feat: evaluation framework - score, compare, and regression-test agent sessions (v0.8.0)

[Siddhant-K-code]

Closes #10

What

Adds agent-strace eval - a subcommand for scoring sessions, comparing them across runs, and integrating quality checks into CI.

Commands

# Score a session against configured scorers
agent-strace eval run abc123
agent-strace eval run abc123 --format json

# Compare two sessions side-by-side
agent-strace eval compare session-a session-b

# CI integration - exits 1 if any scorer fails
agent-strace eval ci abc123

# Dataset management
agent-strace eval dataset add --session abc123 --label "fix auth bug"
agent-strace eval dataset list
agent-strace eval dataset export > evals.jsonl

Scorers

All built-in scorers use zero new dependencies:

Scorer	What it checks
no_errors	No ERROR events in the session
regex	Pattern match against any event type
cost_under	Estimated cost <= max_dollars
files_scoped	All file ops within allowed path prefixes
duration_under	Session duration <= max_seconds
custom	Any callable returning float in [0, 1]

Example output

Session: abc123
----------------------------------------------------------------------
  Scorer              Score  Threshold    Status    Reason
----------------------------------------------------------------------
  no_errors            1.00       1.00  v pass    no errors
  cost_under           0.82       1.00  x fail    $0.61 actual > $0.50 limit
  files_scoped         1.00       1.00  v pass    all files within allowed paths
  duration_under       1.00       1.00  v pass    4.0s <= 120.0s
----------------------------------------------------------------------
Overall: 3/4 passed

Config (.agent-evals.yaml)

scorers:
  - type: no_errors
    threshold: 1.0
  - type: cost_under
    max_dollars: 0.50
    threshold: 1.0
  - type: files_scoped
    allowed_paths: ["src/", "tests/"]
    threshold: 1.0
  - type: duration_under
    max_seconds: 120
    threshold: 0.8

thresholds:
  pass: 0.85
  warn: 0.70

Parsed with a stdlib-only YAML parser - no PyYAML dependency.

Implementation

src/agent_trace/eval/
├── __init__.py     # cmd_eval dispatcher
├── scorers.py      # built-in scorer implementations
├── config.py       # .agent-evals.yaml loader (stdlib YAML parser)
├── dataset.py      # dataset CRUD (JSONL-backed)
└── runner.py       # eval execution, formatting, compare, CI

Dataset entries stored as JSONL in .agent-traces/datasets/. No database, no external service.

41 new tests covering all scorers, dataset CRUD, runner formatting, compare output, and CI exit codes.