Compile capability-aware host component bundles

[ScriptedAlchemy]

Summary

Evolve the existing target adapter registry beyond skills, hooks, MCP documents,
and opaque payloads. Model host-neutral plugin components—skills, commands, agents, rules,
event routes, MCP servers, LSP/native diagnostics, Apps, and native extensions—
then lower only genuinely supported surfaces into each host.

Problem

TraceDecay ships materially different integrations for Claude, Cursor, Codex,
Kimi, OpenCode, Gemini, Copilot, Kiro, Cline, Roo, and Kilo. Some support hooks,
some only MCP; one uses an LSP document, another a native VS Code extension;
commands and agents intentionally differ by host.

Agent Bundle currently has pinned adapters for Claude, Cursor, Codex, and a
portable profile, with Boolean capability checks centered on skills, hooks,
and MCP. Opaque prebuilt payloads preserve bytes but cannot validate component
meaning, ownership, capability requirements, or host-specific omissions.

Canonical component model

type AgentComponentKind =
  | 'skill'
  | 'command'
  | 'agent'
  | 'rule'
  | 'event-route'
  | 'mcp'
  | 'mcp-app'
  | 'lsp'
  | 'native-diagnostics'
  | 'native-extension';

Projects use conventional roots such as src/commands, src/agents,
src/rules, and src/extensions, while existing skills, MCP, Apps, and
events feed the same semantic manifest.

Authors may group components into installable feature sets, but paths and
ownership remain compiler-derived. A component cannot silently write another
component's artifact.

Capability catalog

Target adapters publish versioned capability records:

type TargetCapabilityState =
  | { state: 'supported'; evidence: string }
  | { state: 'degraded'; reason: string; evidence?: string }
  | { state: 'unavailable'; reason: string }
  | { state: 'prohibited'; reason: string };

Capability evidence points to a pinned schema, official host contract, or
checked-in native fixture. Similar branding or payload shape is not evidence.

Projection

The compiler generates a host component plan containing:

• selected semantic components and target restrictions;
• native paths, schemas, manifests, and command shapes;
• host-specific omissions with reasons;
• content and effective-behavior digests;
• source and host-contract provenance;
• protocol and runtime requirements;
• the lifecycle ownership class for every artifact.

Adapters, not application code, own native manifests and shared-config merge
formats.

Acceptance criteria

• One host-neutral skill/event/MCP definition lowers into different valid
  Claude, Cursor, and Codex artifacts.
• A target may intentionally omit commands or agents without appearing broken.
• LSP and native-extension components are represented explicitly rather than
  hidden in opaque payloads.
• Unsupported components fail before artifact publication.
• Every supported capability has pinned evidence and conformance fixtures.
• Inspect output explains selected, degraded, unavailable, and omitted
  components per target.

Design references

• Current Agent Bundle target adapter contract
• Current Agent Bundle adapter registry
• TraceDecay canonical host-bundle model
• TraceDecay host-bundle registry
• TraceDecay capability admission
• Cursor project rules
• Cursor project commands
• VS Code language-server extension model

Stack position

• Depends on: #93 filesystem/compiler IR and its shared capability-state contract.
• The event-route component family plugs in after #97 semantic events land;
  skill, command, agent, rule, MCP, LSP, and native-extension compilation do not
  depend on event routes.
• Unlocks: #101 managed host-bundle lifecycle and expanded local discovery in #105 Workbench.
• Existing opaque payloads remain an escape hatch, not a semantic success
  claim.

Full meta-framework stack

• Foundations: #93 routes - #94 identity - #95 context - #96 renderer
• Runtime: #97 events - #98 state - #99 notices
• Host: #100 components - #101 lifecycle
• Surfaces and proof: #102 CLI/scripts - #103 Rstest - #104 proximity example - #105 Workbench