Flaky test fixes for failures that surfaced once slow test packages were scheduled first (heavier contention at the start of a full-suite run), plus one pre-existing tail flake.

• engine/access/state_stream/backend: widen per-block exec-data liveness waits (1s -> 10s); they timed out under full-suite load.
• network/alsp/manager RepeatOffender: poll for the first penalty change instead of sleeping exactly one heartbeat interval, which could observe 0 or 2 decays when the heartbeat ticker was delayed.
• ledger/complete/wal checkpoint test: redraw random payloads on duplicate encoded keys. Keys can be as short as 1 byte (~1/256 collision per pair, ~130 payloads per trie), which made the payload comparison nondeterministic.
• engine/verification/fetcher/chunkconsumer: assert the multiset instead of the exact receive order in both "receive N" subtests; the consumer's 3 concurrent workers append in nondeterministic order.

All fixes preserve the original assertions' strength. Validation: per-test stress runs (3x-50x, -race where relevant) and clean 10x full-suite campaigns.

Related: #8639

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}