excluding an already-absent test is now a hard error in test_runner.py, not a no-op. Removed from the exclude list in ci/test/00_setup_env_native_previous_releases.sh.
• Same file — --extended,feature_fee_estimation,... argparse corruption

---

2. RPATH — "libsecp256k1.so.6: cannot open shared object file" bug

top-level CMakeLists.txt sets CMAKE_SKIP_BUILD_RPATH TRUE globally (deliberate upstream policy, tied to Guix reproducible builds). This is invisible under a normal (static-secp256k1) build, but any build enabling libbitcoinkernel forces secp256k1 to build as a shared library — and every other executable target then fails at runtime unless it gets the same opt-out.

• elements-tx, elements-util (src/CMakeLists.txt)
• test_elements (src/test/CMakeLists.txt)
• bench_bitcoin (src/bench/CMakeLists.txt)
• elements-qt (src/qt/CMakeLists.txt)
• test_elements-qt (src/qt/test/CMakeLists.txt)
• elementsd (src/CMakeLists.txt)
• elements-cli (src/CMakeLists.txt)

---

3. Code/sanitizer bugs

• CTxOut 2-arg vs 3-arg constructor — ported Bitcoin Core test/IPC code used upstream's CTxOut(amount, script); fixed to Elements' CTxOut(CAsset(), amount, script) in transaction_tests.cpp, test/ipc_test.cpp.
• checkqueue.h::Loop() ownership bug — failing/skipped-after-failure checks never deleted in Elements' pointer-based (T*) CCheckQueue; real leak reachable from production validation code.
• spend.cpp — vtxinwit/vin invariant violation — vtxinwit.emplace_back() only ran inside a conditional branch; moved to run unconditionally alongside every vin.emplace_back().
• wallet.cpp::SignPSBT — missing tx.witness.vtxinwit.resize(tx.vin.size()) (had the vtxoutwit equivalent but not the vtxinwit one); found via gdb backtrace.
• coins.h::AddFlags — Assume(flags & (DIRTY|FRESH)) didn't account for Elements' third PEGIN flag; fixed to Assume(flags & (DIRTY|FRESH|PEGIN)).
• kernel/chainparams.h::CChainParams — missing virtual destructor; new-delete-type-mismatch deleting a CCustomParams through unique_ptr<const CChainParams>.
• validation.cpp::CheckInputScripts synchronous path — leaked the heap-allocated CScriptCheck on every call (missing delete).
• test/fuzz/deserialize.cpp — generic round-trip fuzz check inappropriately applied to Coin, which asserts !IsSpent() on serialize; fuzzed bytes trivially violate this.
• test/fuzz/package_eval.cpp — duplicate tx_mut.vin.push_back(in) — an earlier fix (adding vtxinwit.emplace_back() alongside the push_back) was applied additively instead of replacing the original line, leaving every selected input pushed into vin twice. Traced via reproducible fuzz crash + gdb + targeted fprintf instrumentation to a outpoints.insert(in.prevout).second assertion failure; root-caused to the duplicate push, not (as initially suspected) OutpointsUpdater or cross-iteration state.
• kernel/chainparams.cpp::CTestNet4Params — entire Elements-specific initialization block (enforce_pak, accept_unlimited_issuances, multi_data_permitted, accept_discount_ct, create_discount_ct, pegin_subsidy, pegin_minimum, anyonecanspend_aremine, consensus.has_parent_chain, consensus.genesis_subsidy, consensus.connect_genesis_outputs, consensus.subsidy_asset, g_signed_blocks) missing entirely — testnet4 is a brand-new upstream 29.x chain type, and the new class was adapted from upstream's version (which has no concept of any of these fields) without ever adding Elements' equivalent block. Caught independently by both UBSan (invalid-bool-load) and MSan (use-of-uninitialized-value) in two separate CI jobs.
• validation.cpp::GuessVerificationProgress — unguarded division (pindex->nHeight / (pindex->nHeight + moreBlocksExpected)) with a denominator that can hit exactly zero when a block's timestamp is far enough ahead of "now" (e.g. during feature_taproot.py's test-vector generation). Added a guard clause returning 1.0 when the extrapolation is degenerate.
• wallet/rpc/spend.cpp — psbtbumpfee not-mine-input path — FillPSBT(...) call omitted Elements' imbalance_ok parameter (defaults to false), causing a legitimate CT-balance-check failure for external/not-owned inputs to trip a CHECK_NONFATAL(!err) that assumed the call could never fail. Every other Elements-specific FillPSBT call site in the same file already passes /*imbalance_ok=*/true; this one didn't. Root-caused by diffing against upstream (confirmed the CHECK_NONFATAL pattern itself is unmodified) and against Elements' own FillPSBT declaration.

---

4. Type/size mismatches

Several issues traced to CoinsCachePair/Coin/CTxOut being substantially larger than upstream's equivalents (confidential asset/value/nonce commitments), breaking constants and assumptions tuned for upstream's smaller structs:

• CTxMemPoolEntry::discountSizeWithAncestors — declared uint64_t while its three sibling ancestor-state accumulators are int64_t; the odd type caused a UBSan implicit-integer-sign-change when a negative delta was added during removal. Retyped to int64_t to match siblings.
  • Follow-up compile fix: GetDiscountTxSize() returns size_t; brace-init into the now-int64_t member needed an explicit static_cast<int64_t>(...).
  • Confirmed the six external call sites (node/miner.h, node/miner.cpp, txmempool.h) did not need changes — they only ever read the post-assert, always-non-negative value into uint64_t, which is a value-preserving conversion UBSan doesn't flag.
• coins.h::CCoinsMapMemoryResource — PoolResource's generic default constructor hardcodes a 256 KiB chunk, sized for upstream's smaller CoinsCachePair; no longer big enough to hold 1000 of Elements' larger nodes without overflowing into a second chunk. Fixed by passing an explicit chunk_size_bytes scaled to sizeof(CoinsCachePair).
  • Follow-up: validation_flush_tests.cpp's getcoinscachesizestate test hardcoded MAX_COINS_CACHE_BYTES = 262144 + 512 based on the (now-stale) "PoolResource defaults to 256 KiB" assumption; updated to compute the same way as the fixed default.
• 32-bit ARM validation_flush_tests.cpp — the test's own "unsupported architecture" fallback path hardcoded 1000 coins as "obviously enough" to cross the CRITICAL cache-size threshold; doesn't hold on 32-bit ARM where COIN_SIZE is smaller (32-bit pointers). Fixed by computing the coin count from MAX_COINS_CACHE_BYTES / COIN_SIZE instead of a fixed constant.

---

5. Build configuration / CMake

• src/CMakeLists.txt — simplicity subtree warnings — -Wconditional-uninitialized/-Wimplicit-fallthrough (Clang) and -Wtype-limits (GCC/mingw) in vendored dag.c/rsort.c; suppression was gated if(APPLE) only, extended to all Clang toolchains plus a GCC-specific suppression.
• src/qt/CMakeLists.txt — WITH_MULTIPROCESS block called import_plugins(bitcoin-gui), never renamed to Elements' elements-gui target.
• src/kernel/CMakeLists.txt — libbitcoinkernel missing several source files (addresstype.cpp, key.cpp, script/miniscript.cpp, common/{config,settings}.cpp, chainparamsbase.cpp, mainchainrpc.cpp) and libevent linkage, needed because pegins.cpp (pulled into the kernel lib) calls IsConfirmedBitcoinBlock unconditionally.
• Windows cross-compilation — secp256k1 __imp_ linker symbols (fixed via PUBLIC CMake linkage); pegins.cpp/confidential_validation.cpp needed to move from bitcoin_node to bitcoin_common to satisfy the dependency graph; NSIS installer asset staging.

6. Toolchain / fuzz harness

• test/fuzz/simplicity_tx.cpp — missing SeedRandomStateForTest(SeedRand::ZEROS); as the first line of the FUZZ_TARGET body, required by the CheckGlobals reproducibility guard whenever a target touches the global RNG state. Also needed the corresponding #include <test/util/random.h>

---

7. Python functional-test gaps

• test/functional/p2p_headers_sync_with_minchainwork.py and p2p_headers_presync.py — both call self.nodes[0].createwallet(...) unconditionally (an Elements-specific addition needed because Elements' generatetoaddress-based mining approach requires a wallet address, unlike upstream's self.generate()), but neither had the standard self.skip_if_no_wallet() guard that every other wallet-touching test in the suite already has, causing Method not found (-32601) failures in no-wallet CI builds. Both fixed by adding the guard as the first line of run_test(). Flagged for a broader sweep (grep -rL skip_if_no_wallet against files that call wallet methods) since both were adapted from the same original source and the same gap may recur elsewhere.

---

8. Static analysis / clang-tidy

Batch of ~9 findings from a -warnings-as-errors clang-tidy run,

• rpc/util.cpp — two hand-written trivial default constructors → = default.
• qt/networkstyle.cpp — legitimate bounded single-level recursion (instantiate(), guarded by an assert immediately before the recursive call) → NOLINT(misc-no-recursion) rather than a rewrite.
• wallet/coinselection.cpp — two occurrences of auto input_set/set = result.GetInputSet(); copy-constructed but only used by reference (.find()/.end() need to operate on the same instance, so the copy can't simply be inlined away either) → const auto&. One occurrence was caught by CI in a second pass after the first fix (§10 recurrence).
• assetsdir.cpp — range-for copying each string → const std::string&.
• primitives/bitcoin/merkleblock.cpp — three inherently-recursive Merkle-tree traversal functions, straight from unmodified upstream → NOLINT(misc-no-recursion) rather than rewriting a correctness-sensitive algorithm.
• validation.cpp — emplace_back in a loop of known size without reserve().
• bitcoin-tx.cpp — three occurrences of the same unnecessary-copy pattern as coinselection.cpp.
• test/blind_tests.cpp — push_back in a fixed 6-iteration loop without reserve().