feat(campaign): llmJudge single-call bridge + prune dead exports

[drewstone]

What

Builds the llmJudge(name, prompt, opts?) helper that the JudgeConfig doc-comment (src/campaign/types.ts) promised but that never existed, and prunes a set of confirmed 0-consumer dead/duplicate exports.

A) llmJudge — the single-LLM-call judge bridge

llmJudge(name, prompt, opts?) returns a canonical JudgeConfig whose score():

1. makes one call against prompt through an injected ChatClient (the substrate's transport-agnostic LLM seam — router / sandbox / cli-bridge / mock),
2. parses the model's per-dimension scores, validating every declared dimension is present and numeric,
3. normalizes by scale (unit [0,1] default, or ten [0,10]→[0,1]) and clamps to [0,1],
4. composites via weightedComposite (the same sum-normalized reducer ensembleJudge uses),
5. returns the canonical JudgeScore { dimensions, composite ∈ [0,1], notes }.

Fail-loud: an unparseable response or a missing/non-numeric dimension throws JudgeParseError, so the campaign engine records a failed cell instead of a silent zero. Exported from the root barrel and /campaign. New test (tests/llm-judge.test.ts, 12 cases) proves a well-formed canonical JudgeScore, scale normalization, weights, fenced-JSON parsing, single-call behavior, and the fail-loud paths.

This is the landing pad the 6 deprecated judges migrate to next (separate pass).

B) Pruned confirmed 0-consumer dead exports

Each re-verified with a fleet-wide grep (excluding node_modules + agent-eval's own checkouts):

• adversarialScenarioSearch removed — fuzzAgent is the live twin. AdversarialMutation is kept (the fuzz harness consumes it).
• proposeAutomatedPullRequest removed — superseded by campaign openAutoPr. Its input validation (path-traversal / duplicate-path / branch==base / empty-title guards) is folded into each AutoPrClient.proposeChange so the safety invariant stays on the live path. httpGithubClient / ghCliClient are kept — they have live fleet consumers (tax-agent, gtm-agent, creative-agent production loops).
• buildByAxis / summariseRows un-exported from the /matrix barrel — internal aggregation helpers behind runAgentMatrix, with no external consumers.

C) Investigated the 0-consumer yellow set — kept all 7

Each is a deliberate published-API affordance and/or internally consumed:

• runAgentMatrix — live external consumers (tax-agent, blueprint-agent import it from /matrix).
• mcnemar — the exact paired-binary significance test, paired with the live mcnemarPower/mcnemarRequiredN, has its own test.
• exportRewardModel — tier-3 RL reward-model export product surface (tested, referenced in docs).
• summarizeWorkflowTrace — heavily used internally + the /workflow public surface.
• fromClaudeCodeSession / fromOpenCodeSession — members of the documented harness-intake adapter family + internal belief-state consumer.
• parseAgentTrace — documented /contract/intake adapter affordance.

Kept

The 6 deprecated judges (createCustomJudge/createDomainExpertJudge/codeExecutionJudge/coherenceJudge/adversarialJudge/defaultJudges) and the legacy JudgeFn type — their consumer migration onto llmJudge is the next pass.

Verification

• pnpm run typecheck clean
• pnpm run lint clean (no errors)
• pnpm test — 248 files, 2542 passed, 2 skipped, 0 failed
• pnpm run build clean; verified llmJudge is a function exported from the built dist/index.js + dist/campaign/index.js, and score() returns a canonical JudgeScore from the built artifact

Version

0.98.0 → 0.99.0 (additive helper + removals = minor). npm + PyPI trio bumped together (package.json, clients/python/pyproject.toml, __init__.py).

[tangletools]

✅ Auto-approved PR — 98c9d027

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T08:05:26Z}