Add ETH as a launch spoke (SOL<->ETH swaps)

[anderdc]

What

Adds Ethereum as the third spoke: LAUNCH_SPOKES = ('btc', 'tao', 'eth'), so SOL↔ETH swaps launch with direction pools, miner quote flags, and CLI network config all derived from the registry. No contract changes — chains are opaque strings on-chain and leg amounts were already u128 (wei-scale).

Provider design (allways/chain_providers/ethereum.py)

• Lean deps: eth-account + raw JSON-RPC over requests with a narrated failover ladder (publicnode → drpc defaults, ETH_RPC_URLS override) — same shape as the BTC provider's Esplora client, no web3.py.
• Inclusion ≠ settlement: a mined tx must have receipt status == 1 (a reverted tx still decodes intact to/value). Mirrors the TAO provider's Balances.Transfer-event rule.
• Unknown ≠ absent: a mined tx whose receipt can't be read raises ProviderUnreachableError instead of returning None — on a dest leg, None is a slash-eligible verdict.
• Plain EOA value transfers only — internal/contract-mediated transfers don't populate a provable from/to/value and the sender-pin defense needs a provable EOA sender.
• check_connection verifies eth_chainId against ETH_NETWORK, so a wrong-network RPC fails at startup instead of verifying the wrong chain.
• Double-send guard: broadcasted_txids + an incremental head-cursor find_recent_outgoing (TAO-style; plain RPC has no address index).

Base change

New ChainProvider.normalize_address hook (identity default), used by verify_transaction's sender-mismatch and self-transfer checks. ETH overrides with lowercase: EIP-55 is display-only casing, and a checksummed reservation address vs a lowercase RPC sender must not false-reject a real deposit.

ChainDefinition choices

field	value	why
decimals	18	first spoke with more decimals than the hub; integer rate paths already handle negative decimal_diff (tests added)
min_confirmations	32	~1 beacon epoch; post-merge reorgs deeper than 2 blocks are vanishingly rare
min_onchain_amount	5e13 wei	no protocol dust floor; economic floor ≈ a few × a 21k-gas transfer fee
replay_grace_secs	0	block timestamps strictly increase (no BTC median-time lag)

Tests / verification

• 41 new offline provider tests (RPC layer stubbed): revert rejection, receipt-unavailable → unreachable, pending/reorg/underpay/wrong-recipient, EIP-55 casing across verify_transaction, failover ladder ("null result is authoritative"), send guards, scanner cursor.
• eth_utils 5+ is_address accepts unprefixed hex and stopped validating EIP-55 checksums — address validation is hand-rolled and tested.
• Scoring tests now derive pool floors from len(LAUNCH_SPOKES) instead of a hardcoded 2-pair world; the unsupported-chain sentinel moved 'eth' → 'doge'.
• Full suite: 886 passed. Live read-only mainnet smoke: real transfer verified through verify_transaction with an uppercased recipient + pinned sender; chain-id check; balance read.

Notes for rollout

• Miners quoting ETH need ETH_PRIVATE_KEY + a funded wallet; validators only need RPC reachability (public defaults work).
• Third pair dilutes each pair's zero-volume emission floor from (1−α)/2 to (1−α)/3 of the miner pool — expected consequence of the volume-blend design.

[anderdc]

Added c6c34a9 — settled-tx read cache, prompted by the RPC budget review against RPC_OPTIMIZATION.md's 30-concurrent-swap average model:

The original implementation spent 3 RPCs per re-verify (getTransactionByHash + getTransactionReceipt + getBlockByNumber) every 12s pass, vs BTC's single /tx. But a mined tx's receipt and its block's timestamp are immutable per block hash — so they're now fetched once and cached keyed by the blockHash the per-pass tx fetch already returns. Hash change (reorg) = miss = full refetch; pending/absent/reverted reads are never cached (the unknown≠absent and settlement rules are untouched). Same pattern family as the shipped SOL/BTC tip hoists (#542–#544).

Effect at the doc's average model (30 concurrent, all-ETH worst case): ~655k → ~225k req/day per validator; at the realistic 15-leg mix ~110k/day — inside dRPC's free keyed tier with 6× headroom. Covered by 4 new unit tests (call-count assertion, reorg invalidation, never-cache pending/reverted, bound). 890 tests green.