issue2225_fu2_dod_points_fig.py imports numpy before load_dotenv — reds test_no_new_torch_before_dotenv_vm_entrypoints on main

scripts/issue2225_fu2_dod_points_fig.py violates the #847 thread-cap invariant, reddening test_no_new_torch_before_dotenv_vm_entrypoints on origin/main

What is red

tests/test_shared_vm_thread_caps.py::test_no_new_torch_before_dotenv_vm_entrypoints FAILS on a pristine origin/main tree:

E   scripts/issue2225_fu2_dod_points_fig.py (module-top heavy import at line 23,
                                             first load_dotenv( at line None)

Evidence (all verified against origin/main, not inferred)

• The file is TRACKED on origin/main — git cat-file -e origin/main:scripts/issue2225_fu2_dod_points_fig.py succeeds.
• It landed in faeb45f5e3, 2026-08-14 17:11:30 -0700, "task #2225 fu2: per-question dose-change view behind the direction-specificity contrasts".
• Line 23 is import numpy as np at module top; grep -n 'load_dotenv' over the file returns NOTHING.
• It is NOT in GRANDFATHERED_TORCH_BEFORE_DOTENV.
• The failure was produced by a full-suite run on a scratch worktree detached at origin/main exactly (6f72f73d84), with cwd + PYTHONPATH inside that scratch — so it is main-side red, not a worktree artifact. Result of that run: 1 failed, 3867 passed in 1861.17s.
• The test resolves its scan root tree-locally (root = Path(__file__).resolve().parents[1]) and the violator is tracked, so this is NOT the #2209 cross-worktree untracked-stray path.

Why it matters

The invariant exists because env.load_dotenv() binds the shared-VM BLAS/intra-op thread caps IN-PROCESS only when it runs BEFORE the import that freezes the pools (#847). A module-top numpy import with no prior load_dotenv() means running this figure script on the shared VM gets no thread caps — the #847 incident shape (5-6 uncapped jobs drove load to 186-226 while each realized ~5-6 cores).

Secondary effect: any session running the full suite at the shared root sees this as red. The #1388 class (inline-landed lint-red scripts breaking the Step 9c gate fleet-wide) is the precedent for treating a main-side red as its own task.

Fix options (implementer's call)

1. PREFERRED — add explore_persona_space.orchestrate.env.load_dotenv() before the heavy import, matching the convention in sibling figure scripts. The test's own docstring is explicit that the #847 offender "was FIXED, not grandfathered — keep it that way", so grandfathering is the weaker option.
2. Grandfather the path only if there is a real argument the script never executes on the shared VM.

Either way the currency tests over the grandfather list must stay green.

Scope note

Filed rather than fixed in-flight by the #2296 session that found it: the file belongs to another task's payload, and editing it mid-landing would invalidate the tip that session's Step 10d gate is currently certifying. #2296's own landing is unaffected — its mapped-invariant baseline found this failure on the pristine main tree, so the both-trees-red subtraction strips it rather than blocking.

Discovered by #2296's Step 10d lint gate round 3 (baseline leg, 2026-08-15T06:37Z). Owning task of the file: #2225 (parked at awaiting_promotion).