Restructure CLAUDE.md into modular rules + add pr-shepherd command (#750)

* refactor: restructure CLAUDE.md into modular rules and add pr-shepherd command

Trim CLAUDE.md from 710 lines to 92 by moving detailed guidance into
path-scoped `.claude/rules/` files that load on demand. Add a new
`/pr-shepherd` command that consolidates the full PR lifecycle
(review, fix, quality gate, CI fix loop, merge) into one workflow.

Changes:
- CLAUDE.md: keep only essentials (build commands, code style, architecture,
  module specs, config reference, debugging)
- .claude/rules/review-discipline.md: 15+ review rules, scoped to src/**/*.rs
- .claude/rules/database.md: dual-backend rules with SQL dialect translation
  table, scoped to src/db/** and migrations/**
- .claude/rules/safety-and-sandbox.md: safety layer and sandbox rules, scoped
  to src/safety/**, src/sandbox/**, src/secrets/**
- .claude/rules/testing.md: test tiers and patterns, scoped to src/** and tests/**
- .claude/rules/tools.md: tool architecture and implementation pattern, scoped
  to src/tools/** and tools-src/**
- .claude/commands/pr-shepherd.md: 7-phase PR lifecycle command that subsumes
  review-pr, respond-pr, ship, and manual CI fix loops

[skip-regression-check]

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address review feedback on CLAUDE.md restructure

- Restore project structure tree in CLAUDE.md (zmanian blocking)
- Create .claude/rules/skills.md with trust model, SKILL.md format,
  selection pipeline, and skill tools (zmanian blocking)
- Restore configuration section with key env vars (zmanian medium)
- Restore "Adding a New Channel" guide (zmanian medium)
- Add heartbeat mention to Workspace & Memory section (zmanian low)
- Fix pr-shepherd: replace `git add -A` with specific file staging (zmanian)
- Fix pr-shepherd: ask user for merge strategy instead of hardcoding --squash (zmanian)
- Fix pr-shepherd: replace `--watch` with polling + 10min timeout (zmanian)
- Fix testing.md: "skipped if DB is unreachable" not "expected to fail" (Copilot)

[skip-regression-check]

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address review comments on PR #750

- Narrow `crate::` import rule: `super::` is fine in tests and intra-module refs
- Fix capabilities file naming: `<name>.capabilities.json` sidecar, not bare `capabilities.json`
- Update mechanical verification checklist to match narrowed import rule

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* refactor: move Bedrock docs from CLAUDE.md to src/llm/CLAUDE.md

Bedrock provider details (auth, config, feature flag) belong in the
LLM module spec, not the top-level guide. Added file map entry,
provider table row, and dedicated section in src/llm/CLAUDE.md.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* refactor: move env var config block out of CLAUDE.md

Replace 20-line config block with one-liner pointing to .env.example
and src/llm/CLAUDE.md. Config details are only needed during deployment,
not everyday coding.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: use gh pr checkout for fork-safe PR checkout in pr-shepherd

Replaces git fetch/checkout with gh pr checkout {number} which
handles both same-repo and fork-based PRs automatically.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address Copilot review round 5 on PR #750

- Add gh pr list and gh pr checkout to pr-shepherd allowed-tools
- Align crate:: import rule in pr-shepherd with updated CLAUDE.md guidance
- Fix vector type in database.md: BLOB (flexible dims), not F32_BLOB(1536)
- Update MCP limitation: stdio/HTTP/Unix transports exist, no streaming

Co-Authored-By: Claude Opus 4.6 <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>

.claude/rules/skills.md

@@ -0,0 +1,56 @@
+---
+paths:
+  - "src/skills/**"
+  - "skills/**"
+---
+# Skills System
+
+SKILL.md files extend the agent's prompt with domain-specific instructions. Each skill is a YAML frontmatter block (metadata, activation criteria, required tools) followed by a markdown body injected into the LLM context.
+
+## Trust Model
+
+| Trust Level | Source | Tool Access |
+|-------------|--------|-------------|
+| **Trusted** | User-placed in `~/.ironclaw/skills/` or workspace `skills/` | All tools available to the agent |
+| **Installed** | Downloaded from ClawHub registry (`~/.ironclaw/installed_skills/`) | Read-only tools only (no shell, file write, HTTP) |
+
+## SKILL.md Format
+
+```yaml
+---
+name: my-skill
+version: 0.1.0
+description: Does something useful
+activation:
+  patterns:
+    - "deploy to.*production"
+  keywords:
+    - "deployment"
+  exclude_keywords:
+    - "rollback"
+  tags:
+    - "devops"
+  max_context_tokens: 2000
+metadata:
+  openclaw:
+    requires:
+      bins: [docker, kubectl]
+      env: [KUBECONFIG]
+---
+
+# Skill instructions here...
+```
+
+## Selection Pipeline
+
+1. **Gating** -- Check binary/env/config requirements; skip skills whose prerequisites are missing
+2. **Scoring** -- Deterministic scoring: keywords (10/5 pts, cap 30) + patterns (20 pts, cap 40) + tags (3 pts, cap 15). `exclude_keywords` veto (score = 0 if any present)
+3. **Budget** -- Select top-scoring skills within `SKILLS_MAX_TOKENS` prompt budget
+4. **Attenuation** -- Minimum trust across active skills determines tool ceiling; installed skills lose dangerous tools
+
+## Skill Tools
+
+- `skill_list` -- List all discovered skills with trust level and status
+- `skill_search` -- Search ClawHub registry for available skills
+- `skill_install` -- Download and install a skill from ClawHub
+- `skill_remove` -- Remove an installed skill