feat(agent): complete middleware pipeline with config-driven ordering

[buchenberg]

Summary

Completes the middleware pipeline plan (Phases 2-3), removes the legacy code path, and adds comprehensive documentation.

Phase 2 — Config-driven middleware

• MiddlewareConfig with enabled/disabled lists in config.yaml
• builtinMiddleware registry maps canonical names to constructors via MiddlewareBuilder func(l *Loop) Middleware
• resolvedPipelineNames() applies config overrides to default set
• Scaffold config updated in create.go

Phase 3 — New middleware (4 files)

Middleware	File	Hook	What it does
PermissionMiddleware	middleware_permission.go	PostModel	Path-glob allow/deny tool filtering
ToolConcurrencyMiddleware	middleware_toolconcurrency.go	—	Semaphore-capped goroutine pool via Loop.toolSem
SubAgentMiddleware	middleware_subagent.go	PostModel	Task depth cap enforcement
PromptCachingMiddleware	middleware_promptcaching.go	PrepareStep	Anthropic cache-control on system + tool msgs

Phase 3e — Compaction enhancements

• CompactionMiddleware takes configurable threshold (default 0.8)
• compactContext() signature changed to accept threshold float64
• Preflight (PrepareStep) + post-iteration (PostTool) hooks

Cleanup

• Removed legacy code path: runLegacy() (115 lines), executeToolsParallel() (97 lines), denyTool() (23 lines)
• Removed AgentMode field from Loop — single unified middleware pipeline
• Added CacheControl to types.Message (omitempty, no-op for non-Anthropic)

Docs

• docs/architecture.md: full architecture reference covering agent loop, middleware pipeline, tool execution, streaming, compaction, hook events, provider interface
• README.md: new Development section (build, test, lint, cross-compile, repo layout, architecture link), updated config docs
• .kilo/agent-loop-comparison.md: implementation status for all 10 recommendations

Verification

gofmt -l .       → empty
go vet ./...     → clean
go test ./...    → all pass
yaah doctor      → OK
yaah "say hi"   → agent loop executes, streams response