ci: stop security-scan crashing on fork PRs; de-flake trust payload test

[prakashUXtech]

What

Two CI failures were landing on contributor PRs, and neither was caused by the contributor's code. Both are fixed here.

1. security-scan fails on PRs from forks (#295, #296)

The dangerous-pattern scan grep used eval\s*\(, which matches retrieval( (and set_retrieval_weight(, ...) all over the memory modules. So an innocent diff touching manager.py or soul.py set FOUND=1, and the step then called gh pr comment. Fork PRs run with a read-only GITHUB_TOKEN, so the comment returns GraphQL: Resource not accessible by integration (addComment), and under set -e that failed the whole security-scan job. It surfaced as a red security check on PRs that had no security issue.

Fixes:

• Word-boundary the eval / exec patterns (\beval\s*\(), so they match real eval( / obj.eval( calls but not retrieval(. Verified against a fixture.
• Guard both gh pr comment calls with a || echo "::warning::..." fallback, so a fork-token 403 degrades to a log warning instead of crashing the job. This is the same fork-token fix ci: run tests on dev PRs and stop the quality gate crashing on forks #280 applied to the check-quality job, now extended to security-scan. Real-secret detection still exits non-zero.

2. Flaky trust test (#297)

test_payload_is_hashed_not_stored asserted assert "t1" not in raw, where raw is a serialized chain that includes the base64 Ed25519 signature. The signature covers a wall-clock timestamp, so it changes every run, and roughly 2% of the time the base64 happens to contain the 2-char substring t1. That is what failed #297's run (the signature was ...xfatt1Jo...); the PR just rolled the unlucky 2%. The same suite passed on dev and on #295/#296 by luck.

Fix: use long hyphenated sentinels (do-not-leak-token-1) that cannot appear in base64 (no hyphen in the alphabet) or hex output. Ran the test 40x locally with zero failures; by construction it can no longer false-match.

Why now

Sidhant's #295/#296/#297 (which fix #281/#282/#283) were blocked by red CI that had nothing to do with their changes. This unblocks them and every future fork PR.

Testing

• YAML parses; all three security-scan shell blocks pass bash -n.
• \beval\s*\( matches eval(code) and obj.eval(s), not retrieval(q) / set_retrieval_weight(.
• The de-flaked test passes 40/40 locally.

[github-actions]

Heads up

• This PR touches sensitive files (CI/CD workflows). These paths require maintainer review via CODEOWNERS.

Please update your PR to address these points.