[1/3] Add shared structured action discovery and contracts

[George Ng (GeorgeNgMsft)]

This PR gives Copilot a shared way to discover what TypeAgent can do and look up the information needed to use a specific action: its inputs, current availability, and whether user confirmation is required. It also adds a way to detect when an action's definition has changed, so later calls can avoid using outdated instructions. This is the foundation for calling actions directly; it does not execute them yet or change how ordinary natural-language requests work.

Summary

Adds the shared discovery and contract foundation for structured action invocation from Copilot into TypeAgent. This is layer 1 of 3 in the refactor of #2973, following the canonical design.

Direct and MCP integrations will use the same dispatcher service rather than implementing their own discovery or serializing actions into command strings.

Changes

• Two-level discovery: Dispatcher.searchActions(request?) returns compact, filterable action summaries; Dispatcher.getActionContract({ schemaName, actionName }) retrieves one contract directly by exact identity. Both methods are wired through the existing dispatcher RPC client and server.
• Closed action contracts: reuse the existing schema parser, serializer, and TypeScript generator to include the selected action's required/optional parameters and transitive referenced types, including nested objects, enums, and recursive references. Unrelated sibling actions and internal parser types are not exposed.
• Versioning and scope: responses include protocol version 1, an opaque scope identifier, and a semantic contract fingerprint. Fingerprints cover execution-relevant schema and policy, excluding descriptions and transient availability. Trusted logical bindings can retain their scope across authorized reconnects; distinct sessions, bindings, and anonymous facades remain isolated.
• Conservative effect policy: optional SchemaManifest.actionPolicies metadata defaults to unknown effects requiring confirmation. Only an explicitly declared read-only action is exempt, unless its policy explicitly requires confirmation. No agents are speculatively classified.
• Honest availability and results metadata: report exact schema/action enablement and cached readiness without enabling agents, running setup, or probing readiness. Command enablement does not imply action availability. Contracts describe the standard ActionResult envelope, leave agent-specific result values explicitly unknown, and acknowledge possible runtime interactions.
• Trusted visibility: host-supplied schema restrictions are applied before parsing or disclosure. Scope identifiers are reuse boundaries, not authorization grants; owning-service authentication and resource authorization remain execution-time requirements.

Stack boundary

This PR exposes discovery only. It does not add action execution, continuation/cancellation endpoints, command-string wrappers, per-action MCP tools, or a batch/plan API. Existing natural-language behavior is unchanged.

Layer 2 adds guarded execution, stale-contract rejection, confirmation, structured results, and interaction lifecycle handling. Layer 3 adds thin Direct/MCP integrations over that same service, conversation mapping, and integration guidance.

Validation

• Dispatcher and dispatcher-RPC builds passed.
• 94 dispatcher tests passed across structured-action discovery, readiness, describe, and schema-cache suites.
• 18 dispatcher RPC tests passed.
• Committed-diff lint, complexity, circular-dependency, and test-debt ratchets passed against origin/main.
• Changed-file formatting passed.

Coverage includes exact identity resolution, dependency closure and sibling exclusion, optional parameters, recursive types, semantic fingerprint changes versus description/readiness changes, conservative confirmation defaults, visibility filtering before parsing, unavailable states without probes or effects, trusted reconnect scope reuse, and RPC forwarding/error propagation.