fix: harden audit evidence and evaluation

[solivane-bluedot]

Summary

Hardens DarkPrompt's audit-evidence, evaluation, media, regex, provider-error, and reproducibility boundaries.

• makes raw prompts, responses, traces, metadata, tool calls, and artifacts opt-in via --include-raw-evidence
• preserves structural report keys while recursively redacting retained values
• fails closed on incomplete, skipped, errored, inconclusive, or missing evaluations unless --allow-incomplete is explicit
• rejects mixed refusal/actionable-compliance responses without penalizing benign fenced context
• confines media reads to explicit roots and validates/reads through the same no-follow descriptor
• sanitizes credential-bearing provider failures
• bounds regex count, pattern length, input size, and execution time
• propagates generated media roots through adaptive OCR variants
• commits uv.lock, uses locked CI installs, and corrects repository/security metadata

Validation

• default Python: 70 passed
• Python 3.9: 70 passed
• Ruff: passed
• uv lock --check: passed
• git diff --check: passed
• wheel and sdist builds: passed
• dependency audit: zero known vulnerabilities
• no real provider credentials or live provider calls used

Current-head proof

• base main: 65c1b75df171cae3f6b2bfd21382709fdef62f7e
• remote branch tip: 081f85be1b8302bbe76fee11a4b017b41b0a3c63
• local branch tip: 081f85be1b8302bbe76fee11a4b017b41b0a3c63
• Python 3.9 compatibility fix parent: 037808d4663ae3fe32276723e8a8c269a0ed0e50

Independent review and hosted PR gates remain required before merge.

Rollback

Revert the merge commit (or both branch commits if merged without squash). Before merge, reset/delete solivane/darkprompt-security-completion; the base branch is unchanged.