feat(testing): make SellerA2AClient event drain configurable

[bokelley]

Summary

Mirrors the contribution from #763 onto a branch in the base repository so the repository's required CodeQL check can run.

Closes #698.
Supersedes #763.

What changed

• Adds keyword-only max_events: int = 32 to SellerA2AClient.invoke().
• Replaces the hardcoded 32-event A2A drain cap with the caller-provided cap.
• Updates the exhaustion error to report both timeout_seconds and max_events.
• Adds regression coverage for default-cap exhaustion and success with a larger cap.

Original patch authored by @sangilish in #763.

Testing

• uv run --extra dev python -m pytest tests/test_seller_a2a_client.py -q
• uv run --extra dev ruff check src/adcp/testing/harness.py tests/test_seller_a2a_client.py
• uv run --extra dev mypy src/adcp/testing tests/test_seller_a2a_client.py

[aao-ipr-bot]

LGTM. Additive kwarg with a default that preserves the prior range(32) — non-breaking on the public testing surface, conventional-commit prefix matches.

Things I checked

• src/adcp/testing/harness.py:296 — default max_events=32 matches the prior hardcoded bound. Loop at L356 and error message at L370-L374 are consistent.
• SellerA2AClient is re-exported from adcp.testing (src/adcp/testing/__init__.py:26,48) — public surface. Change is additive, no ! / BREAKING CHANGE: needed.
• Regression coverage covers both arms: default-cap exhaustion at tests/test_seller_a2a_client.py:170-177, larger cap success at L181-L191.
• Incidental supported_billing=("operator",) → supported_billing=["operator"] in the shared fixture aligns with the declared list[str] type at src/adcp/decisioning/platform.py:198. Drift cleanup, not a hidden break.

Follow-ups (non-blocking — file as issues)

• The new error message format "within {timeout_seconds}s x {max_events} events" (src/adcp/testing/harness.py:372) uses a bare x as a separator. Reads fine to humans; a log scanner could mis-tokenize it next to a float. Something like "within {timeout_seconds}s per event, max {max_events} events" would be unambiguous. Purely cosmetic.

Minor nits (non-blocking)

1. Substring assertion is implicitly format-coupled. tests/test_seller_a2a_client.py:177 asserts "within 0.01s x 32 events". Stable today because CPython's shortest-round-trip repr of 0.01 is "0.01", but if anyone ever adds a format spec to the f-string the test silently passes for 0.01 and breaks for other values. Splitting into separate "x 32 events" and "0.01" substring checks would be more robust.

Mirroring #763 onto the base branch for the CodeQL gate is the right move — happy path is unchanged, regression coverage is in place. Safe to merge.