fix(node): evict imported blocks' transactions from mempool + periodic expiry

[lai3d]

Fixes #135.

Problem

Mempool eviction only happened at self-production (producer.rs:457). Blocks imported from other validators never evicted their txs, and select_with_nonce skips stale-nonce txs without removing them — so with 3 validators in round-robin, ~2/3 of every node's pool became immortal zombies. Observed on the post-reset testnet: oldest_tx_age climbing linearly past 45 min (T2 alert at 15m permanently red), node-3 depth stepping up monotonically.

Fix

• SyncManager::evict_imported_txs — on every successful import_block (all 4 sites: gossip handle_block, fetched-block task, pending-blocks drain, range-sync catch-up) remove the block's txs from the local pool, keyed by blake3(tx.to_bytes_without_signature()) — the same key space the pool uses at insert and producer.rs uses at production.
• Periodic remove_expired() in the catch-up tick (5s) as a backstop for txs that will never be mined under their exact hash (e.g. superseded duplicates). Previously the pool's 1h tx_lifetime had zero callers.
• Contract test in pool.rs pinning the recomputed-hash key space, so a future serialization change that silently breaks eviction fails CI.

Verification

• cargo test -p qfc-mempool 7/7 (incl. new test)
• cargo test -p qfc-node unit 18/18, integration 6/6 (release binary)
• Post-deploy check: T2 dashboard oldest pending tx age should saturate and reset instead of climbing forever; node-3 depth should drain. Will verify on the next image rollout during the soak.

🤖 Generated with Claude Code