test(signing): add cross-SDK signature vectors

[ogasurfproject-jpg]

Adds tests that pin what a change to _canonicalize_agent_card (#1278) has to keep: a card signed by another SDK still verifies here. Test only; no library code changes.

• tests/utils/signing_interop_vectors.json: the five s0-control vectors of the a2a-card-sign-v01 corpus (test(conformance): add Agent Card signature vectors (a2a-card-sign-v01) a2a-tck#246), each lifted verbatim, the same way jcs_vectors.json was lifted from Add RFC 8785 canonicalization conformance vectors (a2a-jcs-v01, Layer A) a2a-tck#228. One card with every REQUIRED field and no field at its default value, signed by this SDK, @a2a-js/sdk 1.3.0, a2a-go (main 534a60fc) and a reference signer, plus the reference signature on the card edited after signing. These expectations hold however [Bug]: §8.4.1 canonicalization is under-determined - two SDKs produce different bytes for the same card, and neither reproduces the section's own worked example A2A#2122 is settled.
• tests/utils/test_signing_interop.py:
  • each of the four signed cards verifies;
  • each canonicalizes to the shared bytes;
  • the edited card raises InvalidSignaturesError;
  • the frozen JS-signed production card from AgentCard JWS: v2.6.0 canonicalizes the served JSON, the JS SDK canonicalizes the proto projection; signatures do not cross-verify a2a-go#445 verifies and keeps 6410 bytes / c5d5384a…, the anchors @kuangmi-bit set out in Signing/verification canonicalization drops REQUIRED fields at their default value, breaks cross-SDK Agent Card signature verification #1278. Its file hashes are asserted too, so the fixture cannot drift.

One caveat on that last test: the frozen card carries securityRequirements: [{}], a nested empty value whose canonical form is part of #2122. So it pins today's bytes, the ones this SDK and @a2a-js/sdk agree on, as agreed in #1278, and it needs updating if #2122 settles the other way. The docstring says so.

The reading-dependent group (s1-default-valued: description: "", skills: [], tags: [], extensions: []) is deliberately not lifted. Its expectations depend on #2122, and adding them now would lock in one answer before the spec does.

All 11 tests pass on main (fad0482). ruff check, ruff format --check and ty check are clean.

The production card is ours (HORIZON SHIELD publishes it). If you would rather not carry it in the repo, test_frozen_js_signed_production_card and its two files drop out cleanly.

cc @aeoess @kuangmi-bit

[github-actions]

🧪 Code Coverage (vs main)

⬇️ Download Full Report

No coverage changes.

Generated by coverage-comment.yml

[aeoess]

I ran the 11 tests from this PR (at 82e4dde) against both commits in #1287. They pass on the descriptor-only commit 3f90712 and on the Rule 1 commit cdee28e.

The frozen JS-signed card also canonicalizes to 6410 bytes / c5d5384a… on both, matching your anchors. No changes were needed to your tests.

[ogasurfproject-jpg]

Thanks for running them against both commits. That is the result this PR is built for: the five s0 vectors and the frozen card carry every REQUIRED field, so they hold under every reading of §8.4.1 and should stay green whichever way A2A#2122 is settled.

The reading-dependent cases stay out of this PR on purpose. The s2 group added to a2a-tck#246 (a REQUIRED field absent from the served JSON) is where the Rule 1 commit cdee28e departs from the shipping SDKs, and it is scored per reading there rather than pinned here. The numbers are in #1287 (comment).