feat(sf): PR 3c LLM-as-judge wiring on Saturday Step Function

[cipher813]

Summary

Wires the alpha-engine-research-eval-judge Lambda (PR #91) into the Saturday Step Function. Adds two-tier sampling driven by the SF execution date — monthly Sonnet sweep on the first Saturday of each month, Haiku-only with per-artifact escalation on the others.

SF flow update

CheckBacktesterStatus (Success)
  → CheckSkipEvalJudge        (skip_eval_judge bypass → SaturdayHealthCheck)
  → ComputeEvalCadence        (Pass: extract day_of_month + eval_date from
                               $$.Execution.StartTime via intrinsic StringSplit)
  → CheckMonthlyCadence       (Choice: day_of_month < '08' lex compare)
      → EvalJudgeFirstSaturday  (Lambda: force_sonnet_pass=true)
      → EvalJudgeWeekly         (Lambda: force_sonnet_pass=false)
  → SaturdayHealthCheck → NotifyComplete

Both eval Tasks Catch: States.ALL → SaturdayHealthCheck. Eval is observability per ROADMAP §1635 — pipeline must NOT halt on eval failure.

Cadence rationale

Monthly Sonnet sweep on the first Saturday of each month (matches ROADMAP §1626's "every 4th run" with a deterministic, stateless rule that approximates ~4-week intervals to within one week). The lex-compare on zero-padded '01'–'07' < '08' is correct because all day_of_month values are 2-char strings.

Other Saturdays run Haiku-only with the per-artifact <3 escalation gate inside the Lambda — borderline artifacts still get a Sonnet pass; only the full sweep is gated by cadence.

IAM

Adds alpha-engine-research-eval-judge* to the LambdaInvoke resource list in both deploy_step_function.sh and deploy_step_function_daily.sh. The shared-policy convention (deploy_step_function.sh:91-92) means the last script to run wins, so both must stay in sync.

Test plan

• pytest tests/test_sf_eval_judge_wiring.py (21 tests pin transition + payload shape).
• Full suite: 406 → 427 passing.
• Manual JSON validation: state graph reachable; Choice defaults correct (weekly not monthly).
• Live verification: Sat 2026-05-09 SF run. 5/9 is the first Saturday of May → ships the monthly Sonnet sweep on the first eval pipeline run, which becomes the calibration corpus for PR 4.

Deploy order

1. From alpha-engine-research: ./infrastructure/deploy.sh eval_judge (creates alpha-engine-research-eval-judge:live)
2. From alpha-engine-data: ./infrastructure/deploy_step_function.sh (updates SF JSON + IAM policy)

Order matters — if the SF picks up the new state before the Lambda exists, the EvalJudge state would fail-fast on first invocation (caught by the non-blocking Catch, but unnecessary noise in CloudWatch).

Out of scope (PR 4+)

• CloudWatch metric AlphaEngine/Eval/agent_quality_score + SNS rolling-4-week-mean alarm (ROADMAP §1634).
• Streamlit dashboard quality-trend page + prompt-version → quality-score correlation chart (ROADMAP §1632–1633).
• Human cross-validation writeup (ROADMAP §1627 calibration record).

🤖 Generated with Claude Code