fix: SF stages experiment-package predictor.yaml + pins ALPHA_ENGINE_EXPERIMENT_ID (config#1066)

[cipher813]

What

Aligns the Saturday SF model-zoo + predictor-training cp/staging with alpha-engine-predictor/config.py's experiment-first search path so the rotation child spot deterministically loads a predictor.yaml WITH model_specs populated.

Closes the root cause behind config#1051 / alpha-engine-predictor#265 (the 6/13 inert ModelZooRotation: empty MODEL_SPECS → select_rotation_specs() returned [] → 0 challengers trained).

Diagnosed mismatch

The 2026-06-12 experiment-package adoption made config.py (L24-32) search ~/alpha-engine-config/experiments/$ALPHA_ENGINE_EXPERIMENT_ID/predictor/predictor.yaml FIRST, then fall back to the legacy predictor/predictor.yaml, then config/predictor.yaml. On the always-on box ALPHA_ENGINE_EXPERIMENT_ID is unset → defaults to reference, and the experiment-package yaml exists, so config.py loads it and ignores the file the SF cp writes (config/predictor.yaml, the last/lowest search path). Verified live on the box: config.py._CONFIG_PATH = .../experiments/reference/predictor/predictor.yaml, _EXPERIMENT_ID = reference. The legacy + experiment yamls are identical today only by coincidence; the cp writes a file config.py never reads on the box, and the child spot inherits the same ambiguity.

Fix (deterministic, no config.py edit)

• cp source → experiments/reference/predictor/predictor.yaml (the path config.py actually resolves on the box), so the box config/predictor.yaml and the spot's S3-staged copy of it equal config.py's resolution.
• export ALPHA_ENGINE_EXPERIMENT_ID=reference before spot_train.sh so the child spot bootstrap stages the yaml to the experiment-package path config.py searches first.

Both PredictorTraining (--full-only) and ModelZooRotation (--model-zoo-weekly) get the same treatment. Regenerated the frozen tests/fixtures/sf_prekeystone_spot_commands.json PredictorTraining baseline (friday-shell-run byte-identicality guard) for the deliberate command change.

Deploy

This data SF change AUTO-DEPLOYS on merge to main via deploy-infrastructure.yml.

Companion PR

alpha-engine-predictor#PREDPR — spot_train.sh stages to BOTH the experiment-package path AND config/predictor.yaml, pins the env in every spot heredoc.

Tests

bash -n clean; JSON valid; 653 SF/deploy/pin tests pass (incl. test_sf_friday_shell_run_wiring, test_sf_research_predictor_parallel_wiring).

Refs config#1066, config#1051. Re-exam 2026-06-20 — next Saturday rotation must register ≥1 spec-* challenger.

🤖 Generated with Claude Code

[cipher813]

Companion predictor PR: cipher813/alpha-engine-predictor#267 (spot_train.sh staging + env pin). Merge order does not matter — both must land for the next Saturday rotation; the data SF auto-deploys on merge.