fix(sdjwt): honor expected aud/nonce on every KB hop

[SashaMIT]

Summary

verify() only called verify_expected_claims under if typ in TYP_TERMINAL, so on an intermediate hop (typ kb+sd-jwt+kb) a caller that passes expected_aud or expected_nonce got no check at all. That is why test_verify_rejects_aud_mismatch and test_verify_rejects_nonce_mismatch in kb_sd_jwt_intermediate_tests.py fail on main.

Move the check out of the terminal guard so it runs on every hop. Matches the verify() docstring (expected match is a general check) and create() (which already requires aud and nonce on every hop).

Scope note (split)

Per discussion with @chopmob-cloud (and the contract / deployment / CI sequencing from @giorgioroth / AlgoVoi / @arjun2075), this PR is only the supplied-expectation fix. The stronger behaviour change — a terminal hop that carries aud/nonce must be bound even when the caller passed no expectation — lands as a follow-up with its own compatibility note and tests, so this change can green the two intermediate-hop tests and unblock #325 without settling a protocol question by default.

Test plan

• Intermediate-hop mismatch tests now pass: test_verify_rejects_aud_mismatch, test_verify_rejects_nonce_mismatch
• Existing terminal-hop tests that already pass expected_aud/expected_nonce unchanged
• No new fail-closed-on-unbound behaviour in this PR

CI

cla/google is green. Lint Code Base is the repo-wide Biome scan from #306 and also fails on main; unrelated to this Python-only diff.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[SashaMIT]

@googlebot I signed it! (commits re-authored to the signing email)

[SashaMIT]

@googlebot rescan

[SashaMIT]

@googlebot rescan

[SashaMIT]

@googlebot rescan

[SashaMIT]

CI note: the Lint Code Base failure is the repo-wide Biome run (75 errors in code/web-client TS files, also failing on main at run 25120930358). This PR touches Python SDK + tests only, so none of the reported diagnostics come from this change. The remaining blocker here is cla/google.

[chopmob-cloud]

Reviewed this independently against current main.

The intermediate-hop gap is real and reproduces. verify() only calls verify_expected_claims under if typ in TYP_TERMINAL, so on an intermediate hop (typ kb+sd-jwt+kb) a caller that passes expected_aud or expected_nonce gets no check at all. That is why test_verify_rejects_aud_mismatch and test_verify_rejects_nonce_mismatch in kb_sd_jwt_intermediate_tests.py fail on main. Moving verify_expected_claims out of the terminal guard so it runs on every hop is the correct minimal fix, and it matches the documented contract: the verify() docstring states the expected match as a general check, and create() already requires aud and nonce on every hop, so both claims are always issued and should always be checkable.

The second change (a terminal hop that carries aud or nonce but is verified with no expected value now raises) is a stronger and separate behaviour change. It converts a previously silent pass into an error for existing callers, which is why seven in-tree tests had to start binding. It is defensible on replay grounds, but it decides a protocol-semantics point rather than only turning the two red tests green, so it reads as a maintainer call. Splitting it into its own follow-up would let the two-test fix land immediately as a clean green baseline (which the CI-only work in #325 depends on) without waiting on the semantics discussion.

On CI: the two red checks are cla/google (your signing) and Lint Code Base, which is the repo-wide Biome scan from #306 and also fails on main. That Biome result is unrelated to this diff and clears once the linter-workflow change reaches main and this branch rebases onto it.

[SashaMIT]

Thanks for the careful review. Happy to split: I can move the fail-closed-on-terminal change into a follow-up so the two-test fix lands clean ahead of #325. I'll wait for a maintainer to confirm that's the preferred shape before reshuffling, since the replay exposure only fully closes with the second half.

[chopmob-cloud]

@SashaMIT that shape works, and splitting sequences the hardening rather than dropping it. Your caveat is the right one: the replay exposure only fully closes with the second half, so the mandatory-presence change (a terminal hop carrying aud or nonce must be bound even when the caller passed no expectation) should still land, as its own PR with a compatibility note and its own tests.

The split reads the same from every side that has looked at it. On the contract side, @giorgioroth set out on #310 why these are two properties: the supplied-expectation fix restores what the verify() docstring already documents (expected values are matched when provided, and the next docstring line conditions on hop type only for cnf), whereas mandatory presence moves a rule from issuer to verifier. On the production side, we at AlgoVoi run live payment verification, and mandatory presence is the half that would reject tokens deployed verifiers and issuers accept today, including ones this SDK did not emit, so it needs its own compatibility window. And the sequencing is @arjun2075's: land the behavioural fix first, then run the CI-only #325 against a green baseline, which also keeps a red test from settling a protocol question by default. Contract reading, deployment view and CI ordering all point the same way.

So the shape for a maintainer to confirm: keep #313 to the supplied-expectation fix that greens the two intermediate-hop tests and unblocks @arjun2075's #325, then mandatory presence as a follow-up. Once you reshuffle I am happy to review both.

[SashaMIT]

Reshuffled as agreed.

• fix(sdjwt): honor expected aud/nonce on every KB hop #313 is now only the supplied-expectation fix (honor expected_aud/expected_nonce on every hop). Diff is the one-line move of verify_expected_claims out of the terminal guard, plus a short comment. Title/body updated.
• Follow-up opened as fix(sdjwt): require expected aud/nonce when a terminal KB hop carries them #326 for mandatory presence (terminal hop carrying aud/nonce must be bound), with the compatibility note and its own tests. Based on this branch so it can merge after fix(sdjwt): honor expected aud/nonce on every KB hop #313.

Happy for a re-review of both whenever you have a moment.

[chopmob-cloud]

Re-reviewed both against current main (e1ea56d), built and run in a clean venv.

#313 — supplied-expectation fix. The reshuffle is exactly the shape agreed: the diff is the one-line move of verify_expected_claims out of the if typ in TYP_TERMINAL guard plus the explanatory comment (the only other change is two cspell allowlist entries). Baseline main reproduces the gap, test_verify_rejects_aud_mismatch and test_verify_rejects_nonce_mismatch both fail "DID NOT RAISE". On this branch both go green and nothing regresses: the two kb_sd_jwt test files are 23/23 and the whole Python SDK suite is 188 passed. verify_expected_claims compares aud/nonce with exact equality and is a no-op when the caller passes no expectation, so this only tightens intermediate hops for callers who already bind, with no new false rejects. LGTM to land as the clean baseline #325 depends on.

#326 — mandatory presence. This is the correct home for the stronger change. The terminal-only guard is right, it runs after the supplied-expectation check so a mismatched value still gets the specific error, and the two new negative tests (test_verify_rejects_terminal_aud_left_unbound, test_verify_rejects_terminal_nonce_left_unbound) plus the in-tree call sites that now bind bring the full suite to 190 passed. The compatibility note names the real exposure precisely: an unbound terminal verifier turned any captured presentation into a replayable bearer credential. The RFC 9901 §7.3 citation checks out, "Verification by the Verifier", step 5(f) requires the Verifier validate nonce and aud for replay detection, exactly the property this enforces. Approve the approach; merge after #313 since it is based on that branch.

One scope note for a possible later follow-up (not a change request on #326): the mandatory check fires on claim presence ('aud' in payload), so a terminal token minted with no aud/nonce at all still verifies. AP2-profile create() always emits both, so conformant tokens are covered; enforcing that terminal hops MUST carry them is an issuer-side invariant that could be a separate verify-path check if the group wants defence against non-conformant issuers.

[SashaMIT]

Thanks for the re-review @chopmob-cloud — glad the reshuffle matches the agreed shape and the suite stays green. Holding #313 to the supplied-expectation fix only; happy to wait on a Google code-owner pass for merge ahead of #325.