feat: add declarative review loops

[apresmoi]

Summary

• Add an optional declarative .loop system with a built-in serial review workflow: maker → repository check → fresh reviewer → convergence gate → CLI-owned item completion.
• Compile Loop packages into pinned, closed IR; freeze assignments, agent routes, trusted check capabilities, budgets, and launch authority into append-only recoverable Runs.
• Add the complete CLI flow for agent profiles, routes, capability trust, Loop assignment and ASCII inspection, Run creation/execution/recovery, and idempotent completion.
• Show the active Run directly in the read-only Checklist dashboard, including the graph cycle, current node and attempt, candidate-bound maker/check/reviewer evidence, transitions, timestamps, and corrupt or interrupted states.
• Document setup, operation, recovery, authority boundaries, and the unchanged direct burnlist burn path in the skill, README, CLI reference, and website.

Validation

• npm run verify
• npm run verify:package — exact 202-file package manifest; 1,942,426 bytes unpacked
• npm run test:global-install on Node 22.14.0
• BURNLIST_TEST_NODE=<node-18-binary> npm run test:global-install on Node 18.20.8
• npm run build:dashboard
• git diff origin/main...HEAD --check
• Two independent final reviews cover runtime/state/completion correctness and CLI/UI/product scope; both approve with no open P0–P2 findings.

Structure

• Read order: loops/review/review.loop (shipped workflow) → src/loops/dsl/ (compiler and frozen IR) → src/loops/run/ (journal, state machine, authority, runner) → src/loops/completion/ (CLI-owned completion transaction) → src/cli/loop-cli.mjs (consumer flow) → Checklist dashboard projection and rendering → skill and website guidance.
• Canonical Burnlist mutations remain CLI-owned. The dashboard only reads bounded projections and refetches canonical state after invalidation events.
• Runs use an append-only hash-chained journal, recoverable publication, candidate-bound evidence, one writer, and an independently launched read-authority reviewer.
• Root runtime dependencies remain zero, and every new authored file is at most 400 lines.

Notes

• This is deliberately the minimal Stage 1 slice: one built-in serial workflow, one Codex CLI adapter, one writer, and foreground execution.
• Reviewer write denial is supervised rather than an OS sandbox. Parallel branches, Docker isolation, metrics gates, custom adapters, forecasting, automatic worktrees, and background execution remain deferred.
• Items without a Loop assignment continue to use the existing direct completion flow unchanged.