Unblock the nightly Official TCK run (SLIMRPC re-vendor) and fix two intermittent server tests

[tomtom215]

What this changes

Two independent commits.

f51c776 — SLIMRPC spec re-vendor; the binding does not change. Upstream a2aproject/experimental-cpb-slimrpc merged its collaborative-task work to main on 2026-09-29 as 1328426 (its PR #5). That added three documents to main and changed both vendored ones, so check_slimrpc_spec.sh failed the scheduled Official TCK run (36666242899). That step comes before "Start the SUT", so the official suite and the baseline gate were skipped: main has had no full official conformance result since the upstream merge.

• All five upstream documents are now vendored.
• Everything the merge added to the two implemented documents is A2A 1.1: SendLiveMessage, its multicast form (new §8), and the slimrpc-context-map key used only on that call. The newest a2aproject/A2A tag is v1.0.1, so the binding stays on the A2A 1.0 surface. The spec itself requires implementations targeting an older A2A version to use that version's methods. The eleven 1.0 rows are unchanged.
• The three new documents are byte-identical to branch tip 36b03a7, which was already triaged as not followed. Only "never reached main" stopped being true.
• The reasoning is in spec/slimrpc_v1/README.md. The binding README, the book chapter and an e2e doc comment now say "A2A 1.0 surface". The three triage entries in KNOWN_BRANCH_SPECS are removed, since a file on main can never match one.
• Corrected along the way: spec/slimrpc_v1/README.md said the StreamRequest vocabulary is absent from both vendored specs. Both now name StreamRequest, but only as the request type of SendLiveMessage.

baa8ec8 — two intermittent tests in a2a-protocol-server, each failing because of the test. These are the failures on #145 (Windows) and #146 (macOS). Neither PR's diff could have caused them.

• the_rps_unit_is_multiplied_by_the_window: the 121st request was admitted. The limiter's windows are fixed and aligned to the wall clock, so a boundary inside the loop resets the count, as it should. This test and the 1-second-window tenant test now use refused_within_one_window, which discards an attempt that straddles a boundary and retries it.
• concurrent_claims_of_one_key_produce_exactly_one_winner: 0 winners, which only a pre-existing claim can produce. The database path was a2a-idem-{pid}, and it was removed while the store still held the file open. Windows refuses that, and the error was discarded, so a later test process with the same PID found the key already claimed. This test and a_migrated_store_can_take_an_artifact_append now use a per-run directory and drop the store first.

How it was verified

• ./scripts/check_slimrpc_spec.sh: exit 1 before; after, "5 file(s) on upstream main, all vendored and matching; 4 branch-only spec(s), all triaged".
• check_method_denominator.py --slimrpc-spec spec/slimrpc_v1/slimrpc.md: 2 sources agree on the same 11 methods.
• Both test failures were reproduced first, with scratch tests that are not committed:
  • A database left behind with the key already claimed gives claimed=0 replays=16, exactly the CI signature.
  • A run forced across a 1-second boundary admits the over-limit request under the old logic. The helper discards that attempt and gets a refusal on the retry.
• cargo test -p a2a-protocol-server --lib passes with default features (1,021 tests), --features sqlite (1,183) and --all-features (1,374 passed, 2 ignored).
• cargo fmt --all -- --check, and rustfmt --edition 2024 --check on the binding's tests/e2e.rs.
• cargo clippy -p a2a-protocol-server --all-features --all-targets -D warnings.
• The static-job scripts from ci.yml (7 shell, 18 Python) and check_book_code.sh all pass.
• scripts/check_dco.sh: both commits carry a matching sign-off.

Checklist

• Every commit is signed off (git commit -s) by a human git author
• SPDX header on every new file: the three new files in spec/slimrpc_v1/ are verbatim upstream copies and carry none, like the two already vendored
• No file exceeds 500 lines (check_file_lengths.sh passes)
• cargo fmt --all passes
• cargo clippy --workspace --all-targets -- -D warnings: run for a2a-protocol-server with all features and all targets, not the whole workspace; CI runs the rest
• cargo test --workspace: run for a2a-protocol-server --lib under three feature sets, not the whole workspace
• cargo doc --workspace --no-deps: not run; no doc comments on public items changed
• New public types/functions have doc comments (none added)
• New code has tests (the changes are tests)
• cargo mutants: not run; only test code and documents changed, so there is no production line to mutate
• CHANGELOG.md updated ([Unreleased] → ### Internal)
• ADR not needed

🤖 Generated with Claude Code

https://claude.ai/code/session_01CmrtmTtr47ip388nrSaduU

---

Generated by Claude Code