CacheInvariant v0.3.0

CacheInvariant v0.3.0 adds a registered direct-token prefix-divergence lane
for the exact pinned llama.cpp b10107 CPU runtime.

Evidence in this release:

• main and lightweight v0.3.0 both resolve to
  203f84a42bb7fb841abd3c1f5d221b0bed6de289;
• main CI run 30293701488 and tag CI run 30293833809 each passed 9/9
  jobs, including fresh Windows and Ubuntu exact-runtime integrations;
• the source-bound Windows schema-v3 reference verifies 77/77 registered
  checks, while the preserved schema-v2 and schema-v1 references still verify
  57/57 and 29/29 checks; and
• exact, shared-prefix, and first-token-divergence rows are compared with
  matched cache-off requests using bounded cache/prompt counters.

The direct-token lane explicitly sends n_predict=1, requires one reported
predicted token, then discards generated content and generated token values.
Every registered direct-token request also pins n_cache_reuse=0, disabling
that separate optional KV-shifting chunk-reuse path; the paired
cache_prompt=true and cache_prompt=false requests remain the tested prompt
reuse switch. This is not an output-quality or performance benchmark.

This release contains only the wheel and source distribution. Pinned runtime
binaries, fixture inputs, and the converted model are fetched and checked by
the reproducible integration path; they are not redistributed. There is no
PyPI publication or paid-service dependency.

Development and validation were AI-assisted and owner-operated. These results
do not establish independent reproduction, review, users, adoption,
production use, security, cross-runtime equivalence, model quality, latency,
throughput, memory savings, or recruiting signal.