Add scripts/test-downstream.sh for pre-release downstream smoke checks

[arch-colony]

Why

The v1.7.0 → v1.7.1 fiasco happened because the SDK's own unit tests passed but every downstream framework integration broke under strict mypy. The unit tests use raw dict access and never tripped the dict | Model union; downstream tools.py files do, because they process responses as dicts on behalf of LLMs.

This script catches that class of regression by actually running each downstream's test suite against a wheel built from the local SDK source, in an isolated venv per repo. It would have caught the v1.7.0 regression locally before tagging.

Usage

```bash
./scripts/test-downstream.sh # all known repos
./scripts/test-downstream.sh langchain-colony # one repo
COLONY_DOWNSTREAM_DIR=~/code ./scripts/test-downstream.sh
```

Repos auto-discovered in $COLONY_DOWNSTREAM_DIR, ../<repo>, /tmp/<repo> in that order. Missing repos skipped with clear messages.

Behaviour

• Builds the wheel once via python -m build
• Per repo: creates a fresh venv (uv if available, else python -m venv), installs the downstream's [dev] extras, force-reinstalls the local SDK wheel on top, runs pytest (excluding integration tests), and runs mypy if installed
• pytest failures are release blockers
• mypy errors are advisory — downstream packages have their own type-stub noise from upstream deps (smolagents/langgraph etc.) that lack py.typed
• Reports a pass/fail/skip summary and exits non-zero on any pytest failure

Verified locally

Repo	Result
smolagents-colony	68 tests pass against local SDK 1.7.1
langchain-colony	377 tests pass against local SDK 1.7.1

Also updates RELEASING.md

Added as step 4 in the pre-release checklist, between integration tests (step 3) and version bump (step 5). Renumbered the rest of the steps. Step text explicitly references the v1.7.0 fiasco as the motivating example.

Test plan

• Script is executable and shellcheck-syntax-valid
• Builds wheel successfully
• Creates isolated venvs per repo
• Installs local wheel via force-reinstall
• Runs pytest from each repo's directory
• Reports failures clearly with red/green output (or plain when not a TTY)
• Skips missing repos cleanly
• Verified against smolagents-colony and langchain-colony

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[ColonistOne]

Closing — opened from the wrong account. Reopening as ColonistOne.