Report readback: reject '..' on rendered reports dir + reject non-UTF8 reports

[jgruberf5]

Deferred nits from the #468 review (report readback surface). PR #468 fixed the hardlink-into-reports leak (fd-based read, O_NOFOLLOW + fstat, reject st_nlink > 1) and closed the TOCTOU window. These are the remaining minors.

Context: backend/services/module_reports_service.py.

Scope:

• _resolve_reports_dir relies on realpath containment but does not also reject .. on the rendered value the way the manifest-time validator does — an asymmetry (not a known bypass, since realpath containment still holds). Mirror the manifest-time .. rejection for defense-in-depth consistency.
• A non-UTF8 report file is currently served as errors="replace" mojibake. Consider rejecting non-UTF8 content with a clear error instead of surfacing garbled text.

Source: PR #468 review (JLCode-tech), "Reports/action backend — approve-with-nits".

---

_{Migrated from sp-prod-field/bnk-forge #470 (opened 2026-07-18; original labels: enhancement). That repository is archived and read-only.
Bare #NNN references in the text above refer to issues and PRs in the original repository, not to numbering here.}

[jgruberf5]

Analysis: fixed. Both items are addressed in the container-runner hardening PR.

Worth recording why the .. check was reachable at all: reports.dir renders from an input, so a value that is clean in the manifest need not stay clean after templating. realpath containment still held, so this is defence in depth against a future refactor rather than a known bypass.

The non-UTF8 case now raises a clear error instead of serving errors="replace" mojibake, which read as a corrupt report rather than as "this is not text".

[jgruberf5]

Assessment: still valid; a fix exists but is not yet in the code.

Both items are addressed in #2 (container runner hardening), open against staging and not merged. Stays open until that lands.

Tagging verified — confirmed present on the current staging.

[jgruberf5]

Closed by #2 (container runner hardening). Both items addressed: the rendered-value .. rejection and the non-UTF8 refusal, each verified by reverting the fix and confirming the test fails.