feat(research): shell-run dry path for eval-judge chain + rationale clustering (closes SF skip-exceptions)

[cipher813]

Summary

Closes the two Saturday-SF Friday-PM shell-run keystone skip-exceptions (alpha-engine-data #260): skip_eval_judge + skip_rationale_clustering were hard-skipped under the Friday shell_run because no clean no-write dry path existed — so their import/bootstrap was never exercised on the Friday smoke.

• eval_judge_submit always _persist_client_side_skips (S3 put) + creates an Anthropic Message Batch.
• rationale_clustering's _persist_analysis S3 put_object was not gated by its existing dry_run flag (that flag only suppressed the CW metric).

Event flag (verbatim — the SF rewire passes this exact key)

dry_run_llm

Matches Research's existing convention (step_function.json: "dry_run_llm.$": "$.research_dry"). Single canonical shared substrate in evals/lambda_dry.py — no per-handler bespoke logic, no copy-paste, not coupled to dry_run.py (that installs LangGraph agent stubs, irrelevant here).

Per-handler short-circuit points

Boot + module-level imports run for real first (the keystone's whole point), then handler() returns before any external call:

Handler	Returns before	Dry result
eval_judge_submit_handler	build_batch_plan / _persist_client_side_skips (S3 put) / submit_batch (Anthropic Batch create)	{status:"EMPTY", batch_id:"dry-run-no-batch", processing_status:"ended_empty", plan_s3_key:None}
eval_judge_poll_handler	anthropic.Anthropic / poll_batch	terminal processing_status:"ended"
eval_judge_process_handler	anthropic.Anthropic / process_batch_results (S3 plan get_object + results stream + per-artifact persist + CW)	status:"OK"
rationale_clustering_handler	compute_and_emit (decision_artifacts/ read + _analysis/ S3 persist + CW emit), after the evals.rationale_clustering import	status:"OK"

Dry sentinel threading submit → poll → process

submit returns batch_id="dry-run-no-batch" + status="EMPTY". The existing EvalJudgePollChoice SF Choice routes status == "EMPTY" straight to EvalJudgeProcess, skipping the poll loop entirely. Poll + Process also treat the threaded sentinel batch_id as dry defensively (belt-and-braces if the SF wiring reaches Poll).

No-Anthropic / no-write proof

tests/test_eval_judge_shell_run_dry.py asserts, per handler under dry_run_llm:

• anthropic.Anthropic not called (assert_not_called)
• boto3.client / process_batch_results / compute_and_emit / submit_batch / _persist_client_side_skips / poll_batch / build_batch_plan not called → zero S3 put/get, zero CW
• success status returned (EMPTY/ended/OK)
• boot-still-runs smoke: the dry short-circuit is inside handler() after module imports, so a broken import still fails the dry pass
• legacy rationale dry_run flag behavior preserved (still runs compute_and_emit, only suppresses CW) — only the new dry_run_llm is the full short-circuit; non-dry path still takes the real track

Tests / suite

+14 tests. Full research suite: 1357 passed (1343 → 1357). No new deps, no secrets/prompts committed.

Follow-on (separate repo)

The SF rewire to route EvalJudge* / RationaleClustering → dry (pass "dry_run_llm.$": "$.research_dry", drop the skip_eval_judge / skip_rationale_clustering hard-skips from ApplyShellRunDefaults) is a separate alpha-engine-data follow-on — not in this PR.

🤖 Generated with Claude Code