[DSP-7] Patterns & Telemetry: orchestrator/synthesizer, task chains, review-fix termination, anonymous telemetry from devswarm

[justrach]

Child of #53. Source: justrach/devswarm.

This child is the glue — the patterns DSP-1 needs and the observability surface that makes swarm runs debuggable.

Patterns to port

1. Orchestrator → Workers → Synthesizer

• Orchestrator (Opus-tier): decomposes a swarm task into N sub-tasks with role assignment per worker.
• Workers (Sonnet-tier): execute in parallel in isolated worktrees (depends on [P2-9] ExecutionEnvironment trait + Local impl + SDK exec.registerEnvironment (route shell tools through trait) #43).
• Synthesizer (Sonnet-tier): consumes worker outputs, produces one cohesive response.

2. Task chain presets

Four chains, configurable, runnable via run_task:

• finder_fixer — locate → fix
• reviewer_fixer — verify → correct
• explore_report — investigate → structured summary
• architect_build — design → implement

3. Review-fix loop termination

• Reviewer returns NO_ISSUES_FOUND → terminate.
• Iteration count hits max_iter (default 5) → terminate with BAILED.

4. Role + mode routing config

• .codegraff/swarm.toml with [roles.<role>] model = ... and [modes.<mode>] strategy = ....
• Per-call override via SDK / TUI command flags.

Telemetry (port from devswarm, opt-out by default? decide here)

DevSwarm collects (anonymous):

• Agent roles invoked
• Model names used
• Token counts (in/out per worker)
• Execution time, estimated cost
• Worker parallelism metrics

DevSwarm explicitly never collects:

• Code, diffs, file contents
• Prompts, task descriptions, outputs
• Repo names, file paths, branches
• PII

Opt-out: enabled = false in .devswarm/config.toml or DEVSWARM_TELEMETRY=false.

Open question for this issue: mirror this telemetry shape in codegraff, or fold it into the existing forge_tracker crate? Decide before implementation lands.

Acceptance criteria

• Task chain presets land as configurable structures, not hard-coded in code paths.
• Review-fix loop has both termination conditions tested.
• Telemetry decision recorded in this issue (forge_tracker vs new) and implemented.
• Adversarial benchmark verification — when the orchestrator reports per-worker metrics, they're independently verifiable, not LLM-self-reported (devswarm #330 was the lesson; codegraff's own Interactive multi-turn test harness for agents #20 multi-turn harness is adjacent).

Related codegraff issues

• feat(observability): trajectory_events SQLite table for subagent debugging #28 trajectory_events SQLite table for subagent debugging — telemetry sink could be the same table.
• feat(observability): subagent runs bypass TrajectoryRecorder #33 subagent runs bypass TrajectoryRecorder — fix as part of this child since swarm workers ARE subagents.
• Interactive multi-turn test harness for agents #20 interactive multi-turn test harness — needed to evaluate the role prompts.