feat(runtime): persist a per-subscription log cursor for gap-free resume across restarts

[mfw78]

What

Makes chain-log delivery gap-free across engine restarts, not just within a process. A module opts a subscription in with one manifest flag:

[[subscription]]
kind = "chain-log"
chain_id = 11155111
address = "0x..."
event_signature = "0x..."
resume = true

The host then persists a durable cursor (the last dispatched block for that subscription) and, on boot, re-opens the watch_canonical_logs_from poller from that block instead of at the current head, clamped by MAX_SYNC_BACK_BLOCKS.

Why

Closes #300. #299 made log delivery gap-free within a process (the canonical poller + retry layer + sync-back-on-reopen). Across a process restart the poller still opened at head, dropping logs mined during downtime. This closes that gap for subscriptions that want it.

Design (host owns the cursor, module owns idempotency)

• Host-owned cursor. Keyed by keccak(chain_id | address | topic) under the module's own store namespace, alongside the existing last_dispatched_block block marker. It is not derived from the alloy Filter: Filter's Hash uses a process-randomized HashSet, so it is not reproducible across restarts. The module never manages a cursor or a boot handshake.
• Written after dispatch, not on send. The reconnect task's last_seen_block advances when it enqueues a log, but a crash with logs still buffered would record un-dispatched blocks as done. So the supervisor writes the cursor only after a successful dispatch_chain_log, mirroring the block marker. Single writer (supervisor), one-shot boot read (supervisor); the spawned task never touches the store.
• At-least-once. Boot resumes at the cursor block and replays it in full (a mid-block crash pre-restart loses nothing); redelivery is deduped by the module's chassis idempotency journal. Exactly-once (a transactional cursor + module-state commit) is explicitly out of scope.
• Default off. resume = false reproduces today's behaviour exactly (initial_cursor = None -> poller starts at head, no history replay).

Testing

cargo clippy --workspace --all-targets --all-features --locked -- -D warnings clean, cargo test -p nexum-runtime (206 passed, including chainlog_cursor_key stability/uniqueness tests), full cargo build --workspace, fmt, no em-dashes, all under the pinned 1.94 toolchain.

Depends on

#299 (the canonical poller, poller_resume_block, and MAX_SYNC_BACK_BLOCKS). This PR is stacked on it.

AI Assistance

Claude Code used for the seam mapping, the implementation, and this description.