feat(scripts): add check-encoding-safety.py (sibling of check-windows-footguns)

[pnascimento9596]

feat(scripts): add check-encoding-safety.py (sibling of check-windows-footguns)

Addresses #66668

What / why

The env-class cluster is not hypothetical: #66607 closed, #66609
merged to main via salvage #66825 (2026-07-18), authorship preserved
(51e1fb8) shows a UTF-8 BOM in jobs.json crashing load_jobs;
three open fix PRs (#65124, #66475, #66483) cover the same
user-writable encoding/write class; accepted-class #57247 / #57249 is
the same .env family Hermes users edit by hand. The salvage itself
widened the fix to three additional readers (backup.py,
curator_backup.py, prompt_builder.py context scan). This checker still
flags residual prompt_builder sites on the hit list
(prompt_builder.py:1401 and prompt_builder.py:1858), demonstrating the
class is live and the checker's coverage is real. CONTRIBUTING already
documents Notepad/Windows BOM provenance and mandates utf-8-sig for
Windows-touched config.

This PR lands step 1 of a smaller-ask ladder: an opt-in, stdlib-only
AST checker (scripts/check-encoding-safety.py) sibling of
scripts/check-windows-footguns.py. It flags reintroduction at review
time (R1 plain utf-8 on user-writable reads, R2 latin-1 fallback without
BOM strip, R3 errors="replace" rewrite) without changing runtime
behavior or wiring CI/docs yet.

Known recall boundary: opaque path locals (e.g. the two readers #66825 widened to) are not bound to the allowlist - precision-first tradeoff, assignment tracing is a follow-on; details in #66668.

How to test

venv/bin/python -m pytest tests/scripts/test_check_encoding_safety.py -q
# → 22 passed (Gate 1: 6/6 historical fixtures flagged; fixed forms clean)

venv/bin/python scripts/check-encoding-safety.py --all
# → 85 hits / 786 files @ d59b79fad at push; 81 hits / 791 files @ c78aa0bad as of 2026-07-18 (R1=75, R2=2, R3=4)
Update: 85/808 @ 614dc194e (root-module scan gap fixed per sweeper review; +4 hits are newly-visible top-level modules). -> 97/850 @ bd6437d60 (rebase only, no rule changes)
# Classification (hardened): A=24 / B=49 / C=12

venv/bin/python scripts/check-windows-footguns.py scripts/check-encoding-safety.py
# → no footguns, 1 file scanned, exit 0

Footprint

Implementation footprint (vs upstream/main @ d59b79f): 16 files,
+1700/-0 - script (+1120), tests (+168), 14 fixtures. No CI,
pyproject, or docs in this PR.

Platforms tested

• macOS 26.5 (Darwin 25.5.0 arm64), Python 3.11.15 - real Gate 1
  (22/22) + full-tree --all (85/786) + C-proof runtime evidence in the
  linked design issue.
• Windows - reasoned only: scripts/check-windows-footguns.py scripts/check-encoding-safety.py clean (exit 0); the lint script itself
  has no OS-specific code paths (AST/stdlib-only; pathlib + allowlist
  path fragments). Not re-run on a native Windows host in this PR.

Non-goals

• No runtime behavior change in this PR.
• No C-site fixes here (C-jobs is the separate crash-class filing
  cron: UTF-8 BOM in jobs.json crashes load_jobs (RuntimeError) #66607 / fix(cron): accept UTF-8 BOM when reading jobs.json #66609; remaining C work is follow-on).
• CI gate and docs wiring (CONTRIBUTING / scripts README) are follow-on
  ladder steps, not day one.
• Do not fail the tree on day one at the current hit count (~80).

Test plan (CI)

• pytest tests/scripts/test_check_encoding_safety.py -q (22 passed)
• Existing suites remain green on Linux CI
• No required encoding-safety CI gate added by this PR

CI note (at push, run 29627150709): all required
checks pass ('All required checks pass' = SUCCESS;
ruff, Windows footguns, all 8 test slices green).
The single red mark is the non-required 'CI timing
report' job, which failed historically on fork PRs; passed since the 2026-07-19 run
(empty AUTOFIX_BOT_PAT in the workflow env - same
residual visible on #66609 and other open fork
PRs). Not specific to this change.