feat(judging): add multi-sample pass^k grading

[slowdini]

What and why

Repeated LLM verdicts over one execution expose judge consistency without paying to rerun the agent. This adds assertion-level and campaign-wide judge sampling, then carries the individual verdicts, vote proportion, and pass^k endpoint through grading and benchmark aggregation.

The implementation builds on the bounded judge evidence bundle from #288: every sample for a run receives the exact same persisted judge-evidence.md, so sampling varies the judgment rather than the evidence under review.

Closes #262.
Part of #244.

Strategy

• Resolve the effective count from llm_judge.samples, then run --judge-samples, then the compatibility default of one.
• Fan one authored LLM assertion out to independently resumable __sample-N judge tasks while reusing one bounded evidence bundle. The framework-owned __skill_invoked meta-check remains single-shot.
• Treat a missing response as a failure for that sample only, preserving the other responses and their evidence and confidence values.
• Store sampled assertions as ordered verdicts plus passed, failed, total, vote proportion p, and pass_power_k = p^N. Mixed gradings weight each authored assertion equally: binary assertions contribute zero or one, and sampled assertions contribute their derived endpoint.
• Aggregate per-condition vote-proportion and pass^k statistics, deltas, and pooled per-assertion votes. Fully binary campaigns retain the existing Fisher exact sample-size floor; sampled campaigns print the non-binary endpoints without presenting them as significance tests.

An eval can opt into sampling directly:

{
  "id": "clear-review",
  "type": "llm_judge",
  "rubric": "The review identifies the most important defect and explains its impact.",
  "samples": 10
}

Operators can instead set a campaign default:

eval-magic run --skill ./my-skill --harness claude-code \
  --judge-samples 10

The assertion value takes precedence over the command-line default. With six passing verdicts out of 10, the retained vote proportion is 0.6 and pass^k is 0.6^10, approximately 0.006047.

Compatibility and safety boundaries

• An effective count of one preserves the response filename, binary assertion fields, grading summary, and benchmark passed / n rollup.
• Sampled task filenames are validated against authored assertion IDs, preventing ambiguous paths such as an authored quality__sample-1 colliding with a generated sample for quality.
• The judge evidence remains bounded, persisted once, and read-only. Sampling does not rerun the agent, rebuild the bundle, or expose held-out command-check results.
• Mode B uses the same sample-aware artifacts and response paths. Promotion retains the evidence bundle and benchmark data, while existing provenance fields continue to carry the source revision.
• The schemas accept the sampled artifact variants while retaining the legacy binary variants for existing consumers and historical data.

Verification

• cargo fmt --check
• cargo build
• cargo test — 1,300 tests passed
• cargo clippy --all-targets -- -D warnings
• git diff --check

Review focus

The main review surface is the compatibility split between one-sample binary artifacts and multi-sample artifacts, plus the equal-assertion weighting used when a run mixes binary and sampled assertions.