feat: align TypeScript SDK with AgentScope Python incrementally

## Context

The initial implementation was submitted as #48 and covered the complete Python parity effort in one pull request. That made the change difficult to review, so #48 is being closed while its source branch remains available as a read-only implementation snapshot:

- Snapshot branch: `rayrayraykk:feat/parity-foundation`
- Snapshot commit: `ad8f0b57d225a4e2570e6011ac2e21dc58aa7ba8`
- Python behavior baseline: `agentscope-ai/agentscope@61cdeae4ffe63182f3343229aa6fbd868e30b0c5`

This issue tracks the same scope as a sequence of small, dependency-ordered pull requests.

## Goals

- Match the public behavior, wire contracts, edge cases, and lifecycle semantics of the pinned AgentScope Python baseline.
- Keep TypeScript-native architecture: discriminated unions, generics, Zod validation, `AsyncIterable`, `AbortSignal`, Web APIs, and conditional exports.
- Preserve existing TypeScript import compatibility.
- Keep each PR focused on one reviewable capability and include its tests and documentation.
- Re-run the parity audit against Python `main` before the final PR.

## Delivery plan

PRs will be submitted and merged in dependency order. A later PR will not be opened against upstream `main` until its prerequisite has merged, so reviewers never see an accumulated stack disguised as a small PR.

### Phase 0 — review infrastructure

- [x] Package export-cycle cleanup and packed ESM/CJS/CLI smoke tests: #50.
- [x] Reproducible Python parity manifest, pinned checkout validation, and mapping-tool tests: #51.

### Phase 1 — Core contracts

- [ ] Types, exceptions, utilities, logging, IDs, and timestamps.
- [ ] Message and event wire contracts.
- [ ] Agent state and legacy migration.
- [ ] Permission evaluation.

### Phase 2 — Tools, models, and agents

- [ ] Tool runtime, middleware, groups, and backend contracts.
- [ ] Built-in shell/file/task tools.
- [ ] Skill loading and MCP transports.
- [ ] Model cards, response/usage contracts, and streaming accumulation.
- [ ] Formatters and provider adapters.
- [ ] Embedding and TTS.
- [ ] Agent state machine, interruption, structured output, and middleware.
- [ ] Tracing, context compression, and memory middleware.

### Phase 3 — Workspace and RAG

- [ ] Local workspace and sandbox gateway.
- [ ] Bubblewrap and Docker adapters.
- [ ] Apple Container, E2B, Daytona, Kubernetes, and OpenSandbox adapters as separate PRs.
- [ ] Documents, chunkers, and parsers.
- [ ] Vector-store adapters.
- [ ] Knowledge base and RAG middleware.
- [ ] Pipeline and console.

### Phase 4 — AgentScope Service

- [ ] Records and storage contracts.
- [ ] Redis and SQL storage adapters.
- [ ] Access policy and message bus.
- [ ] Local and remote workspace managers.
- [ ] Runtime managers, scheduler, sessions, and background services.
- [ ] Knowledge indexing, chat toolkit, and team runtime.
- [ ] Application composition, SSE, AG-UI, and registry hubs.
- [ ] Channel runtime and individual platform adapters.
- [ ] HTTP schemas and routes.

### Phase 5 — Desktop and final audit

- [ ] Desktop migration to the shared Core/Service composition root.
- [ ] Latest Python delta, cross-platform fixes, documentation, and final parity audit.

## How maintainers can review

1. Review only the current PR; the snapshot branch is reference material and is not the review target.
2. Follow the dependency link at the top of each PR. Every PR will state its prerequisite and the next planned unit.
3. Start with production code, then inspect the colocated parity tests and Python source/test mapping.
4. Generated manifests and model-card data will be isolated from handwritten runtime changes whenever possible.
5. Use the exact verification commands in each PR. Live provider tests remain credential-gated; deterministic provider and adapter contracts run in CI.
6. Merge in the order listed here. After each merge, the next branch will be rebased onto the new upstream `main` and submitted.

## Global acceptance criteria

- Every Python source, contract-data, and test file is mapped with no unverified entries.
- Core, Service, and Desktop tests, type checks, builds, and package smoke tests pass.
- Ubuntu, macOS, and Windows CI pass.
- Public documentation and migration guidance are updated.
- The final parity audit reports no unmapped Python delta.

## Reference verification from the snapshot

The complete snapshot previously passed:

- Python: 2,128 passed, 147 skipped, 384 subtests passed.
- TypeScript Core: 702 passed, 13 skipped.
- Service: 398 passed, 5 skipped.
- Desktop: 8 passed.
- 38 ESM/CJS exports and 1 CLI package smoke test.
- Ubuntu, macOS, and Windows CI.

The results above are reference evidence only. Each modular PR must independently pass the checks relevant to its own scope.