tech-debt(redaction): llm_shared credential_redaction.redact_dict is a second recursive redactor that partially reveals keys #18205

Summary

There are two recursive mapping redactors. #18193 added autobot_shared/secret_redaction.py::redact_nested (the canonical module, BROAD policy, value-aware, keeps every non-secret field byte-identical). The older autobot-backend/llm_shared/credential_redaction.py::redact_dict also recurses, but differs in two ways that matter:

• it also runs redact_string over non-credential strings, so non-secret values can change;
• it partially reveals credential strings via redact_api_key (value[:4]...value[-4:]), where the canonical module uses a fixed-width placeholder.

redact_nested could not reuse redact_dict for GET /api/llm/config because of both differences. That leaves one concept with two implementations and different masking, which is what #17336 set out to remove.

Acceptance criteria

• Every redact_dict caller is listed with what it needs: does it rely on content-scanning non-credential strings, or on the partial reveal?
• redact_dict becomes a thin composition over redact_nested (plus redact_text content scanning where a caller genuinely needs it), or is removed with zero-caller evidence, so one recursive mapping redactor remains.
• The partial reveal (value[:4]...value[-4:]) is either justified per caller in REDACTION_BOUNDARY.md or replaced by the fixed-width placeholder. Revealing 8 characters of a key is a disclosure, so this needs a stated reason.
• The masked set at each caller does not shrink, proven with the tech-debt(redaction): two modules in autobot_shared both declare themselves the canonical redactor, and they disagree on 7 of 10 field names #17336 per-caller corpus method.

Related: #17336, #18193.