privacy: redact the capability-lock files and redact before truncating in the n8n auth-risk reader #1079

Deferred by the engineering-review implementation loop from #1077 (issue #1026).

Problem

Two redaction gaps outside #1026's scope:

1. The capability-lock family (capabilities.lock.json, both capability-lock-diff files and capability-delta-attestation.json) publishes tool names unredacted, including the older secret shapes, because src/agents_shipgate/core/capability_lock.py never calls the privacy module. fix(checks): publish secret and injection findings, widen redaction #1077's leak tests had to exclude these files by name.
2. src/agents_shipgate/inputs/n8n/_auth_risk.py:173 cuts text to 200 characters before redacting it, so a credential that straddles the cut is partly published.

Proposed fix

(1) Route every string the lock family publishes through redact_text / redact_data at the point where the lock is assembled, and extend the #1077 leak tests to include these files (remove the by-name exclusion). Because this can move source_set_hash for adopters whose tool names contain a secret shape, add a STABILITY migration note. (2) Redact first, then truncate, in _auth_risk.py, with a test that plants a credential across the 200-character boundary.

Acceptance criteria

• The leak tests cover the capability-lock files and pass.
• A credential straddling the n8n truncation boundary is fully redacted.

Source

Engineering review 2026-10-12 follow-up (PR #1077 implementer notes). Tracked in #1060.