plamentsv / gossip-cache-invariance

Injectable Skill: Gossip Cache Invariance

L1 trigger: P2P flag AND (seen_cache, message_cache, tx_cache, gossipsub, pubsub, dedup, seen_chunks detected)
Inject Into: depth-network-surface, depth-consensus-invariant
Language: Go and Rust
Finding prefix: [GCI-N]

1. Write-After-Validate

For each attacker-controlled cache key, prove the ordering between cache write
and integrity/authenticity validation. Cache insertion before validation is a
poisoning primitive.

Tag: [GOSSIP-CACHE:ORDER]

2. Duplicate Handling

Check whether duplicate suppression is keyed on canonical content, the right
topic/partition, and enough sender identity. Look for cases where legitimate
messages are suppressed or malicious messages can be replayed after eviction.

Tag: [GOSSIP-CACHE:DUPLICATES]

3. Eviction Invariants

Under burst traffic or memory pressure:

• what gets evicted first?
• can attacker traffic evict honest messages before processing?
• can eviction reopen a previously rejected malicious message?

Tag: [GOSSIP-CACHE:EVICTION]

4. Cache Key Soundness

Verify the key includes the right tuple of topic / partition / chain / message
identity. Flag collisions where distinct messages alias to the same cache key.

Tag: [GOSSIP-CACHE:KEY]

5. Broken Safety Property

State the resulting invariant failure explicitly:

• valid message can be suppressed
• invalid message can block later valid message
• message can be processed repeatedly
• attacker can force honest work without progress

6. Cache-miss amplification

For each receive-and-rebroadcast path:

1. Does the handler consult a seen/validated cache before re-broadcast?

2. Is the cache key canonical, or can the attacker create many distinct IDs
   for effectively the same payload?

3. If the cache misses, does one inbound message cause fan-out to many peers?

If a cache miss multiplies traffic or work across peers, record it as an
amplification finding rather than a simple duplicate-handling bug.

Tag: [GOSSIP-CACHE:AMPLIFICATION]

7. Partition-poisoning under re-org

When a chunk / block / attestation cache is indexed by a property derived
from the canonical chain (slot, epoch, partition id, shard id), a reorg
can move the canonical chain underneath entries that are no longer valid
for the new chain but still satisfy the cache key.

Check:

1. When the canonical head changes, is the cache invalidated, re-keyed,
   or left stale?

2. Can an attacker deliberately produce two competing valid branches so
   that a later honest reader hits a poisoned cache entry from the
   non-canonical branch?

3. For storage / DA gossip: is the {piece_id, partition_id} tuple the
   full cache key, or is partition_id inferred from head and therefore
   mutable? If inferred, a reorg silently re-labels cached pieces.

4. For mempool / tx gossip: does the seen cache survive a reorg that
   re-opens a tx that was previously mined? If yes, the tx is permanently
   un-gossipable after the reorg — a targeted censorship primitive.

5. For fork-choice attestation caches: after a reorg, do previously-seen
   attestations still count toward weight in the new branch? If yes, the
   weight is double-counted; if no, honest votes are lost.

Tag: [GOSSIP-CACHE:REORG-POISON:{cache}:{scenario}]. Severity defaults
to High when the cache influences consensus weight or censors a legitimate
message class; Medium when it only amplifies work.