feat(bench): add condition matrix, tracing, oracle grading, and report to the agent benchmark

[e54-bot]

Refs #414. Implements the harness side of SPEC-414 (merged as #459). The spec is not complete: see "Not included".

What it adds (benchmarks/agent, split into small modules beside agent_bench.py):

• Conditions: run --wright-level none|bin|bin+skill --knowledge none|wiki|web --network off|on, and matrix (agents x cells x scenarios x trials, shuffled by a seed, resumable, infrastructure exit 75 retried). Each run gets a fresh HOME, an allowlisted environment, and canaries; a failed canary marks the run invalid.
• Adapter contract: the cell reaches the agent as BENCH_* variables; the adapter reports per-turn usage, a transcript, and its loaded context. adapters/claude_code.py is a reference adapter. A pi adapter only needs to implement the same three files.
• Grading: validity per authority (wright check, wright compile, emitted Workshop text, and the pinned upstream OverPy 9.7.10 oracle under oracle/), with disagreement output as an owner-issue reproducer. New check kinds: oracle, wright-compile, compiled-contains/compiled-absent. validate also calibrates negative/ overlays. Results carry a grader hash.
• Trace: the shim records envelope fields (ok, exit, diagnostic codes, input_identity, selection) and tees wright serve sessions; the entry file is snapshotted after each write and every snapshot is graded, giving first-valid index and regressions.
• Analysis: expectation detectors E01-E04, E06, E08, E11, E12 and friction counters; token, context and to-first-valid metrics from adapter usage; estimated Wright output tokens per command. E05, E07, E09, E10 report unavailable until a normalized transcript exists.
• Report: report [--regrade] gives usable rates with Wilson intervals, tokens per usable result, paired comparison (tokens only where both are usable), per-scenario/split tables, expectations, friction, and eval diagnostics from the automated eval-design article (headroom, variance, infrastructure failures, grader consistency).
• Result contract is now wright-agent-bench/v2; the old baseline/wright conditions are replaced by the levels above. docs/agent-benchmark.md and the spec (REQ-025 to REQ-027, Q-004) are updated.

Verification

• python3 -m unittest discover -s benchmarks/agent: 19 tests pass against a wright 0.4.0 binary, including the real oracle after setup-oracle; without node they skip the oracle test only. validate passes for all four scenarios.
• One real end-to-end trial through the Claude Code adapter (bin+skill, understand-score-flow): usage, loaded context, and the guide were recorded, and the report rendered. That is a smoke check, not a benchmark result.
• CI runs validate and unittest as before; neither needs node.

Not included

• New OPY scenarios (ana-paintball and the others in REQ-005), so no scenario uses oracle or compiled-contains yet. They need reference solutions and negatives first.
• The guide-tuning loop (REQ-027).
• The offline network sandbox: the harness runs the canary and invalidates on failure, but enforcement is the adapter's job, and the reference adapter only removes web tools.
• Transcript normalization for E05, E07, E09, E10.