upstream port: tool flow correctness fixes (#3313, #3297)

[mabry1985]

Two upstream bug fixes that improve tool execution correctness.

Upstream PRs

• fix(core): recover from truncated tool calls via multi-turn continuation QwenLM/qwen-code#3313 fix(core): recover from truncated tool calls via multi-turn continuation — when the model's tool-call output gets truncated by max_tokens, synthesize a continuation message so the model can re-invoke with a smaller budget.
• fix(tool-registry): add lazy factory registration with inflight concurrency dedup QwenLM/qwen-code#3297 fix(tool-registry): add lazy factory registration with inflight concurrency dedup — fix concurrency bug where simultaneous ensureTool() calls ran the factory twice; adds inflight promise dedup.

Why

Both are real edge-case fixes:

• fix(core): recover from truncated tool calls via multi-turn continuation QwenLM/qwen-code#3313 catches output-token-overflow cases that today silently produce broken tool calls.
• fix(tool-registry): add lazy factory registration with inflight concurrency dedup QwenLM/qwen-code#3297 prevents listener leaks from duplicate tool instantiation (especially relevant for AgentTool / SkillTool which have lifecycle hooks).

Scope / effort

Medium each, can land independently.

• fix(core): recover from truncated tool calls via multi-turn continuation QwenLM/qwen-code#3313: geminiChat.ts (+180 logic, +299 tests), turn.ts, useGeminiStream.ts. Our fork has reasoning_content handling in geminiChat.ts — recovery logic is isolated but verify against our subagent concurrency work.
• fix(tool-registry): add lazy factory registration with inflight concurrency dedup QwenLM/qwen-code#3297: tool-registry.ts (+128 lines test), config.ts, ~23 trivial integration files. Self-contained. Risk: if our fork has custom factory registration patterns, vet those.

Prerequisites

None.

Links

• fix(core): recover from truncated tool calls via multi-turn continuation QwenLM/qwen-code#3313
• fix(tool-registry): add lazy factory registration with inflight concurrency dedup QwenLM/qwen-code#3297

[mabry1985]

Investigation update: Both PRs in this issue (QwenLM#3297, QwenLM#3313) attempted as path-of-least-resistance during PR #196 work, both aborted:

• fix(tool-registry): add lazy factory registration with inflight concurrency dedup QwenLM/qwen-code#3297 lazy tool registry: 13 unmerged files including package-lock.json, ACP session paths (Session.ts, ToolCallEmitter), config.ts (3+ conflict regions), contextCommand.ts. The lazy factory pattern propagates through far more code paths than the original investigation suggested. Needs 1-2 days focused; not a quick port.
• fix(core): recover from truncated tool calls via multi-turn continuation QwenLM/qwen-code#3313 truncated tool call recovery: depends on ESCALATED_MAX_TOKENS constant added by a prior un-ported upstream PR (search upstream history for tokenLimits.ts changes). Plus 158-line conflict in geminiChat.ts (around our reasoning_content handling). Need to identify and port the prerequisite first.