feat(benchmark): AuditBench eval harness — B1 runnable baseline

[gnanirahulnutakki]

Closes #40 (Workstream B1 — make the benchmark runnable and reproducible).

What is runnable now

The go/benchmark/live/live.go stub that returned "unknown" for all four
arms has been replaced with a real, deterministic evaluator. The harness:

• Loads and validates .scenario.json + .events.jsonl file pairs
• Runs four independent evaluation arms over each trace
• Walks a pack directory, processes pairs in sorted order, and writes
  results.json + summary.csv to an output directory

Four evaluation arms

Arm	Checks
cedar_strict	Declared AllowedActions + AllowedTools — stateless
cedar_state	Same + cumulative tool_calls budget enforcement
visibility	All events must carry visibility: "full"
mcep_reconciliation	Per-event expected_label oracle

Four in-repo AuditBench scenarios (go/benchmark/testdata/)

Scenario	Ground truth	Arm differentiation
AB-01	compliant	All four arms agree
AB-02	violation	cedar_strict/state/mcep catch it; visibility does not (write was full-vis)
AB-03	violation	cedar_strict misses (tool is authorized); visibility + mcep catch it
AB-04	violation	cedar_strict misses (no tool violation); cedar_state + mcep catch budget overrun

The arms return meaningfully different verdicts (cedar_strict 50% accuracy,
cedar_state 75%, mcep_reconciliation 100% on the 4-scenario slice), confirming
the harness is doing discriminative evaluation.

Running it

# run the benchmark
make bench

# or directly:
cd go && go run ./cmd/benchcheck -- ./benchmark/testdata

# results land in bench-results/results.json and bench-results/summary.csv

See REPRODUCE.md for the full step-by-step reproduction guide including
content-address verification of input fixtures.

Tests

All 15 Go packages pass. 7 new tests added (go/benchmark/live/live_test.go):

• Four end-to-end tests (one per scenario, asserting each arm's verdict)
• Pack walker test (ordering and skip counting)
• Two error-path tests (missing scenario / missing events)

ok  github.com/ArdurAI/ardur/go/benchmark/live   0.163s

No Python tests were touched; no Python tests regressed.

What remains for Workstream B2 (not in this PR)

The publicly-described headline corpus (independently human-labeled scenarios
from real agentic-AI traces) is absent from this repo by design — it carries
privacy-sensitive data and requires independent labeling to avoid evaluator
contamination. REPRODUCE.md documents this gap explicitly under
"What is NOT yet runnable."

B2 work (separate, gated PR):

• Ingest the independently-labeled corpus as a pack directory
• Recall/precision curves per arm at scale
• Statistical significance analysis (bootstrap CIs)
• cedar_strict arm backed by a compiled Cedar policy (not just declaration lists)