v1.0.1 — research-grade hardening

[1.0.1] - 2026-06-14

Research-grade hardening after cross-model critique. No behavior change to the
four modes; this release sharpens honesty and adds manifest-validity CI.

• Manifest parse fix. SKILL.md (root and the claude-ai overlay) wraps its
  description in a >- block scalar so the embedded Triggers: colon is
  literal text. The prior unquoted value parsed as a second YAML mapping key and
  failed strict parsers and marketplace frontmatter validation.
• New scripts/check_manifests.py + CI (pre-build and post-build): parses
  every SKILL.md frontmatter, plugin.json, marketplace.json, and the
  SKILL.md inside the shipped zip. A missing root manifest, an absent zip
  member, or an unclosed frontmatter line now fails the gate.
• Honest positioning. The transfer claim is stated as the open hypothesis it
  is (grounded in the contested far-transfer literature and a moderate,
  heterogeneous meta-analytic effect), not a settled result. The README and
  modes/drill.md now state plainly that a drill key is written and self-audited
  by one model with no independent sign-off.
• Drill challenge window. A new Session Flow step: after the dissection the
  coach stops, invites a challenge to the key, and resolves it before the
  passport write, so a key it cannot defend on the merits never enters the
  longitudinal stats. Locked by the invariant lint.
• docs/evals/ evidence framework (templates only): the headline metric is
  cross-model agreement, not single-model self-scored accuracy. No empirical
  effectiveness claim until a result file exists.
• README prose normalized (em dashes removed); invariant-locked sentences kept
  verbatim.