feat(backtester): spot_backtest.sh --preflight-only (Friday shell-run dry path, smoke-only, no auto-apply)

[cipher813]

Owed-item #3 of ROADMAP "Friday shell-run — per-module dry-path activation" (P1).

Under the Friday shell_run, the Backtester / Parity / Evaluator spot state now boots the spot for real, installs deps, runs the existing bootstrap-class smoke harness, then exit 0 — with NO param sweep, NO portfolio sim, NO parity, NO evaluator, NO config/*.json auto-apply, ZERO external API calls, ZERO S3/config writes — catching bootstrap-class breakage (lib-pin drift, sys.path collision, stale ArcticDB universe, missing predictor weights, SSM timeout, image gap) ~12h before the real Saturday Backtester.

Flag name (verbatim — SF keystone follow-on)

--preflight-only — identical to the data (spot_data_weekly.sh #259) and predictor (spot_train.sh #175) siblings. PREFLIGHT_ONLY is a MODIFIER (default 0, orthogonal to RUN_MODE), mirroring the data sibling, so an unflagged Saturday SF run is completely unaffected.

Insertion point

New if [ "$PREFLIGHT_ONLY" = "1" ] branch placed AFTER the real boot/clone/deps/config-upload + ENV_SOURCE/REMOTE_PYTHON resolution (so the bootstrap path is genuinely exercised) and STRICTLY BEFORE both the if [ "$RUN_MODE" = "smoke-only" ] block and the # ── Full backtest heredoc. It runs one run_remote heredoc (backtest.py --mode=smoke) then exit 0.

Reuses the existing --mode=smoke?

Yes. Per backtest.py:4180-4184, --mode=smoke runs BacktesterPreflight (lib-version / imports / predictor-weights presence / executor-config validation — PRs #43-#48, 2026-04-22) + _runtime_smoke (universe symbols + per-ticker ArcticDB read + recent signals.json load + Layer-1A GBM load/predict — all S3 reads, ~30-60s), then returns BEFORE _init_pipeline / the simulation / the optimizer. So --mode=smoke itself performs zero config writes and makes no external API (yfinance/Anthropic) data fetch. No parallel preflight was built.

No-sweep / no-parity / no-auto-apply proof (statically unreachable under the flag)

• The per-phase smoke loop (smoke-simulate / smoke-param-sweep / smoke-predictor-backtest / smoke-phase4 / smoke-predictor-param-sweep) AND the evaluate.py --mode diagnostics S3-probe block both live inside the if [ "$RUN_MODE" = "smoke-only" ] body — past the exit 0.
• The full-backtest heredoc (backtest stage → pit_parity → parity → evaluator) and its config/{executor,scoring,predictor,research,scanner}_params*.json optimizer auto-apply (evaluate.py --upload, non-frozen) live further below — also past the exit 0.
• The CloudWatch heartbeats, parity_report.json / parity_metrics.csv upload, and reporter S3 upload are all past the exit 0.
• The preflight heredoc body references ONLY backtest.py --mode=smoke (no --upload, no full mode, no evaluate.py, no aws s3 cp/sync, no put-metric-data).

The trap cleanup EXIT still fires and terminates the spot.

Tests

New tests/test_spot_backtest_preflight_only.py (7 tests, static-analysis style mirroring test_spot_backtest_aws_region.py): flag parses, MODIFIER default 0, dedicated branch exits 0, branch precedes the smoke-only block + full-backtest heredoc, body invokes ONLY --mode=smoke with zero sweep/sim/parity/evaluator/--upload/auto-apply tokens, ENV_SOURCE preserved (#247 AWS_REGION guard).

• Full suite: 1665 passed, 5 skipped, 1 deselected
• bash -n clean
• No new deps, no secrets

🤖 Generated with Claude Code