Merge origin/iris-implementation into feedback branch
Integrates origin's agent-suggestion + contribution feature with the local
feedback / verification / regression work. Resolves conflicts in
extraction.ts, orchestrator.ts, and sessions.ts by keeping BOTH:
- extraction.ts: loadable page agent + build-time source-fidelity verify and
one-shot self-correction (ours), and collection of `suggested_agent`
suggestions returning { fragments, suggestions } (theirs).
- orchestrator.ts: iterative feedback refinement + feedback-driven page-agent
training (ours), plus githubToken plumbing and a non-blocking
runContribution(suggestions) after the run completes (theirs).
- sessions.ts: origin's refactor (issue-based contributions, upload-name
mirroring, no PR-on-close) with regression-fixture capture re-added on close
and githubToken passed through to runPipeline.
Follow-on fixes the merge required:
- agents/page.md now emits `suggested_agent` so the loadable page agent matches
the suggestions behavior.
- feedback.ts defines AgentUpdateContribution locally; origin removed
src/github/contributions.ts, so the old import is gone.
Note: origin replaced PR-on-close with GitHub issues filed during the run, so
feedback-driven page.md proposals are now recorded to agent-updates.md (gated by
regression fixtures) rather than auto-opened as PRs.
Verified: no conflict markers; all touched files pass
`node --experimental-strip-types --check`.