feat: wire eval-cache + eval-tier into LLM judge, pin E2E model

callJudge/judge now return {result, meta} with SHA-based caching
(~$0.18/run savings when SKILL.md unchanged) and dynamic model
selection via EVAL_JUDGE_TIER env var. E2E tests pass --model from
EVAL_TIER to claude -p. outcomeJudge retains simple return type.
All 8 LLM eval test sites updated with real costs and costs[].

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

TODOS.md

@@ -231,7 +231,7 @@
 
 **Why:** Spot quality trends — is the app getting better or worse?
 
-**Context:** QA already writes structured reports. This adds cross-run comparison.
+**Context:** `eval:trend` now tracks test-level pass rates (eval infrastructure). QA-run-level trending (health scores over time across QA report files) is a separate feature that could reuse `computeTrends` pattern from `lib/cli-eval.ts`.
 
 **Effort:** S
 **Priority:** P2
@@ -335,6 +335,8 @@
 
 **Why:** Reduce E2E test cost and flakiness.
 
+**Status:** Model pinning shipped (session-runner.ts passes `--model` from `EVAL_TIER` env). Retry:2 still TODO.
+
 **Effort:** XS
 **Priority:** P2