Retire pre-rebuild skill names from targets/ and lint against their return

[MaxGhenis]

Summary

The July catalog rebuild (#61) renamed every skill to its short directory name (skills/<name>/), but the dashboard agent family, the app multizone validator, several commands, and the multizone advisory hook still told agents to load 13 -skill-suffixed names that no longer exist. An agent following those instructions gets a failed Skill load and proceeds without the domain knowledge the file promised it.

Mapping

Each stale name was mapped only after verifying the current skill actually carries the cited content:

Pre-rebuild name	Current skill	Coverage checked
policyengine-interactive-tools-skill	policyengine-tools	multizone integration, data patterns A/B/C, embedding, path-mounted vs root-served
policyengine-frontend-builder-spec-skill	policyengine-tools	"Required stack" (Next App Router, Tailwind v4, ui-kit, bun)
policyengine-recharts-skill	policyengine-tools (+ policyengine-design for tokens)	"Charts" section: ResponsiveContainer, var(--chart-N), niceTicks
policyengine-vercel-deployment-skill	policyengine-tools	"Deployment gotchas"
policyengine-modal-deployment-skill	policyengine-tools	"Pattern C: Modal gateway + worker + polling"
policyengine-simulation-mechanics-skill	policyengine	population Simulation, economic_impact_analysis
policyengine-parameter-patterns-skill	policyengine	"Reform dictionaries and parameter paths" replaces the dead "section 6.5" citation
policyengine-us/uk/design/standards/app-skill	short names	direct renames
policyengine-api-v2-skill	policyengine-api	v1 endpoint surface + v2 alpha status

Where several old skills merged into one, load lists are deduplicated rather than left with repeated entries. Section-name citations were updated too ("Multi-zone integration (preferred)" → "Multizone integration"; "Canonical zone config — static export" → static-export zones under the same section). analyze-policy and generate-content plugin paths move off the retired category-directory layout (skills/domain-knowledge/...-skill/ → skills/<name>/). The integrator's household-payload comment now points at policyengine-api (which documents the v1 payload shape) and conditionally loads it; the planner's pattern table follows the skill's current A/B/C lettering with CSV folded into Pattern A. The writing skill's example PR body used two -skill names and was modernized.

Anti-rot

tests/test_no_stale_references.py gains three defenses so this class of rot is caught in CI instead of by an agent at runtime:

1. A FORBIDDEN entry banning [a-z0-9-]+-skill\b (the pre-rebuild naming convention) across skills/, targets/, docs/, bundles/, presets/ — escapable per line with <!-- stale-ok -->.
2. test_skill_invocations_resolve — every `Skill: <name>` instruction must name a directory in skills/ (plugin-qualified names skipped).
3. test_referenced_skills_paths_exist — every skills/... path referenced in the scanned trees (including bundle manifests) must resolve on disk, expanding the {country} template; a negative lookbehind keeps policyengine-skills/... URLs from false-matching.

Verification

• uv run pytest --ignore=tests/test_skill_examples.py: 71 passed
• Mutation check: a planted Skill: policyengine-us-skill + dead path line trips all three tests, each with a file:line message
• Wrapper smoke build (scripts/build_claude_wrapper.py) succeeds; grep of the output finds zero -skill tokens and the dashboard agents present
• snap-bbce-repeal and uk-spring-statement-2026, kept as precomputed-CSV examples, verified to exist via gh repo view

🤖 Generated with Claude Code

[MaxGhenis]

Merging under the cross-family dual-review protocol (Claude Fable 5 + GPT-5.6-sol). Sol's audit found two blockers in the original head — the --country ca flag expanding to nonexistent skills/policyengine-ca/, and replacement-skill/Modal-deploy conditions keyed on custom-backend/api-v2-alpha, values the planner never emits — plus guard-hardening notes. Fixes applied (2e44a2b), and sol's re-verify caught one remaining behavior-gating custom-backend in the planner's no-variables fallback, fixed verbatim (a8f6f2b); targets/ now greps clean of both stale values. Guards mutation-tested (planted typo-invocation, unknown template, and case-variant -Skill all trip; clean tree 31/31). Also rebased over #77/#79's microsim-guard work — union in test_no_stale_references.py, full suite 99/99.