Fix the defects and test failures CI found once it could run

[kwsantiago]

Continuous Integration had never run to completion here: every pull request failed at Check changelog, step 4 of Pre-build checks, so flake8, the builds, the unit tests and the integration suite were all skipped. With that cleared the suite ran, and this is what it found.

Defects

Reachable in production, not test-only.

• hsmd rebuilt the whole transaction for every input. sign_unified_wallet_input() cloned the entire PSBT per input. Held on tmpctx, which is released when the request completes rather than per input, that was 700 copies at once for a 700-input withdrawal: 13.2 GB resident in lightning_hsmd, 200 MB after, and consolidating a few hundred UTXOs was enough to exhaust the machine. Freeing each clone fixed the memory but left the work quadratic, so the clone is now built once per signing pass and dropped only where the anchored-close branch changes a prevout script or amount, which is all of it the digest commits to. test_bitcoin_backend_gianttx goes 358s to 59s, and from the 1800s Valgrind timeout to 561s.
• --dev-warn-on-overgrind did nothing for wallet inputs. The unified signing branch returns before the check the legacy path reaches, so a short signature was never reported.
• hsmd failed open after the seed was freed. tal_bytelen(NULL) is 0, so use_bip86_derivation(0) was false and the signer fell back to legacy derivation instead of refusing. It now clears initialized first.
• gossipd could abort a syncing node. probe_random_scids passes the available block count to pseudorand, which asserts it is non-zero; that count is zero when the tip sits exactly num_blocks past activation. Verified at heights 962,640 and 971,640.
• Signing rehashed every prevout transaction for every signature, quadratic in whole transactions on a large withdrawal.
• A standard BIP174 PSBT took the node down. signpsbt on an input carrying only non_witness_utxo reached the signer without the prevout metadata the unified path authenticates against, and hsmd answers that with master_badmsg(). walletrpc filled witness_utxo in only when both records were absent, so the shape BIP174 recommends, and several tools emit, was fatal. Reproduced and fixed; regression test added.
• The prevout consistency check was fail-open for inputs other than the one being signed. Skipping the wally_txid recomputation to make signing linear also skipped the bounds check and the witness_utxo vs utxo->outputs[index] comparison. Only the hashing is expensive, so the comparisons are back for every input.
• sign_tx_input() read the PSBT before bounds-checking the index, where check_tx_sig() already checked it. Two hsmd handlers pass an index straight off the wire.

Tests

• test_coinmoves: five balances were 7 sat high. One extra witness byte, since 0x21 forces a 65 byte p2tr signature where SIGHASH_DEFAULT uses 64. The same adjustment already exists in test_wallet.py.
• test_gossip: bit 68 is even, so it appears in node_announcement and widens every stored record.
• test_gossip_lease_rates, test_bookkeeping_missed_chans_leases, test_channel_lease_post_expiry: that same witness byte raises the default funding_weight from 584 to 586, and so the lease fee by 4 sat at 2000 perkw.
• test_unified_emergencyrecover, test_unified_penalty: one unlinks lightningd.sqlite3 and the other copies it, so both carry the same sqlite guard the release already uses for its own.
• test_blake2b_required_peer_bit: asserted connect raises, which only holds when the peer's hangup beats the RPC reply.
• test_sendpay_grouping, test_unified_onchain_htlc_success: both raced asynchronous state; the second pinned an exact txid while onchaind was fee-bumping it.
• test_currencyrate, test_pay: an API that now needs a key, and an invoice that expired in May.
• simple_reorg() asked prioritisetransaction to hold a transaction back and this backend mines it anyway; canned blocks are 487 days old, so the node sat in initial block download.
• did_short_sig() polls rather than sleeping a fixed second. This diverges from v26.06.7 deliberately: signing here is slower, so the warning can arrive after the sleep upstream relies on.
• test_openchannel2_inflight_limit: the hand-built open_channel2 pinned a commitment feerate of 253 and a channel type of [12, 22]. This backend's floor is higher, and the type needs unified_sigs; both are now derived rather than hardcoded.
• test_onchain_missing_no_p2tr_migrate, test_rescan_missing_utxo: these had gained a skipIf(TEST_NETWORK == 'regtest') alongside an existing != 'regtest', so they ran nowhere while reporting as skipped. They now set the activation height past their canned SHA256 headers, the way test_db.py already does, and run again.
• test_splice_stored_feerate_repaired_on_upgrade: uses db_manip, which is sqlite-only in pyln. Every other caller in the tree is guarded; this one was not. Upstream has since added the same guard, so this is a backport rather than a divergence.
• pyln: the bitcoind fixture reaped a process that a test had never started. Proposed upstream separately.

CI

.github/workflows/ci.yaml now matches v26.06.7 exactly, apart from BLAKE2B_CI, the Blake2b backend job replacing the minimum-supported-Bitcoin job, and the dropped elements job below. --maxfail=1 is dropped, which the release does not set and which aborted a whole group at its first casualty. paperclip-build.yml targeted a branch this repository does not have.

Elements

The elements integration job is dropped. Elements is a different chain, so Bitcoin's proof of work becoming BLAKE2b does not reach it and every unified-signing path refuses it: the job tested nothing this branch changes; it was also the one remaining red job, failing test_bitcoin_bad_estimatefee on a feerate floor, in a test and plugin identical to v26.06.7.

Removing it surfaced something worth recording rather than losing: default_features() runs before --network is parsed, so a node built from this branch advertises option_blake2b and option_unified_sigs as compulsory on elements while being unable to honor either. Nothing this branch targets runs there, but the job should come back once that is gated.

Changelog-None