L2 corruption/tamper is never evicted (poison persists; L1 self-heals but L2 does not)

[27Bslash6]

Summary

On deserialize/integrity failure, the L1 path invalidates the entry (self-heals), but the L2 path catches SerializationError and returns None (miss → recompute) without deleting the corrupt/tampered backend entry. Every subsequent reader re-pays the full backend read + decompress + xxHash3 (+ decrypt) before recomputing; under thundering-herd every waiter does it.

Evidence

• L1 heals: src/cachekit/decorators/wrapper.py:695 (sync), :955 (async)
• L2 does not delete: src/cachekit/cache_handler.py:822-824 (sync), :855-857 (async); async post-lock double-check wrapper.py:1100-1101 only warns
• Sync L2 also lacks a separate except SerializationError branch (async has one at wrapper.py:1037-1048) — corruption is indistinguishable from a network error on the sync path

Impact

Poisoned L2 entries persist until TTL/overwrite, repeatedly burning full read+verify work and masking tamper as transient miss.

Fix

On L2 integrity/auth failure, delete the offending key and emit a distinct cache_get_deserialize metric on both sync and async paths.

[27Bslash6]

Reopening: 0.9.0 (#177) shipped the eviction primitive at the get_cached_value/get_cached_value_async API, and the sync decorator path is covered (it routes through get_cached_value).

Remaining (follow-up): the async decorator re-implements the L2 read inline at decorators/wrapper.py:998 and the two post-lock double-checks (:1090/:1116), bypassing get_cached_value_async — so those paths do not yet evict poisoned entries. The right fix is to route them through get_cached_value_async (so they inherit eviction) rather than duplicating the delete across sites. Scoped down to that here.