fix(client): ignore the patch version when preferring a 1.0 interface

[kosesena]

Description

ClientFactory._find_best_interface prefers a 1.0 interface with an exact string comparison (protocol_version == '1.0'). When an agent card lists a legacy interface and a 1.0 interface written with a patch version, e.g.

supported_interfaces=[
    AgentInterface(url='http://agent/v03', protocol_binding='JSONRPC', protocol_version='0.3.0'),
    AgentInterface(url='http://agent/v1',  protocol_binding='JSONRPC', protocol_version='1.0.0'),
]

the 1.0.0 entry is neither equal to '1.0' nor greater than 1.0, so the loop falls through to the first >= 0.3 candidate and the client is built with CompatJsonRpcTransport against /v03, i.e. it speaks v0.3 to an agent that supports 1.0. ('1.0' and '1.1' are handled correctly.)

The specification's versioning section says patch versions "MUST not be considered when clients and servers negotiate protocol versions". Patch versions do show up in practice; several tests in this repo build cards with protocol_version='1.0.0'.

• Compare the Major.Minor release (packaging.version.Version(...).release[:2]) instead of the raw string, via a small _major_minor helper (kept out of the loop for ruff's PERF203).
• Regression test: with 0.3.0 listed first, both 1.0 and 1.0.0 must select the 1.0 interface and JsonRpcTransport. The 1.0.0 case fails on main with CompatJsonRpcTransport.

Checks: scripts/lint.sh clean, uv run pytest 2183 passed (plus tests/integration/cross_version 5 passed), coverage 93%.

• Follow the CONTRIBUTING Guide.
• Make your Pull Request title in the https://www.conventionalcommits.org/ specification.
• Ensure the tests and linter pass
• Appropriate docs were updated (if necessary)

[github-actions]

🧪 Code Coverage (vs main)

⬇️ Download Full Report

	Base	PR	Delta
src/a2a/client/client_factory.py	88.67%	88.82%	🟢 +0.15%
src/a2a/server/cluster/database_event_stream.py	96.94%	92.86%	🔴 -4.08%
Total	93.04%	93.00%	🔴 -0.04%

Generated by coverage-comment.yml

[kosesena]

Thanks for the review! Both done in a9e912e: _major_minor now returns (v.major, v.minor), and the >1.0 and >=0.3 fallbacks use it too, so the patch version is out of every comparison. One side effect worth knowing: a pre-release like 0.3rc1 now counts as 0.3 in the fallback, which I think matches the intent. Happy to change that if you'd rather keep pre-releases below the release.

[mykytanetipa]

design comment:
I don't think we need an additional _major_minor_of helper (understand that it's done to satisfy the type checker).
Let's instead have major_minor return just tuple[int, int] with no try / except in it:

def major_minor(version: str) -> tuple[int, int]:
    """Returns the `Major.Minor` release of a protocol version."""
    v = Version(version)
    return (v.major, v.minor)

This allows us to do:

v1_0 = major_minor(PROTOCOL_VERSION_1_0)
v0_3 = major_minor(PROTOCOL_VERSION_0_3)

and catch InvalidVersion in a single candidate loop (which also avoids Ruff's PERF203 and only parses each version once):

for i in candidates:
    if not i.protocol_version:
        if best_no_version is None:
            best_no_version = i
        continue
    try:
        v = major_minor(i.protocol_version)
        if v == v1_0:
            return i
        if best_gt_1_0 is None and v > v1_0:
            best_gt_1_0 = i
        if best_ge_0_3 is None and v >= v0_3:
            best_ge_0_3 = i
    except InvalidVersion:
        pass

Additional note: let's also rename _major_minor to major_minor.

[kosesena]

Done in fdcb6ab: dropped _major_minor_of, _major_minor now returns tuple[int, int] with no try/except, and a single loop parses each version once, returns on the first 1.0 match and skips InvalidVersion. Tests, ruff and ty pass locally.

[mykytanetipa]

can you please also rename _major_minor to major_minor, no need for "_" since this is a private function syntax

per my previous comment "Additional note: let's also rename _major_minor to major_minor."

[kosesena]

Renamed to major_minor in 9e5098b, sorry I missed that note earlier.