Add Gherkin behaviour tests with dummy runtime

[ifireball]

Summary

• Add defaults.runtime org config and --runtime install flag with shared runtime.ResolveFromConfig() selection in fullsend run.
• Implement a dummy runtime that executes scripted sandbox operations and emits behaviour-results.json for deterministic assertions.
• Introduce godog behaviour tests under e2e/behaviour/ with pluggable GitHub/GitHub Actions drivers, triage scenarios, CI job, and ADR/docs.

Test plan

• go test ./...
• go test -tags behaviour -c ./e2e/behaviour/...
• go test -tags e2e -c ./e2e/admin/...
• make behaviour-test against halfsend org pool with GITHUB_TOKEN and --runtime dummy orgs
• CI behaviour job after adding E2E_BEHAVIOUR_GITHUB_TOKEN secret

Made with Cursor

[github-actions]

Site preview

Preview: https://bddbbd25-site.fullsend-ai.workers.dev

Commit: 6a5393475a282d474bd1a161c7a8c22ad952747e

[ifireball]

Closing to reopen from an upstream branch so CI can access repository secrets.

[fullsend-ai-review]

Review skipped — this PR is already closed.

The /fs-review command only reviews open pull requests.

_{Posted by fullsend post-review check}

[fullsend-ai-retro]

Retro: PR #1981 — Wasted review agent run on short-lived draft PR

What happened: PR #1981 was opened as a draft from a fork at 10:40 UTC by a human contributor (ifireball) using Cursor. A review agent was dispatched 15 seconds later (run 27090168627). The author closed the PR at 10:47 UTC (7 minutes later) to reopen from an upstream branch so CI could access secrets. The review agent continued running until ~10:51 UTC, completing a full request-changes analysis with 8 findings — none of which were ever posted to the PR. PR #1982 was then opened at 10:49 UTC with identical content, triggering a second review run.

Waste: ~11 minutes of review agent compute (tokens + GHA minutes) were spent on a PR that was closed before the review finished. The same content was then re-reviewed on PR #1982.

Existing coverage: All improvement opportunities identified are already tracked by open issues:

• #1715 — Skip review dispatch when PR is in draft state
• #1870 — Skip review/fix dispatch when PR action is 'closed'
• #1439 — Skip review dispatch when PR is already closed/merged
• #1357 — Enable cancel-in-progress for review dispatches on the same PR

Recently closed issues #1566 and #887 suggest active work in this area.

No new proposals filed — existing issues comprehensively cover the improvements needed. This PR serves as a concrete real-world example validating the priority of those issues.