fix(core): estimate semantic request context

[rekram1-node]

Summary

• build one token-estimation string from system text, JSON-encoded tool definitions, and semantic message history
• count tool names and inputs plus textual, JSON, and error tool results using provider-aligned serialization
• exclude direct media, tool-result files, and referenced OpenAI image-generation payloads
• avoid counting internal message structure, metadata, identifiers, and media transport bytes
• leave the canonical provider request and overflow-recovery behavior unchanged
• cover image exclusion and tool-input pressure with focused regressions

Testing

• bun test test/session-compaction.test.ts from packages/core
• bun typecheck from packages/core
• focused OpenAI Chat and Anthropic Messages image-lowering tests from packages/llm

[rekram1-node]

Final estimator shape after independent review:

• System-part text is appended directly.
• Tool definitions are appended once as JSON.
• Message history is flattened to model-visible content.
• Tool calls contribute the tool name and safely encoded input.
• Text, JSON, and error tool results follow their provider-visible serialization semantics.
• Direct media and file items nested in multimodal tool results contribute no base64 transport bytes.
• Referenced OpenAI Responses image-generation results omit their opaque image payload because replay sends an item reference.
• Message IDs, metadata, native state, canonical wrapper keys, and other internal structure are not estimated.

The actual provider request and one-shot context-overflow recovery remain unchanged. Focused regressions verify real user image lowering, hosted image-generation payload exclusion, and tool-call input pressure. The independent review found no remaining substantive issue after these corrections; model-specific media token cost remains an accepted estimation limitation.