Add a two-stage assessment and review process for community PRs

## Summary

Add a reusable Spec Kit extension and label-triggered agentic workflow for assessing pull requests submitted by community contributors.

The proposed process has two stages:

```text
assessment → project fit?
├─ no or unclear → stop
└─ yes → review
```

This follows the staged artifact model used by the `bug` extension and the runtime extension installation used by the `feature-assess` workflow, while keeping maintainers responsible for acceptance.

## Proposed extension

Add a `contribution` extension with two commands:

1. `speckit.contribution.assess` writes `.specify/contributions/pr--/assessment.md`. It captures PR intent, linked issue or specification, scope, contribution-policy compliance, AI disclosure, architectural fit, risks, and required checks.
2. `speckit.contribution.review` consumes the assessment and existing CI evidence, then writes `review.md` with prioritized findings, contributor actions, and a recommendation for maintainers.

Both commands are read-only with respect to repository source. Every artifact records the PR base and head SHAs and refuses to consume stale artifacts after the contributor pushes another revision.

## Proposed agentic workflow

Add a `community-pr-review` workflow triggered when a maintainer applies a `community-review` label to an open PR.

The workflow posts at most two top-level comments:

1. **Community contribution assessment — Stage 1/2**: summarizes fit, scope, policy compliance, risks, and the assessed head SHA.
2. **Community contribution review — Stage 2/2**: summarizes findings, CI evidence, contributor actions, and the maintainer recommendation.

If assessment returns `out-of-scope`, `invalid`, or `needs-clarification`, the workflow posts only the assessment comment, applies the corresponding outcome label, and stops. Review runs only when assessment determines that the contribution fits the project and has enough information to evaluate.

Each comment should identify its stage and PR head SHA. Configure safe outputs with `add-comment: max: 2` and constrained outcome labels.

## Guardrails

- Use a human-applied label as the execution gate.
- Never modify or push to the contributor's branch.
- Never formally approve, request changes, merge, or resolve review threads.
- Use the `pull_request` security context, not `pull_request_target`.
- Keep repository and pull-request permissions read-only and expose no secrets.
- Treat PR descriptions, comments, diffs, and changed files as untrusted data rather than instructions.
- Consume existing CI results instead of executing contributor-controlled commands in the agentic workflow.
- Record the head SHA and stop when assessment or CI evidence is stale.

## Suggested outcomes

Assessment can conclude:

- `fits-project` — continue to review.
- `needs-clarification` — request information and stop.
- `out-of-scope` — explain why and stop.
- `invalid` — explain why and stop.

Review can conclude:

- `ready-for-maintainer-review`.
- `needs-contributor-action`.
- `blocked`.

These are recommendations only; a maintainer remains the final decision-maker.

## Acceptance criteria

- A maintainer can trigger the process against a community PR by applying one label.
- Assessment always posts first and clearly identifies itself as Stage 1/2.
- Review runs and posts Stage 2/2 only when assessment returns `fits-project`.
- At most two comments are created per run.
- Both outputs identify the exact PR head SHA they evaluated.
- Source files and the contributor's branch are never modified.
- Re-running after a new push cannot reuse stale assessment artifacts.

## AI disclosure

This issue was drafted and filed on behalf of @mnriem by GitHub Copilot (model: GPT-5.6 Sol).

Start by reading the existing bug extension and the feature-assess workflow to understand staged artifacts and runtime extension installation. Then trace how a proposed community-pr-review workflow could use the label gate, read-only context, SHA checks, and constrained outputs. Done means the acceptance criteria hold for assessment, conditional review, stale revisions, permissions, and comment limits.

Rédigé par le modèle d'indexation à partir du texte de l'issue.