Bench2Agent 0.3.0

Bench2Agent 0.3.0 renames the application, Python package, CLI and MCP server
from the previous project names. The local checkout folder can keep its name.

Install using either method:

uv tool install bench2agent
bench2agent

python -m pip install bench2agent
bench2agent

The first conversation downloads the checksum-verified research snapshot.
Use bench2agent claude or bench2agent codex for an existing client login;
bench2agent mcp exposes the same six read-only benchmark tools.

Existing users can copy research data, model settings and conversations with
bench2agent migrate --from /path/to/previous/data-home. This preserves source
files and skips existing destination files. Re-register MCP under bench2agent
and restart the client. See the migration guide.

The tested data is data-20261007:
28 editions, 58,616 parsed paper-edition observations and 9,332 catalogue entries.
The renamed data download has the same bytes, checksum and snapshot identity
as the original release asset. Benchmark locations and reviewed introduction
paper links remain part of tool results; coverage limits remain explicit.

bench2agent serve retains the local browser prototype. The shared relational
database and hosted web conversation are planned in the architecture design;
this release still queries the JSON snapshot.

Validation: 82 Python tests, 3 VS Code tests, strict package metadata checks,
fresh wheel installation, migration and representative real-snapshot MCP queries.
Python 3.10+; standard-library runtime. The package contains code, not corpus
files or downloaded research papers. Earlier published packages remain available.

Wheel and source-distribution hashes are in the attached SHA256SUMS.