fix(analysis): cfg.RESEARCH_BUCKET → cfg.S3_BUCKET + CLI smoke test

[cipher813]

Summary

Three diagnostic CLIs that gate L2878 Stages 1d / 2d / 3 cutover decisions all referenced cfg.RESEARCH_BUCKET, an attribute that has never been defined in config.py (canonical name is cfg.S3_BUCKET). The stale symbol was introduced 2026-05-10 via PR #117 (variant gate substrate) and copied into the two sibling CLIs since. The CLIs died at import time during the argparse default lookup before any user-facing message:

AttributeError: module 'config' has no attribute 'RESEARCH_BUCKET'

Programmatic API was unaffected — the bug only surfaced when an operator actually ran the CLI for the first time. Latent for 18 days. Caught 2026-05-28 during audit of what's actionable on L2878 mid-observation-soak: Sat 5/30 is the 3rd parallel-observation Saturday firing, and the operator was about to want these CLIs runnable for diagnostic IC-lift checks.

Sites fixed (7 across 3 files)

File	Lines	Form
analysis/variant_cutover_gate.py	503, 504	CLI default + help string
analysis/horizon_battery.py	266, 267	CLI default + help string
analysis/triple_barrier_cutover_runner.py	226 (docstring), 251 (runtime fallback), 347 (CLI help)	Mixed

All replaced with cfg.S3_BUCKET. Default bucket now resolves cleanly to alpha-engine-research.

Chokepoint test (closes the bug CLASS at PR time)

tests/test_analysis_clis_invokable.py:

1. test_analysis_cli_help_works — parametrizes every analysis CLI module with a top-level main() through python -m <module> --help; asserts exit 0 + non-empty usage line. argparse's --help exits 0; any missing-attr error in argparse defaults surfaces as a non-zero exit code + stderr traceback.

2. test_analysis_cli_list_covers_every_module_with_main — walks analysis/*.py looking for any module that defines a top-level def main(; fails if a CLI is missing from the parametrized list. Catches the next "added a new analysis CLI but forgot to register it" recurrence.

Caught a 4th CLI (analysis/compare_modes.py) during test build; confirmed it already uses cfg.S3_BUCKET correctly and added to the parametrized list.

Composes with

• [[feedback_lift_invariants_to_chokepoint_after_second_recurrence]] — three sibling sites = strong second-recurrence signal; lift to PR-time test rather than fix each in isolation.
• L2878 — unblocks operator-driven diagnostic gate runs over the next 6 weeks of observation soak (Sat 5/30 → 6/06 → 6/13 → …).
• [[feedback_no_silent_fails]] — the argparse-default crash was loud, just unreachable until first CLI invocation. The CI test makes the failure surface at PR time.

Test plan

• pytest tests/test_analysis_clis_invokable.py -v — 5/5 passed.
• Touched-test suite — pytest tests/test_variant_cutover_gate.py tests/test_triple_barrier_cutover_runner.py tests/test_triple_barrier_cutover_runner_wave3_readers.py tests/test_analysis_clis_invokable.py -q — 47/47 passed.
• Manual CLI smoke — each of the 4 CLIs --help now exits 0 + prints a clean usage line with default bucket alpha-engine-research.
• CI Tests workflow passes.

🤖 Generated with Claude Code