Let the system recover from a broadcast that never confirms

The code treated a broadcast as final: money moved on-chain and the DB was
updated on the assumption it would either confirm or be fee-bumped until it
did. Neither is guaranteed, and every way that assumption broke was permanent
(BUGS.md B-02, B-03, B-04, B-07, B-08, B-20, B-21).

Persist before broadcasting. place_bet and request_withdrawal now write their
rows in a "building" state and commit, then broadcast, then promote to
broadcast/pending in a second commit. Before, a failure or crash between the
broadcast and the commit left the coins irreversibly spent with no trace: no
participant (so no entry in the draw), no pending row (so no RBF and no
confirmation tracking), and the UTXOs not even marked spent, so the next bet
would try to double-spend them. A refused broadcast now releases the reserved
UTXOs, restores the balance, removes the participant (or marks the withdrawal
failed), audit-logs it, and answers a translatable broadcast_failed — as 502,
since the network refused it, not the caller, where it used to be an opaque 500.

Reconcile what's in flight against the chain. New PendingTransactionReconciler
(app/tx/reconcile.py, every 120s and once at startup) asks whether each
non-terminal tx exists: present -> promote, gone -> mark failed with a reason,
release the inputs, roll the domain row back, audit-log it. Grace periods differ
by state (120s for "building", 6h for "pending", so the RBF bumper gets its
attempts first). It is deliberately biased to inaction: only a server that
positively doesn't know the tx counts as absent, and a transport failure never
abandons anything, because releasing a UTXO whose tx is actually alive would
invite a double-spend. Verified against the live server, which answers "No such
mempool or blockchain transaction" for an unknown txid.

Stop keying on a value that changes. An RBF bump changes the txid, and
_on_bet_confirmed looked the participant up by bet_txid — so a bumped bet
confirmed under a txid no participant carried, the row stayed "broadcast"
forever, and the scheduler waited on it forever: the round could never close and
the lottery stopped. Handlers now resolve by immutable ids (round_id/user_id,
withdrawal_id), and bump_fee retargets every stored txid — bet_txid,
Withdrawal.txid, Round.payout_txid and UtxoEvent.spent_txid — plus records the
previous one in replaced_by_txid, which was never written at all.

One bad row no longer blocks the rest. The confirmation poller's per-tx lookup
is guarded: a txid the server can't resolve used to abort the whole pass, so
nothing confirmed again until an operator intervened. It also selects plain
columns instead of hydrating entities that outlive their session.

Tests: 6 reconciler cases including "a broken connection must not release coins";
the bet-ordering test probes committed state from an independent session during
the broadcast, and caught a real mistake in the first draft of this change (the
_pending_transaction helper still hardcoded status="pending", so rows were born
already-broadcast and would have got the 6-hour grace instead of 120s).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/unit/test_withdrawals.py

@@ -4,7 +4,7 @@ from sqlalchemy.ext.asyncio import async_sessionmaker, create_async_engine
 
 from app.config import settings
 from app.db.base import Base
-from app.db.models import PendingTransaction, User, UtxoEvent
+from app.db.models import PendingTransaction, User, UtxoEvent, Withdrawal
 from app.wallet.hd import derive_user_address
 from app.withdrawals.service import WithdrawalError, request_withdrawal
 
@@ -122,3 +122,44 @@ async def test_withdrawal_rejects_non_plm_address(session_factory, address):
 
     assert exc_info.value.code == "invalid_address"
     assert not client.broadcasted
+
+
+async def test_withdrawal_to_own_address_is_rejected(session_factory):
+    """B-17: allowed before, and it broke two things that assume the recipient and
+    the change are distinguishable by address — the RBF bump would shrink the
+    recipient output, and compute_pending_balance counted the amount twice."""
+    user_id = await _make_funded_user(session_factory, 8, 3_000_000_000)
+    client = FakeElectrumClient()
+
+    async with session_factory() as session:
+        user = await session.get(User, user_id)
+        with pytest.raises(WithdrawalError, match="own deposit address"):
+            await request_withdrawal(session, client, user, user.address, 1_000_000_000)
+
+    assert not client.broadcasted
+    async with session_factory() as session:
+        assert (await session.scalars(select(UtxoEvent).where(UtxoEvent.user_id == user_id))).one().spent_txid is None
+
+
+async def test_failed_broadcast_marks_the_withdrawal_failed_and_frees_the_coins(session_factory):
+    """B-07/B-08: the Withdrawal row is kept (unlike a bet) so the user can see the
+    instruction didn't go through, but the coins must come back."""
+    user_id = await _make_funded_user(session_factory, 9, 3_000_000_000)
+
+    class RejectingClient:
+        async def broadcast(self, raw_tx_hex: str) -> str:
+            raise RuntimeError("min relay fee not met")
+
+    async with session_factory() as session:
+        user = await session.get(User, user_id)
+        external = derive_user_address(99)
+        with pytest.raises(WithdrawalError, match="refused"):
+            await request_withdrawal(session, RejectingClient(), user, external, 1_000_000_000)
+
+    async with session_factory() as session:
+        withdrawal = (await session.scalars(select(Withdrawal))).one()
+        assert withdrawal.status == "failed"
+        assert withdrawal.txid is None
+        assert (await session.scalars(select(UtxoEvent).where(UtxoEvent.user_id == user_id))).one().spent_txid is None
+        user = await session.get(User, user_id)
+        assert user.cached_balance_sats == 3_000_000_000