fix(validation): cascade-invalidate + ReplayCache mergeable-entry safety

[spreston8]

Summary

Two block-validation correctness fixes surfaced during the bonding-bug investigation. Both close concrete failure modes that produced spurious InvalidTransaction cascades on joiners under load.

Stacked on top of #507 (feat(lfs): joiner-side coverage + orchestrator multi-root pagination fix for multi-bonded shards). This PR's diff shows ONLY the 2 commits below; auto-rebases onto rust/staging once #507 merges.

Commits (2)

776fa361 fix(runtime-manager): ReplayCache mergeable-entry safety check
ad1627ac fix(block-processor): cascade-invalidate children of invalid parents

ad1627ac — Cascade-invalidate children of invalid parents

Problem. When a block X is recorded invalid, its children sit in the buffer waiting on X. Each child eventually gets pulled out for validation, individually fails because its parent is missing a mergeable-channels entry (since invalid parents never wrote one — save_mergeable_channels only runs on the validation success path), and gets recorded as InvalidTransaction one-at-a-time. Slashable cascade for what was originally one parent's failure. Honest validators producing children of a block they didn't know was invalid in their view get punished.

Fix.

• New InvalidBlock::InvalidParent variant — non-slashable (the child's creator may not have known the parent was invalid in their view)
• New Readiness enum returned by get_non_validated_dependencies: Ready / Waiting{deps_to_fetch, deps_in_buffer} / InvalidParents(parents)
• Cascade fires only on invalid PARENTS — invalid justifications stay satisfied (justifications acknowledge a validator's latest message and don't require a mergeable entry)
• effects_for_invalid_block refactored to take &KeyValueDagRepresentation directly so the cascade path doesn't need to build a full snapshot
• BlockException arm in validate_with_effects escalated warn → error as a tripwire (should be unreachable post-fix)
• handle_invalid_block routes InvalidParent through handle_invalid_block_effect so descendants also cascade

New test: check_dependencies_must_cascade_invalidate_when_parent_is_invalid.

776fa361 — ReplayCache mergeable-entry safety check

Problem. ReplayCache is a per-node optimization — on a cache hit, return the cached post-state without re-running replay. But if the mergeable-channels entry for (post_state, sender, seq_num) isn't in the local store, returning the cached post-state lets the block enter the DAG with no mergeable entry. Subsequent children that need the entry hit "Missing mergeable entry" during merge — the exact failure mode the existing StateHashCache check (lines 624-643) was designed to prevent.

Fix. On ReplayCache hit, look up the mergeable key first. If present → fast-return as before. If absent → fall through to full replay so save_mergeable_channels runs. Mirrors the existing StateHashCache safety check pattern.

Test plan

• cargo test --release --workspace — 2183 passed, 0 failed (across 70 test suites; +1 over PR feat(lfs): joiner-side coverage + orchestrator multi-root pagination fix for multi-bonded shards #507's 2182, accounting for the new cascade-invalidate test)
• cargo build --release -p casper — clean
• No new files; cherry-picks rustfmt-clean

Co-Authored-By: Claude noreply@anthropic.com