v0.32.0

Continued the bundled blueprint design-chain skill review: closed a second
5-issue batch (#92–#96) hardening the coherence guard, the self-check, and the
template/prompt discipline, each fix shipped with tests. The skill manifest moved
0.9.0 → 0.10.0.

blueprint skill

• Recompute the quality-gate self-check against the final body: a deterministic
  vendor_leak scan (named products, vendor CLIs, wire-level config flags) now
  re-derives the implementation-neutrality verdict on every revision, so a leak
  reintroduced by a later edit cannot survive behind a stale PASS (#92).
• Strengthen the coherence guard — anchor coverage is enforced (MVP-staging prose
  with no anchors, or an Open Questions section with no stage=open anchor, now
  FAILs), a consumes= signal edge catches signal inversions, and
  orphan_reference catches a consumed signal whose producer is never registered
  (#93).
• Add an altitude ceiling for agent/tool-surface sections: WARNING-tier
  mechanism_altitude (dedup keys, compaction, single-PEP, retry-bounding,
  transport outside the §18 handoff) and tool_identifier_altitude (a named tool
  in a policy/eval/MVP row that should bind to a READ/ACT/AUTH class) (#94).
• Enforce actor-channel completeness: §3 tags every actor's channel class and §8
  defines every escalation/authorization path per channel, with a new Gate 10
  that FAILs an escalation assuming an inline human on a headless channel (#95).
• Template discipline: one authoritative home per decision (Appendix B, keyed by
  a stable ID) and a conservative MVP-0 default that keeps an unvalidated,
  non-research-derived capability out of the smallest core behind a cheap gating
  test (#96).