…t window

Two issues surfaced once the full barbacane-test suite ran in CI:

1. The Integration Tests job installed the toolchain without the
   wasm32-unknown-unknown target, so barbacane-test's build.rs could not
   compile the streaming-echo/body-echo fixture WASM plugins (exit 101).
   The streaming/body suites then had no real coverage (they panic on the
   missing .wasm, or lean on a stale cached artifact). Add the target so
   the fixtures build, matching the build and security-suite jobs.

2. ai_gateway::cel_selected_strict_profile_blocks_prompt failed twice with
   StartupFailed("gateway did not become ready in time"). It boots the
   same spec as a sibling test that passes, so it is not a spec/code bug:
   with --test-threads=2 two CEL-heavy gateways cold-boot at once on a
   shared CI runner and the loser exceeds the 60s health window. Widen the
   TestGateway readiness timeout to 120s; a real boot hang still fails.

The shipped-fragment test's /v1/models step exercised the real
api.openai.com / api.anthropic.com hosts (the fragment omits base_url for
those providers). Locally they fast-fail (401) so the aggregator returns
the expected partial 200; in CI they hang, and the accumulated wall-clock
trips the ai-proxy plugin's epoch deadline (added in the WASM sandbox
hardening), so the dispatcher traps and the gateway returns 502.

Redirect all three providers at the wiremock by injecting
base_url: env://<PROVIDER>_BASE_URL into the copied fragment's
OpenAI/Anthropic routes, and mock GET /v1/models -> 500 so the partial
path fires deterministically and fast. No outbound network, no epoch
pressure. The partial-aggregation contract is also covered by the
dedicated hermetic models tests.

cargo test stops at the first test binary that fails, which hides
failures in every binary scheduled after it. Broadening CI surfaced one
new failing binary per run (ai_gateway -> ai_proxy -> auth) instead of
all at once. Run with --no-fail-fast so a single CI run reports the full
set of failures across all integration binaries.

The WASM epoch deadline (max_execution_ms, default 100ms) bounds plugin
CPU time, but the epoch clock keeps advancing while a plugin is blocked
in a host function doing network I/O (broker publish, HTTP call/stream).
A slow or unavailable upstream therefore made the plugin resume past its
deadline and trap, surfacing as a 500 instead of a clean gateway error.

This was observable as an inconsistency: an unavailable NATS broker
returned 502 (it fails fast), while an unavailable Kafka broker returned
500 (rskafka retries the connect for ~5s, blowing the 100ms budget, so
the dispatch trapped).

Store max_execution_ms in PluginState and, in host_kafka_publish,
host_nats_publish, host_http_call and host_http_stream, refresh the
store's epoch deadline once the blocking call returns. Time spent waiting
on native I/O no longer counts against the plugin's CPU-execution budget,
so a slow/unreachable upstream yields a clean error response rather than a
trap. The CPU guard on actual WASM execution is unchanged.

Broadening CI to run every barbacane-test integration binary surfaced 21
tests that had rotted while excluded. All are test-side issues (the
production behavior they now exercise is correct, and several confirm the
security hardening works):

- auth JWT (5): tests used a fake signature relying on
  skip_signature_validation, which the production WASM correctly ignores
  (it is honored only under the plugin's own cfg(test)). Sign tokens with
  a real ES256 key (new p256 dev-dep) and embed the matching public JWK in
  a generated spec.
- auth secrets (1): file:// secrets now require BARBACANE_SECRETS_DIR
  (fail-closed). Inject it via the new TestGateway::from_spec_with_env,
  which sets per-instance child env vars instead of racing on process
  globals.
- mcp (1): a non-initialize method now needs a session; initialize first.
- proxy (3): tests hit the real httpbin.org; point them at a local
  wiremock so they are hermetic and fast.
- plugins/redirect (5): reqwest follows redirects by default, turning the
  3xx into a followed 404/200; use a non-redirect client.
- plugins/ip-restriction (2): X-Forwarded-For is only trusted from
  declared trusted_proxies (anti-spoofing); declare loopback so the
  XFF-based cases are exercised.
- plugins/kafka (1): now passes via the host-I/O epoch fix.
- streaming (3): the streaming-echo fixture declared capabilities with the
  wrong key syntax (http_stream = true) instead of host_functions =
  ["http_call"], and its plugin.toml was not beside the wasm where the
  compiler reads it. Fix the manifest and have build.rs copy plugin.toml
  next to the built fixture wasm.