feat(predictor): spot_train.sh --preflight-only (Friday shell-run dry path, no promotion)

[cipher813]

Owed-item #2 of ROADMAP "Friday shell-run — per-module dry-path activation" (P1).

Under the Friday shell_run, the PredictorTraining spot state must boot the spot, run import + lib-pin + ArcticDB/universe-freshness connectivity checks, then exit 0 — with NO weight training, NO weight promotion, and ZERO S3/config writes. Catches bootstrap-class breakage (lib-pin drift, sys.path, stale ArcticDB, SSM timeout, image gap) before the real Saturday PredictorTraining.

Flag name (verbatim, for the SF keystone follow-on): --preflight-only

Insertion point

New if [ "$MODE" = "preflight-only" ] branch placed after the bootstrap + deps run_ssm steps and before the # ── Smoke test section. It runs one new run_ssm "preflight-only" step then exit 0, short-circuiting before the smoke step (dry_run=True training) and before the full-training step (dry_run=False).

No-train / no-promote / no-write proof — steps SKIPPED under the flag

• run_meta_training() is never invoked — the entire walk-forward + meta-model fit is skipped (it lives inside the smoke/full SSM steps that the exit 0 never reaches).
• meta_trainer.py's if not dry_run: upload/promote block — the only weights/meta/* + manifest + dated-archive/{date}/ writer — is unreachable.
• train_handler.main()'s training_summary_* write, triple-barrier cutover gate, training email, and health-status write all live after run_meta_training in main() — none are reached.
• The probe does only: import alpha_engine_lib + the training package (no main() call), run the existing TrainingPreflight (env-var check + S3 bucket reachability — a head/list, no put_object), and a read-only ArcticDB list_symbols() + single read().tail(1) universe-freshness probe. No put_object, no upload_file, no download_from_arctic, no config/* write, zero external API (yfinance / Anthropic) calls.

Reuses an existing python dry path?

Partially, and deliberately. The repo's python dry_run=True (run_meta_training) still trains the full model and only skips S3 writes — it is a "train-but-don't-promote" smoke, NOT a bootstrap-only preflight, so it does not meet the zero-train/zero-API invariant this item requires. This change reuses the existing training/preflight.py::TrainingPreflight (env + S3 connectivity) instead of rebuilding one, and reuses store.arctic_reader._get_arctic for the read-only ArcticDB probe. No parallel preflight was built.

Tests

tests/test_spot_train_preflight_only.py (mirrors test_spot_train_aws_region.py static-analysis style — the SSM/EC2 path can't run in CI): flag parses; dedicated branch exists + exit 0; branch precedes smoke + full-training; step body references no train/promote/write tokens (run_meta_training(/train_main(/put_object/upload_file/download_from_arctic); reuses TrainingPreflight + imports alpha_engine_lib + read-only list_symbols(); keeps the #247 AWS_REGION export guard. SSM-transport path for normal modes unchanged.

bash -n clean. Full suite 1050 passed (8 new/sibling infra tests green). No new deps, no secrets.

🤖 Generated with Claude Code