fix(release): harden version 7 for stable release - #1025

Request changes on one Major bug. Everything else here is clean, fail-closed hardening — the async JWKS resolver just needs to mirror the sync resolver it diverged from.

Must fix (blocking)

1. Async CachingJwksResolver.__call__ rejects valid signed requests during every refresh window. src/adcp/signing/jwks.py — the new request_signature_jwks_unavailable raise ("cached JWKS is expired and refresh cooldown has not elapsed") lives in the unlocked pre-check and is never re-evaluated inside async with self._lock. _refresh sets self._last_attempt = now before the ~10s awaited fetch, so at each max_age boundary the first request enters refresh and every concurrent verification arriving during the fetch window computes cache_expired and (now - _last_attempt < cooldown) and rejects — legitimate signed inbound requests get request_signature_jwks_unavailable for the duration of the fetch, recurring every 30 minutes under load. The sync CachingJwksResolver handles this correctly: it raises inside self._refresh_lock after re-reading state, so waiters block and observe the fresh cache. Move the async raise inside the lock and re-evaluate after re-reading _last_successful_refresh/_last_attempt, mirroring the sync path. Fails closed (security-reviewer: no replay window opened), but it's a verification-path availability regression this PR introduces on the default async resolver. code-reviewer: Major.

The sync resolver already does exactly this three lines up; the async one didn't get the memo.

Things I checked

• Semver signal is correct. The two breaking hardening changes — PgBackend/lazy PG wiring now requiring a distinct lock_pool, and executor= now requiring timed_sync_get_products_limit= — land under fix(signing)! / fix(decisioning)! with BREAKING CHANGE: footers and a Compatibility section in the body. Both fail at construction, not at runtime. push_sender is additive (feat(a2a)); new signing exports (AtomicReplayStore, ReplayClaimResult, supports_atomic_claim) are additive.
• Sync-executor admission has no permit leak. submit_supervised (time_budget.py) acquires the BoundedSemaphore once and releases exactly once — immediately on executor.submit failure, otherwise via the concurrent.futures done callback through loop.call_soon_threadsafe. Supervised lifecycle tasks are strong-referenced in _SUPERVISED_SYNC_LIFECYCLES / _SUPERVISED_FINALIZATIONS / _SUPERVISED_OPERATIONS so they can't be GC'd mid-flight on the deadline path; the routed-sync ContextVar path acquires its permit inside _run_sync_delegate.
• Idempotency lock contract holds. dispatch.py (262 net lines) read in full: the hold()/put_if_absent() backend surface is concrete-with-NotImplementedError so existing subclasses stay importable, PgBackend._active_connection reuse is keyed on asyncio.current_task() and only reused in the task that set it, and the legacy process-local fallback warns + deprecates rather than silently degrading. lock_pool is pool fails closed at construction.
• Replay store is atomic and never evicts a live nonce. claim() is check-and-set under one RLock (in-memory) / pg_advisory_xact_lock + conditional insert in one txn (PG); the indexed min-heap _purge_expired pops only entries with expiry < now and rejects at capacity instead of evicting; caps validated > 0.
• Bounded fetches close SSRF + decompression-bomb vectors. Redirect handling inside the new client.stream() blocks still rebuilds the IP-pinned transport with follow_redirects=False; _bounded_http counts actual streamed bytes (immune to a missing/lying Content-Length) and rejects non-identity content-encoding before reading.
• Freshness changes are strict tightenings and match AdCP 3.1.8. JWKS max_age=1800, brand.json split, and revocation dropping _slide_next_update on 304 (a 304 authenticates no new signed next_update) all fail closed. verify_from_agent_url passing resolution.key_origins or {} turns a missing brand-json declaration into a fail-closed reject rather than warn-and-skip. ad-tech-protocol-expert: sound.
• Unique disclosure filters are the right shape. Both request schemas declare disclosure_positions/disclosure_persistence as uniqueItems: true, minItems: 1; order-preserving dict.fromkeys dedup via WrapValidator (needed to run after enum coercion and short-circuit None past the rebuilt min_length) is lossless, non-breaking, and keeps the SDK from emitting a schema-invalid array. Generator (scripts/generate_ergonomic_coercion.py) and its output (_ergonomic.py) updated in lockstep; no generated_poc/** hand-edits.
• Webhook accepting request-signing adcp_use is spec-mandated (webhooks reuse the request-signing key; isolation is the tag, not the purpose — expected_tag=WEBHOOK_TAG retained). Closes #1018.
• Test-controller FORBIDDEN closes a real live-account bypass (env_sandbox could previously flip allowed=True for a resolved live account). No new bypass.
• Credential exposure tightened — wire details.caused_by now carries only the exception class name, not str(exc); ctx_metadata fail-close untouched.

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

• brand.json stale-on-error window is dead in the default config. brand_jwks.py:can_serve_stale clamps stale_deadline = min(expires_at + max_stale, fetched_at + DEFAULT_MAX_AGE_SECONDS); with DEFAULT_MAX_AGE_SECONDS=900 and the default no-Cache-Control path where expires_at = fetched_at + 900, the second term collapses the deadline back to expires_at → zero grace. The code comment ("no configuration extends trust past 30 minutes") contradicts the 15-minute value it enforces. Fails closed, so not a block, but the 900+900 split the PR advertises never fires — the ceiling constant should be 1800 (or DEFAULT_MAX_AGE_SECONDS + DEFAULT_MAX_STALE_SECONDS). ad-tech-protocol-expert. Cheap to fix while you're already in the async-JWKS change.
• Shared default replay store caps all counterparties together. agent_resolver._DEFAULT_REPLAY_STORE is one 1M-entry InMemoryReplayStore partitioned only by keyid namespace, so a single valid signer can exhaust the shared global_cap and fail-closed-reject other counterparties' legitimate requests. Net-new protection (old default was replay_store=None), so it's an improvement — but decouple global_cap from per_keyid_cap or expose it for sizing. security-reviewer: Low.
• _execute_locked detached task isolates the handler's contextvars. Running the idempotent handler in asyncio.create_task means ContextVars it mutates are no longer visible after the wrapper returns (previously inline). Worth a one-line note in the PR body / changelog. code-reviewer.

Minor nits (non-blocking)

1. platform_router._run_sync_delegate execution-is-None branch. worker = asyncio.create_task(asyncio.to_thread(...)) then await asyncio.shield(worker) — under caller cancellation the only strong ref is dropped and the task can be GC'd ("Task was destroyed but it is pending"). Not reachable on the deadline-managed get_products path (which always binds an execution), hence a nit.
2. backends.py per-instance dynamic ContextVar. ContextVar(f"adcp_idempotency_connection_{id(self)}", ...) is the documented dynamic-ContextVar anti-pattern; harmless for a long-lived backend, but prefer one module-level var keyed by backend identity.

Fix the async JWKS resolver and I'll approve. security-reviewer clean, ad-tech-protocol-expert sound-with-caveats, code-reviewer one Major.