feat(bench): correction rounds and multi-reference paired lift for the agent benchmark

[e54-bot]

Refs #466. Stacked on #492 (feat/474-mcp-level).

Three additions to the agent benchmark from the issue's punch list, plus the transport normalization they need.

Correction rounds (correctionRounds)

Counts failed validation -> workspace edit rounds per trial: the condition tool's validating op reporting exit 1 followed by an edit (snapshot of a watched file changing).

• Wright: check/lint/analyze/compile — CLI calls under bin, tools/call serve ops under mcp/stdio/jsonrpc. overpy compile counts under the opy cell, so the comparison stays fair across conditions.
• Consecutive failures before one edit count once; a pass resets the sequence.
• Usage errors (exit 2), crashes (exit 3/4), serve refusals, protocol errors, and unanswered requests are all neutral — they neither count as a correction nor clear a pending failure.
• Reported as a corr column in the outcome table and correctionRounds per cell in summary.json.

serve_result_payload normalizes op payloads across transports: {"result": envelope} for stdio/jsonrpc, result.content[] text blocks for MCP; isError refusal payloads ({code, message}, no exit) are detected and skipped.

Repeatable --reference on report and evaluate

--reference LABEL is now repeatable: the report emits one Paired against \LABEL`section per named reference, so lift can be measured againstdocs, a skill-only control, or any other named cell — not only none/none/off`. Default is unchanged (baseline only).

workshop-wiki → workshop-skill rename

The generated wiki skill was named workshop-wiki while the condition vocabulary (SKILLS, cells, docs) calls it workshop-skill — the generated artifact never installed under the name the harness records. SKILL_NAME, the output-dir requirement, BUILD.json identity, help, examples, docs, and fixtures now agree.

(Already in place, per the punch list: clustered-bootstrap intervals via rate_runs/cluster_interval, and model/effort/skill-hash/tool-list recording via agent.id + agentInfo + environment.skills.)

Verification

• 111 unittest tests pass with WRIGHT_BIN (correction-round coverage for CLI calls, stdio serve ops, MCP payload unwraps, overpy compile, refusal/error/unanswered neutrality, multi-reference pairing, and the renamed skill fixtures).
• Scripted agents through the real harness: a deliberately failed check then repair yields correctionRounds: 1 under wright/none/off and 0 under none/none/off; a two-reference report renders both paired sections.
• git diff --check clean.

Independent review

One review pass done on the metrics commit: found that non-0/1 exits cleared pending state like a pass (fixed — only exits 0/1 mark a verdict now), plus nits applied (docstring lag caveat, payload.get("exit"), wording). Deferred pre-existing finding: the jsonrpc transport maps compile requests to jsonrpc:compile rather than the compile op (predates this diff, also affects toolUse/friction/E02/E04) — worth a follow-up issue.

Limits

• Edit markers use the snapshot poller's detection time (~0.2s granularity); sub-interval edit/validation orderings can merge rounds.
• correctionRounds observes only watched files (the scenario watch list); edits elsewhere are invisible.