Electrum chain source stop() discards registered Filter entries #1005

Discovered in #992 (but pre-existing)

Claude:

The new failure-path chain_source.stop() silently discards Electrum Filter registrations, breaking a failed-start → retry cycle for nodes with existing channels.

During Builder::build, channel monitors are loaded (src/builder.rs:1987) and register their watched txs/outputs with the chain source's Filter impl. For Electrum, these are queued in ElectrumRuntimeStatus::Stopped { pending_registered_txs, pending_registered_outputs } until start() applies
them. But ElectrumRuntimeStatus::stop() does *self = Self::new() (src/chain/electrum.rs:456), wiping both the running client and any pending registrations, and nothing re-registers them afterward — the outer ChainSource::registered_txids set survives but is never re-applied (and doesn't
carry scripts/outputs).

Concrete scenario: a node with existing channels calls start(), the fee-rate update or 0FC validation fails transiently (network blip) → the new error path calls chain_source.stop() → the app retries start(), which now succeeds — but the fresh ElectrumRuntimeClient's tx-sync watches nothing.
Confirmations/spends of existing funding outputs (e.g. a counterparty force-close) go undetected until a full process restart.

To be fair, a normal stop() → start() cycle already has this property on main, so it may be a known limitation — but pre-PR the failed-start path left the registrations intact (at the cost of the leak this PR fixes), and a failed first start with a retry loop is a much more common pattern
than a deliberate stop/restart. I'd suggest having ElectrumRuntimeStatus::stop() preserve the registration set (transition back to Stopped carrying the known registrations) rather than resetting to empty — that would fix both this path and the pre-existing stop/start case.

The long-term fix would still be to get rid of the start/stop mechanism entirely (see bitcoindevkit/rust-electrum-client#166), but as a hotfix we probably should make sure we trigger re-registration after restart (via load_outputs_to_watch