fix(ci): make codestyle gates realistic and green-able

[jam-jee]

Problem

codestyle-doc-tests bundles 7 tools (black-check, flake8, pylint, docstyle, twine, sphinx, doc8) and has been red on every PR while merging anyway (non-blocking). Two tools are misconfigured for this codebase, making the check structurally unwinnable:

Tool	Issue	Measured
flake8-future-import (FI11)	Requires from __future__ import absolute_import in every file — a Python-2 relic, meaningless on py3.10+	~107 errors/submodule
pylint --fail-under=9.9	Gate far above actual scores; unreachable without large refactor	core 8.45, train 7.24, serve 7.84, mlops 7.50
pydocstyle	~600 errors/submodule, dominated by low-value/judgment rules (D400 period, D401 imperative mood, D1xx missing docstrings)	~600 → ~45 real after ignore

Changes (config only)

1. Remove flake8-future-import from requirements/tox/flake8_requirements.txt — eliminates the FI11 noise repo-wide. Real flake8 codes (E/W/F — including F401/F811/F841 that catch actual bugs) stay fully enforced.
2. pylint --fail-under → per-submodule floors just below current scores (core 8.0, train 7.0, serve 7.5, mlops 7.0). pylint now blocks regressions (score can't drop) while passing today — instead of a red check everyone ignores.
3. pydocstyle --add-ignore for the noisy/judgment rules (D400/D401/D1xx/…), so docstyle enforces only high-value checks (~45 real issues remain repo-wide, fixed separately) rather than blocking on style opinion.

Gate philosophy (what this establishes)

• Blocking, high-value, auto-fixable: black-check, flake8 (E/W/F), twine.
• Blocking against regressions: pylint at a realistic floor.
• Enforces only meaningful rules: docstyle.

Companion work

• Per-submodule black + flake8 code fixes so those envs pass (serve: fix(serve): fix codestyle-doc-tests (black-check + flake8) for sagemaker-serve #6125; train/core/mlops to follow).
• A CDK buildspec change (internal CR) to split the bundled check into a blocking group (black/flake8/twine) and an advisory group (pylint/docstyle) so a formatting miss doesn't hide a real bug and vice-versa.

Testing

• Verified per submodule with pinned tool versions: FI11 disappears with the plugin removed; pylint passes at the new floors; docstyle drops from ~600 to ~45 (real, high-value) errors.

[github-actions]

Claude finished @jam-jee's task in 2m 57s —— View job

---

PR Review — codestyle gate config

• Read PR diff
• Check flake8 config for orphaned flake8-future-import settings
• Check pylint/pydocstyle consistency across submodules
• Post review

---

Solid, well-motivated change — the rationale for each gate is clear and the per-submodule pylint floors are a good "block regressions, pass today" design. One issue needs attention before merge, plus a couple of smaller notes.

🔴 Blocking: orphaned flake8-future-import config left behind

The PR removes flake8-future-import from flake8_requirements.txt, but that plugin is what registers the require-code option. All four submodule tox.ini files still contain it:

• sagemaker-core/tox.ini:49 → require-code = True
• sagemaker-mlops/tox.ini:50 → require-code = True
• sagemaker-serve/tox.ini:50 → require-code = True
• sagemaker-train/tox.ini:50 → require-code = True

Since flake8 5.0, an unknown option key in the config file is a hard error (it was introduced specifically to catch typos). With the plugin uninstalled, require-code is no longer a registered option, so flake8 will abort at startup with something like Error: unrecognized arguments/critical config error — instead of running. That directly defeats this PR's goal of making the flake8 env green-able.

Fix: delete the require-code = True line from all four tox.ini files in this same PR.

Please verify by running flake8 in one submodule with the plugin uninstalled (i.e. after this PR's requirements change) — if it errors on require-code, this is confirmed. I couldn't execute it in this environment.

🟡 Minor: dead FI* ignore codes

Each tox.ini still lists FI10, FI12–FI18, FI50–FI58 in the [flake8] ignore block (e.g. sagemaker-core/tox.ini:30-46). These codes are only defined by flake8-future-import, so they're now no-ops. flake8 won't error on unknown codes in ignore (unlike require-code), so this isn't blocking — but for hygiene it's worth removing them in the same cleanup so a future reader doesn't think future-import checks are still active.

🟢 Looks good

• pylint floors — per-submodule values (core 8.0, train 7.0, serve 7.5, mlops 7.0) match the measured scores in the description and are consistently applied. Good regression guard.
• pydocstyle --add-ignore — identical ignore list applied uniformly across all four submodules; the selected codes (D1xx missing-docstring, D400/D401 judgment rules, etc.) are reasonable to defer. Consider whether the ignore list belongs in a shared [pydocstyle] config or the .pylintrc-style shared file rather than duplicated inline in four tox.inis, so it can't drift — but that's optional.

Consistency note

The two config knobs (--add-ignore, --fail-under) are applied identically across all four submodules, which is exactly right. The only inconsistency is the leftover require-code/FI* entries above — cleaning those up completes the "remove future-import" intent.