test_no_new_torch_before_dotenv_vm_entrypoints red on main: 4 scripts missing load_dotenv, reds Step 9c gates fleet-wide

Goal

Restore tests/test_shared_vm_thread_caps.py::test_no_new_torch_before_dotenv_vm_entrypoints to green on main by adding the missing load_dotenv calls to the four violating entrypoints, so the Step 9c test gate stops reporting a pre-existing fleet red as if it were the task's own payload failure.

Evidence — reproduced on clean main by the orchestrator, not inherited from a report

$ git rev-parse --abbrev-ref HEAD
main
$ uv run pytest tests/test_shared_vm_thread_caps.py::test_no_new_torch_before_dotenv_vm_entrypoints -q
rc=1
FAILED tests/test_shared_vm_thread_caps.py::test_no_new_torch_before_dotenv_vm_entrypoints

Violators named by the failure, four scripts, all missing a load_dotenv before their torch import:

• scripts/issue1901_mlpdense_fold_analysis.py
• scripts/issue1901_mlpdense_fold_figures.py
• scripts/issue2254_firstk_ctxext_sensitivity.py
• scripts/issue2378_lenmatch_fig.py

The test is on main and so are the violators, so this is not a feature-branch artifact.

Why it matters beyond one task

The Step 9c test-verdict gate selects tests by changed-file mapping, so any task whose payload maps onto test_shared_vm_thread_caps.py inherits a red that has nothing to do with its own diff. That forces every affected session to spend a review round establishing "pre-existing, payload-external" attribution — which is exactly the verification the gate exists to make unnecessary. Task #2546 paid that cost across four consecutive review rounds (rounds 9-12), each one re-deriving the same attribution: the round-9 breadcrumb, round 10's re-fenced 1500 s lint run, round 11's branch-vs-main skew check, and round 12's independent re-confirmation.

Two failure modes it creates, both worse than the red itself:

1. Attribution work repeated per session — every affected task re-establishes the same fact, and a reviewer who skips that step either blocks a clean payload or waves through a genuine failure.
2. Desensitization — a standing red in a gate trains sessions to treat that gate's failures as background noise, which is how a real regression in the same file would ship. This is the same dynamic the .claude/rules/repo-root-uncommitted-state.md § watcher-pass note warns about for recurring alerts.

Scope

• Add the missing load_dotenv (orchestrate.env.load_dotenv, per the lint rule --check-dotenv-before-hf-import and CLAUDE.md § Upload Policy: "New direct-upload scripts use orchestrate.env.load_dotenv, never bare dotenv") to each of the four scripts, placed before the torch/HF import the test checks.
• Re-run the test to green on main.
• Confirm the four scripts still run (or at minimum still import) after the insertion — a load_dotenv at the wrong position can shadow an intentional env override.

Do NOT weaken, skip, or xfail the test — it guards a real property (env must load before torch so thread caps and HF cache redirects apply on the shared VM). The fix is the four missing calls.

Provenance

Surfaced by the /issue 2546 round-12 implementer while attributing its own lint/pin-sweep results, and then independently reproduced on clean main by the orchestrator before filing (rc=1, same four scripts). Related standing rule: .claude/rules/gotchas.md § dotenv-before-torch; lint twin workflow_lint.py --check-dotenv-before-hf-import.