feat: sandcastle refinement loop with critic-based convergence

.sandcastle/critic-prompt.md

@@ -0,0 +1,56 @@
+# Critic Agent
+
+Analyze the implementation on branch `{{BRANCH}}` and produce structured findings.
+
+## Task
+
+Run `git diff main...{{BRANCH}}` to see all changes. Examine the diff carefully. For each issue found, produce a structured finding.
+
+Read `CONTRIBUTING.md` for the project's coding standards.
+
+## Output Format
+
+Output your findings as JSON wrapped in nonce-tagged delimiters. Use EXACTLY this tag format:
+
+```text
+<findings-{{NONCE}}>[...]</findings-{{NONCE}}>
+```
+
+Each finding must have this structure:
+
+```json
+{
+  "file": "path/to/file.ts",
+  "line": 42,
+  "title": "short description of the issue",
+  "severity": "CRITICAL|HIGH|MEDIUM|LOW",
+  "category": "security|logic|performance|architecture|style",
+  "confidence": "HIGH|MEDIUM|LOW",
+  "description": "detailed explanation of why this is a problem",
+  "suggestion": "how to fix it"
+}
+```
+
+If no issues are found, output:
+
+```text
+<findings-{{NONCE}}>[]</findings-{{NONCE}}>
+```
+
+## Rules
+
+- Do NOT modify any files. Do NOT commit. Do NOT push.
+- Only report issues in the CHANGED code (not pre-existing issues).
+- Use HIGH confidence only when you've verified the issue by reading the relevant code.
+- Use MEDIUM confidence for pattern-based detection.
+- Use LOW confidence for style preferences or uncertain issues.
+- Focus on: logic errors, missing edge cases, security issues, type safety violations, test gaps.
+- Do NOT report formatting issues (prettier handles those).
+
+## Completion
+
+After outputting the findings, output:
+
+```text
+<promise>COMPLETE</promise>
+```

.sandcastle/implement-prompt.md

@@ -4,11 +4,11 @@ Implement issue **#{{TASK_ID}}** ("{{ISSUE_TITLE}}") on branch `{{BRANCH}}`.
 
 ## Issue Details
 
-!`gh issue view {{TASK_ID}} --json body,title,labels,comments`
+{{ISSUE_BODY}}
 
-## Recent Commits
+## Review Findings
 
-!`git log -n 10 --format="%h %s" --date=short`
+{{FINDINGS}}
 
 ## Exploration
 
@@ -22,24 +22,26 @@ Read `AGENTS.md` and `CONTRIBUTING.md` for project conventions.
 
 ## Implementation
 
-1. Implement the fix/feature. Follow existing patterns:
+1. If review findings are provided above, cross-validate each one against the code. Fix findings you agree with. Ignore findings that are incorrect or not applicable.
+
+2. If no findings are provided, implement the issue from scratch following existing patterns:
    - Strict TypeScript, JSDoc on public APIs
    - Co-located tests in `*.test.ts` files
    - Zod for runtime validation
 
-2. Before every commit, run the full validation suite:
+3. Before every commit, run the full validation suite:
 
    ```bash
    npm run type-check && npm run test && npm run test:node && npm run test:edge && npm run prettier-check && npm run lint && npm run build && npm run check-build && npm run build:v2 && npm run check-build:v2
    ```
 
-3. Commit with conventional commits:
+4. Commit with conventional commits:
    - `fix: <description>` — bug fix
    - `feat: <description>` — new feature
    - `refactor: <description>` — restructuring
    - `chore: <description>` — tooling/config
 
-4. Push the branch:
+5. Push the branch:
 
    ```bash
    git push -u origin {{BRANCH}}
@@ -51,6 +53,7 @@ Read `AGENTS.md` and `CONTRIBUTING.md` for project conventions.
 - Tests must pass before pushing. Zero type errors, zero test failures.
 - Do not modify unrelated files.
 - Do not bump version numbers.
+- Push BEFORE signaling completion.
 
 ## Completion