Add unified AgentProvider trait, CLI adapters, and e2e pipeline tests

[csells]

Summary

• Unified AgentProvider trait — every backend (HTTP API or CLI subprocess) implements run_to_completion(), so Claude Code, Codex CLI, and Gemini CLI plug into Attractor pipelines as first-class backends
• CLI adapters for claude-code, codex-cli, and gemini-cli with JSONL event parsing, proper UTF-8 truncation (floor_char_boundary), and subprocess lifecycle management
• Attractor spec compliance — skipped status, failure_reason, retry presets, node timeouts, auto_status, manifest.json, real shell execution in ToolHandler, executor-based parallel branches, LLM-backed fan-in consolidation, goal-gate enforcement on all graph nodes (including unvisited), 4-level stylesheet specificity
• 8 full-stack e2e pipeline tests — DOT file → forge-cli → real CLI agent → JSONL events → disk artifacts → CXDB persistence, all running in isolated git sandboxes in /tmp
• Test hygiene — #[ignore] is the only acceptable gating mechanism; removed all banned double-gating patterns; tests fail hard when prerequisites are missing

What changed

forge-llm

• AgentProvider trait, AgentRunOptions/AgentRunResult types
• CLI adapters: claude_code.rs, codex.rs, gemini.rs with JSONL parsing
• Fix: truncate_json/truncate panic on multi-byte UTF-8 char boundaries (used floor_char_boundary())

forge-agent

• HttpApiAgentProvider extracting Session's tool loop into the AgentProvider interface
• pub(crate) visibility on session utils for reuse

forge-attractor

• AgentProviderSubmitter bridging AgentProvider → AgentSubmitter for pipeline integration
• Condition evaluation fixes: missing keys → empty string, bare key truthiness, context prefix resolution
• Goal-gate enforcement checks ALL graph nodes (including unvisited)
• Stylesheet specificity: universal (0) < shape (1) < class (2) < node_id (3)
• Lint: exactly one terminal node (was "at least one")
• manifest.json now includes cxdb_context_id
• spec_complete test suite + 5 agent_provider_integration tests

forge-cli

• --backend claude-code|codex-cli|gemini-cli modes
• 8 e2e pipeline tests covering linear (×3 backends), HITL auto-approve, parallel fan-out, CXDB persistence, cross-provider parity, resume-from-checkpoint
• Each test uses init_sandbox() for isolated git repos — no project files touched

Specs & docs

• New spec/06-unified-agent-provider-spec.md
• Updated spec/03-attractor-spec.md DoD (all checkboxes complete except optional HTTP server mode)
• Updated AGENTS.md and README.md with test commands and three-tier test taxonomy

Test coverage

Tier	Count	What
Default (cargo test)	647	Unit + integration, no external deps
Infrastructure (--ignored)	51	CXDB server, CLI agents (OAuth, no API keys)
API keys (--ignored)	25	OpenAI/Anthropic live calls (costs money)

Test plan

• cargo test — 647 passed, 0 failed
• cargo test -p forge-cli --test e2e_pipeline -- --ignored — 8 passed (requires CLI agents + CXDB)
• cargo test -p forge-llm --test cli_agent_e2e -- --ignored — 9 passed (requires CLI agents)
• Verify against upstream CI if configured

🤖 Generated with Claude Code