fix(benchmarks): parse both v1 and v2 BENCHMARK.md formats, and unify agent identifiers

[mosheabr]

Holding this pending BI review — see the agent-identifier section. Not for merge yet.

Problem 1: re-signed skills drop out of the feed

SkillEvaluator 0.9.x emits a redesigned BENCHMARK.md. aggregate_benchmarks.py was written against the older layout, so every skill that re-signs silently loses its evaluation data: has_results goes false, with null tasks, no agents and no uplift, even though the new report carries richer data than the one it replaced.

Two skills re-signed this week and both lost their numbers:

Skill	before	after re-sign
nemo-automodel-distributed-training	PASS, uplift 44.6	PASS, has_results: false
nemo-automodel-model-onboarding	PASS, uplift 40.7	PASS, has_results: false

7 of 323 reports are v2 today. With 21 skills still to re-sign, the feed behind the adoption dashboard would empty out as the re-sign campaign succeeded.

Problem 2: every v2 report parsed as PASS

In v2 the verdict is a callout above the report body. The old pattern didn't match it, and instead matched the methodology bullet present in every v2 report:

- Overall verdict: PASS only when every configured dimension passes for at least one supported agent.

So a failing v2 skill would have been recorded as PASS. The v1 pattern is now anchored to end-of-line, and the v2 callout is matched explicitly.

Problem 3: the feed carries four identifiers for two agents

Independent of the format work, the results rows today contain:

identifier	rows
claude-code	1145
codex	1145
Claude Code (\aws/anthropic/bedrock-claude-opus-4-8`)`	335
Codex (\openai/openai/gpt-5.5`)`	330

67 skills were signed by an evaluator version that wrote the model ID into the results-table header. Anything grouping the feed by agent sees four agents, so per-agent aggregates split and undercount.

This PR normalises all four to claude-code and codex (1500 rows each). This rewrites 665 existing rows across 67 skills and is the change that needs BI sign-off — it is a data-shape change to a field the dashboard joins on, not just a parser fix.

What changed

• Evaluation Summary (v1) and Evaluation Metadata (v2) field lists
• ## Agents Used bullets (v1) and the inline - Agents: line (v2)
• ## Results and ## Results at a Glance tables, keyed on Dimension or Measure
• v1 cells (100% (+70%)) and v2 cells (45% → 98% (+53 points)), recording the skill-assisted score in both so the series stays comparable
• v2's aggregate Overall row skipped, so average_uplift_pct stays comparable with v1
• agent identifiers normalised (Problem 3)

Verification (local)

• skills_without_results 27 → 22; result_row_count 2955 → 3000
• Report census: 316 v1, 7 v2, 0 unrecognised
• All 7 changed skill records are v2 reports; no v1 skill record altered
• 665 result rows change, all from the agent-identifier normalisation, across 67 skills
• Verdict regression test: v2 report with a FAIL callout now parses FAIL (previously PASS); v1 FAIL still parses FAIL
• Schema unchanged: same keys on skill records and result rows, schema_version still 2
• aggregate_benchmarks.py --check passes against the committed file

🤖 Generated with Claude Code