sim-rs: Improve simulation performance, observability, and memory management

[sandtreader]

Summary

This PR addresses three categories of issues encountered during long-running simulations:
a clock coordinator performance bottleneck, unbounded memory growth from accumulated
transactions, and gaps in end-of-run observability.

Performance: Clock coordinator optimisation

The ClockCoordinator previously received FinishTask events through its mpsc channel —
the same channel used for actor registration and time-wait requests. In simulations with
high CPU task throughput, this created a bottleneck: every task completion required the
coordinator to wake, match the event, and update its counter.

Fix: Replace the channel-based FinishTask path with an AtomicUsize counter
decremented directly by the completing actor, plus a tokio::sync::Notify to wake the
coordinator only when it's actually blocked waiting for tasks to drain. This eliminates
per-task-completion channel traffic entirely.

Also fixes a bug in ClockBarrier::wait() where ts == self.now() should have been
ts <= self.now(), causing waits for already-passed timestamps to block instead of
completing immediately.

Memory: Slot-based transaction pruning

During long simulations (hundreds of slots), every node's
txs: HashMap<TransactionId, TransactionView> grows without bound — nodes accumulate
every transaction they've ever seen. For 100-node simulations running 500+ slots, this
causes significant memory pressure.

Fix: Add a prune_old_txs() pass at the start of each slot that removes transactions
older than linear-tx-max-age-slots (configurable, default disabled). Transactions still
in the mempool are retained regardless of age. The Mempool struct gains a
HashSet<TransactionId> for O(1) membership checks.

The age threshold of 23 slots used in linear.yaml is derived from: vote stage (5) +
diffuse stage (5) + 3× header diffusion time (3) + buffer (10).

Observability improvements

• Suppress IB stats for non-IB variants: LeiosVariant::has_ibs() gates IB-related
  statistics, so variants like FullWithoutIbs and Linear no longer show misleading
  zero-IB stats.
• Report uncertified EBs: Counts and reports EBs whose votes fell below the
  certification threshold.
• Vote failure breakdown: End-of-run stats now show a per-reason breakdown of why
  votes were not generated (e.g., InvalidSlot, ExtraIB, MissingIB).

Cleanup

Remove the vestigial -t/--timescale CLI flag — the simulator runs in virtual time and
this flag was unused.

New configuration

Parameter	Type	Default	Description
linear-tx-max-age-slots	number | null	null	Max slot age before a TX is eligible for pruning. null disables pruning.

Test plan

• cargo test passes
• Run a long simulation (500+ slots) and verify memory usage stabilises
• Verify end-of-run stats show vote failure breakdown and uncertified EB count
• Verify IB stats are suppressed for Linear variant

🤖 Generated with Claude Code