Motivation

The agent-facing instructions had drifted from the code (nonexistent APIs in skills, an incomplete architecture map) and the root CLAUDE.md had grown past 4k words of always-loaded context, much of it deep per-feature reference.

Changes

• Root CLAUDE.md goes from 4,159 to 3,082 words (-26%): ~1,650 words of deep reference moved into nested files (loaded only when working in those packages) while ~570 words of new conventions were added
• Restructure CLAUDE.md: complete the internal/ package map and command groups, correct the test-integration and mock-generation claims (mockgen via go:generate), and add sections for commit/PR/Linear conventions, the release process, single-test/lint/JUnit workflows, and openspec
• Move deep reference sections (snapshots, Azure setup/interception, volume mounts, GATEWAY_LISTEN + offline fallbacks) into nested CLAUDE.md files under internal/{snapshot,azureconfig,config,container}/, each with an AGENTS.md symlink mirroring the root convention
• Refresh skills: fix add-command (the sink.Emit(output.XxxEvent{...}) pattern instead of nonexistent output.EmitXxx helpers, initConfig(nil), Cobra group registration, requireSubcommand, isInteractiveMode, testEnvWithHome), fix the review-pr events checklist (sealedEvent() marker, no emit helper), and update create-pr (current Linear MCP tool name, Linear-first ticket step)
• Make test-integration depend on build (one-line Makefile change) so the binary is always rebuilt before integration tests — this is what makes the updated build docs accurate; previously a stale bin/lstk could silently run after source edits

Closes PRO-365