Run the live smoke harness as a release gate, not by memory

[yuanhao]

Second of the trustworthiness items. #160 did the first.

Why

examples/release_smoke.rs found a real bug on its first run — tools with zero parameters were uncallable on Anthropic, in shipped releases. examples/long_horizon.rs then found the LlmCompaction fallback behaviour that became #150.

Both currently run when someone remembers. That is the same class of problem the price audit had before it was written into the release checklist: a check that depends on memory is a check you do not have.

Why not per-PR CI

• Needs provider keys as secrets
• Costs real money per run
• Live calls are flaky — two long_horizon runs hung mid-session during Compaction's MIN_HEADROOM_RATIO floor produces a post-compaction context cliff #150 work

Per-PR is the wrong granularity. Per-release is right: a few cents, a few minutes, once per version.

Proposal

A manual-dispatch GitHub Action (workflow_dispatch, optionally also on push to release/**) that runs:

cargo run --example release_smoke                      # ANTHROPIC_API_KEY
SMOKE_MODEL=deepseek cargo run --example release_smoke # DEEPSEEK_API_KEY

Both already exit non-zero on failure, so they gate correctly as-is.

DeepSeek matters as a second target: it exercises the OpenAI-compat implementation, which is a separate SSE parser and tool-call accumulator from Anthropic's. The zero-arg bug lived in exactly that layer.

Acceptance

• Workflow exists and can be dispatched against a release branch.
• Keys stored as repository secrets; the job is skipped rather than failed when absent, so forks are not broken.
• memory/release.md's checklist references it as a numbered step alongside the price audit.
• Timeout set — a hung run should fail the gate, not hang the queue.

Not in scope

long_horizon is not a gate. It is a diagnostic: slower, and its interesting output is a judgement call rather than a pass/fail. Keep it manual.