ci: run the live smoke harness as a release gate

[yuanhao]

Closes #161.

Why

release_smoke found a real bug on its first ever run — tools with zero parameters were uncallable on Anthropic, in shipped releases. Since then it has run when someone remembered.

That is the same failure mode the price audit had before it went into the release checklist: a check that depends on memory is a check you do not have.

The gate

Manual dispatch, plus automatic on release/** push. Both providers run independently — they are separate SSE parsers and tool-call accumulators, and the zero-argument bug lived in exactly that layer.

decision	why
Not per-PR CI	needs keys, costs money per run, and live calls are flaky — two long_horizon runs hung mid-session during #150
15-minute timeout	the harness has none, and a hung provider call should fail the gate rather than block the queue
Missing key skips, not fails	a fork without secrets shouldn't see a red X for a gate it can't run
long_horizon excluded	it's a diagnostic — slower, and its output wants reading rather than a pass/fail

The harness already exits non-zero on any failed check, so it gates as-is with no changes.

Verified

• Workflow YAML parses; triggers, matrix, and timeout confirmed
• SMOKE_MODEL="" (the Anthropic row passes an empty string, not unset) falls through to the Sonnet 5 default — checked, not assumed

Documented in CONTRIBUTING.md and added to the release checklist as step 3b, next to the price audit.

Note on a sibling issue

While writing the CONTRIBUTING section I found that #162's premise was wrong — I'd claimed the SSE parsing layer had no coverage, but tests/anthropic_stream_test.rs has 23 wiremock-based tests, google_stream_test.rs 8, openai_compat_stream_test.rs 3. I've corrected that issue and re-scoped it to the actual gap: empty tool arguments are untested while malformed ones are pinned, and openai_compat has thin coverage for something backing 15+ providers.

CI-and-docs only — no library change.

🤖 Generated with Claude Code