fix: resolve traceability, safety, compatibility, benchmark, and auto-recall bugs

[mayankpandey0]

This Pull Request resolves 5 bugs identified in TencentDB-Agent-Memory:

1. Traceability Chain Fix: Uses normalizeToolCallIdForLookup inside flushL1 to prevent tool call ID lookup failures when LLMs mangle or strip underscores when echoing IDs. Also adds a defensive ?? "" fallback for result_ref in replaceWithSummary.
2. Parser Type-Safety: Adds the missing required result_ref: "" to the OffloadEntry object literal construction in l1-parser.ts to satisfy typing, and adds a parser unit test.
3. Dead Compatibility Options: Removes compatibility: "compatible" configurations from llm-caller.ts and llm-runner.ts (as the latest @ai-sdk/openai does not accept this key) and adds type checks via satisfies OpenAIProviderSettings.
4. Benchmark Path Error: Corrects the relative import path in benchmark-token-estimate.ts for fast-token-estimate.ts.
5. Auto-Recall Guard: Type-guards the optional searchL1Hybrid vector store capability call check in auto-recall.ts to prevent runtime TypeError failures.

[Maxwell-Code07]

Thank you for your contribution! We will review your code and get back to you as soon as possible.