feat: add --core tasks and judge result caching

[ScuttleBot]

Summary

Two speedup features for faster benchmark iteration:

1. Core Tasks (--core flag)

Runs ~21 representative tasks (~17% of full suite) for quick signal:

• At least one task per category
• Favors automated grading for speed
• Includes mix of difficulty levels

benchmark.py --model anthropic/claude-sonnet-4 --core

2. Judge Result Caching

Caches LLM judge responses by hash of (task_id, transcript, rubric, model):

• Skips re-grading if inputs unchanged
• Persistent cache in results/.judge_cache/
• Logs hit rate at end of run

# Normal run with caching (default)
benchmark.py --model ... 

# Disable caching
benchmark.py --model ... --no-judge-cache

# Clear cache before run
benchmark.py --model ... --clear-judge-cache

Benefits

Scenario	Before	After
Full run, first time	~2h	~2h
Full run, same model	~2h	~30m (cache hits)
Quick iteration	~2h	~20m (--core)
Dev testing	~2h	~5m (--core + cache)

Closes #211 (partially - batch judging in separate PR)

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Solid additions across both reviews. The original judge cache and --core flag are well-structured, and the new Axiom observability layer properly opts in via env var (silently no-ops without AXIOM_API_TOKEN). The workspace snapshot fix correctly addresses the race condition between parallel grading and workspace teardown. The lib_agent.py session retry bump (15→90 attempts) and .trajectory file filter are clean targeted fixes.

Files Reviewed (8 files)

• scripts/benchmark.py - Axiom integration, workspace snapshot for parallel grading, score vars
• scripts/lib_axiom.py - new Axiom logging module
• scripts/lib_grading.py - judge cache + judge prompt clarification
• scripts/lib_agent.py - transcript retry bump, trajectory file filter, model validation fix
• scripts/lib_tasks.py - core_tasks field
• tasks/manifest.yaml - core task list
• BENCHMARK_VERSION - version bump
• README.md - whitespace

---

_{Reviewed by claude-4.6-sonnet-20260217 · 167,850 tokens}