v0.5.0

• add: contextualizer selection for the locator skills — /skill-engine:{discover,refresh,status,self-audit,new-reference} <name> names which contextualizer to operate on, unblocking multi-contextualizer installs that previously hard-stopped on ambiguity (plugin/skill-engine/skills/*/SKILL.md).
• add: an error run outcome in the eval harness for CLI-invocation failures, excluded from the pass-vs-fail vote in both the renderer and the HTML viewer so an expired token reads as runner trouble rather than a routing regression; the harness exits 70 when every run errored, and pass detection is scoped to Read tool_use so catalog text flowing through a transcript cannot false-pass (plugin/skill-engine/engine-bootstrap-templates/eval/).
• add: a live-tree gate in apply — promotion halts when a live file's hash no longer matches the manifest's sha_before, so a stale proposal cannot silently revert edits that landed after staging, such as approved self-audit fixes (plugin/skill-engine/skills/apply/SKILL.md).
• add: make ci-local (scripts/ci-local.sh) as the single validator entry point shared by CI, the Makefile, and /release — one check inventory with no transcribed job lists to drift; in CI a missing check-jsonschema fails loud instead of green-skipping the schema gates (scripts/ci-local.sh, .github/workflows/{lint,security}.yml, Makefile).
• change: the clone cache is unified on the kind-partitioned layout everywhere it is taught, with the doctrine lint extended from skills to docs and the README so the flat path cannot reappear unmarked (plugin/skill-engine/docs/, plugin/skill-engine/tests/doctrine.sh).
• change: stamped *.sh.template files are now covered by shellcheck and the doctrine curl lint; the curl allowlist is comment-safe and accepts -I anywhere in a flag cluster, and the source-paths schema is tested against verify.sh for equivalence via negative fixtures (plugin/skill-engine/tests/doctrine.sh, scripts/ci-local.sh).
• change: CI toolchain versions are pinned for reproducible runs (.github/workflows/).
• fix: the shared locator block no longer dies silently under pipefail when no contextualizer matches — the zero-match diagnostics ("Run /skill-engine:engine-bootstrap first.") actually print (plugin/skill-engine/skills/{discover,refresh,status,self-audit,new-reference,using-skill-engine}/SKILL.md).
• fix: eval-pipeline robustness — an interrupted harness run exits 130 instead of publishing a corrupt results file, pass detection cannot record a SIGPIPE false fail on Read-heavy transcripts, and the renderer handles any runs-per-query count instead of silently dropping runs past the third (plugin/skill-engine/engine-bootstrap-templates/eval/).
• fix: the engine-bootstrap overwrite guard triggers on any existing contextualizer files, not only a parseable state marker, so a corrupt marker cannot bypass it (plugin/skill-engine/skills/engine-bootstrap/SKILL.md).
• fix: docs reconciled to shipped behavior — the quickstart gained the review/apply promotion step required since the v0.3.0 staging gate, 12-evaluation.md documents the error outcome and the jq requirement, issue templates and broken links repaired, and prose counts swept (plugin/skill-engine/docs/, .github/ISSUE_TEMPLATE/).
• remove: the stale worker-agent stratum (agents/worker-{refresh,verify}.md) retired in favor of goal-given delegation, with the templates README rewritten to match what actually ships (plugin/skill-engine/agents/, plugin/skill-engine/engine-bootstrap-templates/README.md).