fix: improve forwarder lifecycle recovery and Cursor transcript sync

[philau2512]

Summary

This PR brings the fork's forwarder lifecycle fixes, Cursor transcript synchronization improvements, and prompt/tool compatibility updates back to leookun/cursor-byok.

The changes are cumulative because this branch contains several related fork improvements. The main goal is to make Agent turns, tool execution, shell sessions, cancellation, and transcript projection more reliable under streaming, reconnect, and interrupted-transport scenarios.

Why

The Cursor client can receive events from multiple sources during an Agent turn:

• live provider/tool stream events;
• shell lifecycle events;
• persisted conversation history;
• transcript projection and replay.

Without strict lifecycle handling, this can result in:

• duplicated reasoning or progress rows;
• transcript content being overwritten by an older snapshot;
• hidden PatchEdit or Write operations remaining incomplete;
• terminal events being emitted more than once;
• shell output or tool results being lost after cancellation or stream closure;
• MCP tool calls failing when Cursor uses a server name instead of its canonical identifier.

This PR addresses these cases while preserving the existing upstream protocol and tool contracts.

Main changes

Forwarder lifecycle and transcript consistency

• Defer transcript synchronization while an Agent turn is still active.
• Project the final transcript only after a terminal state is reached.
• Preserve terminal markers such as turn_ended.
• Improve handling of continuation runs, checkpoint replay, and history backfill.
• Prevent duplicated reasoning/progress content across multiple tool calls.
• Preserve tool output and edit content during replay and compaction.
• Ignore late deltas and duplicate invocations after a tool has completed.

Tool, edit, and shell recovery

• Complete pending hidden PatchEdit and Write operations when a turn is canceled.
• Recover tool results when the transport closes before the terminal result.
• Improve shell streaming and shell lifecycle recovery.
• Prevent duplicate terminal events and invalid subscriptions after stream termination.
• Improve terminal tool handling and interrupted output recovery.

MCP compatibility

• Normalize MCP lookups using canonical server identifiers and tool names.
• Resolve server-name aliases to the corresponding canonical server identifier.
• Add regression coverage for direct invocation, persisted registries, and alias resolution.

Prompt and tool contract updates

• Update shared prompt behavior and language-policy handling.
• Synchronize prompt/tool definitions across Agent, Ask, Debug, Plan, and Multitask modes.
• Keep tool invocation and response-language behavior consistent across modes.
• Lock language response with rule:

---

alwaysApply: true
response_language: en
lock_response_language: true

Build, localization, and documentation

• Update build scripts and packaging configuration.
• Refresh the generated localization catalog.
• Update README and add the English CHANGELOG.md.

Review focus

The most important areas to review are:

1. internal/backend/forwarder/

   • lifecycle transitions;
   • terminal event ordering;
   • transcript projection timing;
   • cancellation and transport-closure recovery.

2. MCP tool normalization

   • canonical server identifier handling;
   • server-name alias resolution;
   • persisted registry compatibility.

3. Prompt/tool contract changes

   • consistency across prompt modes;
   • compatibility with the existing Cursor client.

Validation

The following checks pass:

• go build ./...
• go test ./...
• go test ./internal/backend/forwarder
• yarn build
• git diff --cached --check

The branch is currently 49 files changed, with approximately 3317 additions and 369 deletions compared with upstream/main.

Known limitation

go vet ./... still reports the existing lock-copy warning at:

internal/backend/agent/bridge/interaction/bridge.go:328

This warning is unrelated to the changes in this PR.

E2E validation with the installed Cursor client and a real provider has not yet been completed. The remaining risk is primarily around provider-specific streaming behavior and client-side transcript rendering.