fix(cache): prevent unreadable oversized fingerprints (#553)

## Motivation

wincode 0.5.2 enforced its default preallocation limit when
deserializing dynamic maps but not when serializing them. A large
inferred-input fingerprint could therefore be persisted successfully and
then fail on the next cache lookup with `Encoded sequence length
exceeded preallocation limit`, leaving a cache entry that its writer
could not read back.

## Changes

- bump wincode from 0.5.2 to 0.6.0, which applies the preallocation
check symmetrically

This preserves the 4 MiB preallocation policy while preventing oversized
fingerprints from poisoning later cache lookups. Formatting, workspace
checks, default tests, Clippy, and rustdoc pass locally.

Co-authored-by: GPT-5.6 Codex <codex@openai.com>

Author: wan9chi

Cargo.lock

@@ -45,7 +45,7 @@ assert2 = "0.4.0"
 assertables = "10.0.0"
 async-trait = "0.1.89"
 base64 = "0.22.1"
-wincode = "0.5.2"
+wincode = "0.6.0"
 bindgen = "0.72.1"
 bitflags = "2.10.0"
 brush-parser = "0.4.0"