chore(work): decompose 2026-05-14-library-evaluation-policy epic

Decomposer produced 9 stories under docs/work/2026-05-14-library-
evaluation-policy/, ordered to land the schema foundation first and
the cross-referencing content (skill, guide, backfill) last:

01 - trace schema module + docs/library-decisions/ foundation
02 - pre-commit check script
03 - Claude PreToolUse / PostToolUse hooks
04 - evaluate-library skill
05 - human guide (docs/guides/adding-a-library.md)
06 - sandcastle reviewer prompt update
07 - generator pre-shipped traces for optional cores
08 - backfill traces for existing runtime deps
09 - CLAUDE.md Key Conventions bullet

Anchored by ADR-022 + the approved PRD at
docs/work/prds/2026-05-14-library-evaluation-policy.prd.md.

docs/work/2026-05-14-library-evaluation-policy/03-claude-hooks/_story.md

@@ -0,0 +1,42 @@
+---
+id: 03-claude-hooks
+epic: 2026-05-14-library-evaluation-policy
+title: Claude PreToolUse / PostToolUse hooks for library-policy nudge
+type: technical-story
+status: todo
+feature: tooling
+depends-on: []
+blocks: []
+---
+
+## Goal
+
+Write `.claude/hooks/library-policy-nudge.sh` — a single hook script that dispatches on `tool_use_type` to handle both `PreToolUse` (Bash invocations matching `pnpm add` / `pnpm i <pkg>`) and `PostToolUse` (Edit/Write on any `**/package.json`). On match, emit a non-blocking system-reminder pointing the agent at the `evaluate-library` skill with the exact invocation pattern. Register the hook in `.claude/settings.json` following the same pattern as `generator-first-nudge.sh`.
+
+## Why
+
+Agents (and developers) running `pnpm add` by reflex bypass the policy before the pre-commit gate fires. The PreToolUse hook injects the skill reminder _before_ the install runs — the cheapest possible intervention point. The PostToolUse hook catches the rarer case where an agent edits `package.json` directly without running the install command.
+
+## Done when
+
+- `.claude/hooks/library-policy-nudge.sh` exists, is executable, dispatches on `CLAUDE_TOOL_USE_TYPE` (or equivalent hook env var), and emits the reminder to stdout when a `pnpm add`/`pnpm i <pkg>` bash command is detected or when an Edit/Write tool targets a path matching `**/package.json`.
+- The hook is registered in `.claude/settings.json` (or `.claude/settings.local.json` if per-repo convention) under both `PreToolUse` and `PostToolUse` event types, matching the registration pattern of `generator-first-nudge.sh`.
+- The reminder text includes the literal string `/evaluate-library` and the argument template `<name> --tier <feature|core|app> --target <package-path>` so the agent sees the exact invocation.
+- Bash smoke tests (same style as any existing hook tests): pipe a mocked Claude hook payload `{ "tool_input": { "command": "pnpm add foo" } }` into the script and assert stdout contains the skill-reminder marker; pipe a payload for `pnpm add -D foo` (devdep) and assert no reminder emitted; pipe an Edit payload on `src/feature.manifest.ts` and assert no reminder emitted.
+- `pnpm lint && pnpm fallow:audit` pass.
+
+## In scope
+
+- `.claude/hooks/library-policy-nudge.sh` — the hook script.
+- `.claude/settings.json` (or equivalent) — hook registration lines.
+- Bash smoke tests for the hook script.
+
+## Out of scope
+
+- Hook auto-deny or blocking behavior — the hook is advisory only (non-blocking system-reminder).
+- Hooks for `npm install` / `yarn add` — the repo is pnpm-only.
+- Integration with the skill implementation (Story 04) — the hook emits a text reminder; the skill is a separate file.
+
+## Tasks
+
+- [ ] Investigate how `generator-first-nudge.sh` is registered (read `.claude/settings.json` and the existing hook script) to confirm the env-var names, payload shape, and registration keys; then write `.claude/hooks/library-policy-nudge.sh` + register it in the settings file + write bash smoke tests covering `pnpm add <pkg>` (reminder), `pnpm add -D <pkg>` (no reminder), Edit on non-package.json (no reminder), Edit on `package.json` (reminder); all gates pass on this single commit.