L4472: split Saturday Backtester SF state into 3 (simulate / predictor / optimizer)

[cipher813]

Context — ROADMAP L4472 (⏰ soft deadline: 2026-06-06 09:00 UTC Saturday SF)

The single Saturday Backtester state ran simulate+param_sweep+predictor-backtest+Phase4+optimizer/cov/gamma in one SSM command whose summed runtime exceeded the SSM execution timeout on a fresh date (L4470) → no backtester/parity/evaluator completion → config auto-apply froze (L4473). Each post-sweep phase re-runs the predictor pipeline (GBM inference + 10y ArcticDB read), so it was death-by-sum-of-runtime with no single per-phase cap tripped. The 06-06 run repeats this failure unless this lands first.

The split

Decompose the backtest stage by --mode into three sequential SF states, each with its own SSM execution timeout + independent redrive (mirrors the 2026-05-07/05-16 simulate/parity/evaluator split precedent):

State	Invocation	Runs
Backtester	--mode=param-sweep --no-pit-parity	simulate + param_sweep
PredictorBacktest (new)	--mode=predictor-backtest	predictor pipeline + Phase 4a/b/c; pit_parity here
PortfolioOptimizerBacktest (new)	--mode=portfolio-optimizer-backtest --no-pit-parity	optimizer_gate + cov_sweep + gamma_sweep

Happy path: CheckSkipBacktester → Backtester → PredictorBacktest → PortfolioOptimizerBacktest → CheckSkipParity → Parity → CheckSkipEvaluator → Evaluator.

• skip_backtester still skips the whole backtest-family (routes past CheckSkipParity to CheckSkipEvaluator) — unchanged.
• pit_parity fires exactly once (default-ON in PredictorBacktest; the other two pass --no-pit-parity).
• Each new state mirrors Backtester's Retry / Catch(HandleFailure) / Timeout posture, distinct ResultPaths; on failure the next state is not entered (anti-auto-promote-garbage). Decouples Backtester→Parity so pit_parity.json finally lands.
• No backtester-repo logic change needed for dispatch — spot_backtest.sh already threads --mode/--no-pit-parity/--skip-stages.

Safety verified before wiring

• Each --mode is self-contained (re-reads from S3/ArcticDB; no in-memory handoff between phases) → 3 sequential processes == 1 --mode=all process for artifacts.
• The one semantic fork (executor-recommendation artifact collision) is closed by the companion alpha-engine-backtester PR (gate predictor's executor-apply on mode=="all").

Testing

• Updated SF pins: test_sf_backtest_parity_split_wiring (3-way split + new TestL4472PhaseSplit), test_sf_eval_judge_wiring, test_sf_payload_uniqueness (spot-state count 8→10), test_sf_friday_shell_run_wiring (_SPOT_STATES + main-thread trace).
• Regenerated tests/fixtures/sf_prekeystone_spot_commands.json (documented workflow for a deliberate spot-command change).
• SF graph validated (62→72 states, no dangling transitions). Full data suite: 1731 passed, 1 skipped.

Deploy

SF deploy is a manual operator step (infrastructure/deploy_step_function.sh → aws stepfunctions update-state-machine). Must be deployed before the 06-06 09:00 UTC Saturday cron.

Companion PRs

• alpha-engine-backtester#269: executor-apply mode-gate + simulation cap (correctness).
• alpha-engine-config: enable assembler.cutover_enabled + ROADMAP follow-ups.

🤖 Generated with Claude Code