v0.4.0

21 commits since v0.3.0. Two full protocol capabilities go from planned to shipped and
field-validated — trip-wire / institutional memory and the ult-cep-retrofit metaskill —
exercised together end-to-end against a real, unrelated 62-unit skill library. Also closes
out runtime field-validation (Cursor is the 4th and last supported runtime confirmed
working end-to-end), ships an installable Claude Code plugin package, and fixes a real
Windows PowerShell CLI-ergonomics bug found via a cross-runtime (GitHub Copilot) re-run of
the retrofit experiment.

Added

• Trip-wire, implemented — ult-institutional-memory-distill ships a real, tested
  decision-ledger mechanism (add-entry, alias, advance-cursor, reject-source,
  query, disposition, show; 33 tests) integrated into ult-context-generate's Step
  7.7. Every revise/escalate hit requires explicit human disposition before the
  package's approval gate closes — never auto-applied, never auto-suppressed. See
  PROTOCOL.md §7.
• ult-cep-retrofit — a new metaskill that brings an existing third-party skill library
  under this protocol: deterministic inventory across both flat-file and skill-directory
  conventions, per-unit code/task classification, a recommend() step for contract fit, and
  idempotent frontmatter-pointer insertion — without vendoring or rewriting the third-party
  library. 32 tests (2 skipped, Windows symlink privilege). See
  PROTOCOL.md §8.
• New case study: cep-retrofit-superpowers
  (obra/superpowers, MIT, 62 units, pristine clone) — 0 misclassifications, a deep
  with/without-CEP comparison of writing-plans's generated output, and trip-wire exercised
  on top. Real defects found in the no-CEP baseline that the retrofitted run avoided (a
  duplicate AVP dictionary declaration; a fabricated test target), plus a humbling finding
  about tier: revise hits' required_evidence field being load-bearing, not decorative.
  First case to close the Trip-wire and Metaskill-retrofit origin columns in the
  case-study feature-coverage table.
• Claude Code plugin package — claude-plugin/ packages all 7 real skills, generated
  from .github/skills/ so it can't drift; --check runs in CI on every push. Verified
  with a real local one-click install, not just claude plugin validate.
• Cursor field-validated (closes ROADMAP.md item 5 — 4/4 supported runtimes now
  confirmed working end-to-end). See
  issue #35.
• EVIDENCE.md — a headline-first, shareable Evidence page leading with the
  cross-domain hallucination-recurrence finding (26 real citations vs. 0, the same
  fabricated concept recurring identically in two unrelated codebases).

Fixed

• Installer no longer leaks __pycache__/.pytest_cache into installed projects
  (PR #37).
• CONSUMING-CONTEXT-PACKAGE.md now documents how to consume institutional_memory_hits[]
  — found and fixed during the case study's trip-wire rung, before it could produce a
  silently-ignored field in a real consuming skill.
• recommend --description silently mangled on Windows PowerShell for descriptions
  containing embedded double-quotes, producing a false "neither" classification with no
  error surfaced. Found via a GitHub Copilot cross-runtime re-run of the mattpocock/skills
  retrofit. Fixed by adding recommend --description-file <path> as a Windows-safe
  alternative, with 4 new regression tests.

Removed

• A second ult-cep-retrofit case study against mattpocock/skills, run alongside
  cep-retrofit-superpowers but deliberately not kept — the Copilot re-run that followed it
  surfaced the real CLI bug fixed above, which was worth keeping; the write-up itself
  wasn't. This release ships one metaskill-retrofit case study, not two.

Full changelog: CHANGELOG.md §0.4.0