skills: add per-skill agents: and related: frontmatter contract

[t4sh]

Closes #6

Adds two optional top-level frontmatter fields to SKILL.md (sibling to `name:` / `description:` / `version:`), and populates them on the 31 skills erphq/lab-sites currently composes.

What lands

Spec (in README.md)

New "Frontmatter — `agents:` and `related:` (optional, downstream-facing)" subsection under "For agents" defining:

```yaml

name: accounts-receivable
description: …
version: 1.0.0
agents: # ← optional

• collections
• reconciliation
• approvals
  related: # ← optional
• accounts-payable
• general-ledger
• period-close
  metadata:
  …

---

```

• `agents:` — agent capability slugs that work on this skill (reusable across apps). Renderers display as cards/chips.
• `related:` — other skill slugs this skill commonly works with. Slugs MUST resolve to real `SKILL.md` files in this repo. Renderers use for "works with" cross-link grids.
• Both optional — skills without populated fields render the standard page sans those sections.
• Top-level (not nested under `metadata:`) because they describe skill relationships, not file metadata.

Data (31 SKILL.md files updated)

Populated agents+related on every SKILL.md at the `03-org-1k-plus` tier across the 5 departments lab-sites composes (8 finance + 8 supply-chain + 2 HR + 8 sales-crm + 5 IT-plumbing under `information-technology/`).

Source: lab-sites' `apps/web-lab/src/_data/erpFeatures.json` + `crmFeatures.json`. For skills composed by multiple apps (e.g. `security-roles` shows up in both ERP and CRM), values are the union across composing apps. Sorted for deterministic diffs.

Other ~50 SKILL.md files

Untouched. Optional means optional — maintainers backfill at other tiers / depts as time permits. No drift introduced; consumers gracefully render missing fields.

Why this is the right shape

• Canonical — single source of truth for skill relationships, replacing per-consumer hand-curation
• Drift-validatable — existing SKILL.md drift checks now cover this data automatically
• Forward-compatible — both fields are optional; new SKILL.md files don't need to populate them on day one
• Cross-app — `agents:` from a skill composed by 5 different apps still resolves correctly per app, since slugs are globally addressable

Downstream impact

Lab-sites can drop the `agents` / `related` columns from its `erpFeatures.json` / `crmFeatures.json` after this lands and read them from `skillSpecs.mjs` (which already loads SKILL.md frontmatter). Cleanup PR will follow on the lab-sites side.

Test plan

• All 31 modified SKILL.md files parse as valid YAML frontmatter
• Spot-check 5 random skills — `agents:` and `related:` insert in the right position (between `version:` and `metadata:`), with deterministic alphabetical ordering
• Lab-sites' `pnpm verify` (which runs `check-skills-drift.mjs`) still passes when the corresponding mirrors are pulled
• All `related:` slugs resolve to real SKILL.md files in this repo (no orphan refs introduced)

Spec aware

Surfaced during work on erphq/lab-sites#118, which built the appskills+skills→lab-sites pipeline and noted this as the highest-value upstream contract gap.