feat: per-turn agent trace table for latency and outcome tracking

[cgcardona]

Background

Inspired by Kit-Daemon's trace-based learning pattern. We store ACAgentRun rows but capture almost nothing about what happens inside a run at turn granularity.

Problem

We cannot answer:

• Which issues take the most turns to complete?
• Which roles fail (cancelled/error) most often?
• What is the p50/p95 LLM response latency for local vs. cloud?
• Is Qwen 9B meaningfully slower than 14B for reviewer tasks?

All of this data flows through agent_loop.py and is silently discarded.

Proposed solution

Add a lightweight ACAgentTrace table that records one row per LLM turn:

id            UUID PK
run_id        FK -> ACAgentRun
turn_number   int
role          str  (developer, reviewer, etc.)
model         str  (which model was actually called)
input_tokens  int
output_tokens int
cache_read    int
cache_write   int
latency_ms    int  (wall-clock time for the LLM call)
stop_reason   str  (tool_use, stop, max_tokens)
tool_calls    int  (count of tool calls in this turn)
recorded_at   timestamp

agent_loop.py already has all of this data at the point it processes each LLM response -- it just does not persist it.

Acceptance criteria

• ACAgentTrace SQLAlchemy model in agentception/db/models.py
• Alembic migration
• persist.py: record_agent_trace() async function
• agent_loop.py: call record_agent_trace() after each LLM response
• queries.py: get_run_traces(run_id) and get_model_latency_stats() queries
• Unit tests for persist and queries
• mypy clean

Future use

This table feeds the complexity-based routing decision (issue #1011) and the degradation detection monitor (separate issue). It is also the foundation for any future prompt optimization work.