fix: correct token double-counting for Anthropic and Bedrock providers

[amitksingh1490]

Summary

Fix token double-counting for Anthropic providers and incorrect prompt_tokens mapping for Bedrock, bringing token reporting.

Context

1. Anthropic double-counting: Anthropic streams usage as cumulative values across message_start and message_delta events (ref). Our code was using accumulate() (sum) to combine them, producing 1 + N instead of the correct N for output tokens.

2. Bedrock prompt_tokens: prompt_tokens was set to total_tokens instead of input_tokens, inflating the reported input count by including output tokens.

Changes

Bug 1 – Anthropic cumulative usage (6 providers affected)

• crates/forge_domain/src/message.rs: Added Usage::merge() which uses max() per token field instead of +. Cost is still summed since cost events are additive.
• crates/forge_domain/src/context.rs: Added TokenCount::max() for comparing two token counts by inner value while preserving Actual/Approx semantics.
• crates/forge_domain/src/result_stream_ext.rs: Changed partial-usage branch from accumulate() to merge(). Also fixed cost-only events to sum costs instead of replacing.

Affected providers: anthropic, claude_code, anthropic_compatible, vertex_ai_anthropic, minimax, alibaba_coding.

Bug 2 – Bedrock prompt_tokens

• crates/forge_repo/src/provider/bedrock.rs: Changed u.total_tokens → u.input_tokens.

Affected providers: all Bedrock models.

Key Implementation Details

The distinction between accumulate and merge:

Method	Strategy	Use case
accumulate()	Sum all fields	Session-level totals across independent requests
merge()	Max per token field, sum cost	Combining partial streaming events within one response

Documentation references

• Anthropic streaming event types: https://platform.claude.com/docs/en/build-with-claude/streaming#event-types
• OpenAI Responses API usage schema: https://developers.openai.com/api/reference/resources/responses#(resource)%20responses%20%3E%20(model)%20response_usage%20%3E%20(schema)

Testing

All existing tests updated + new tests added:

• test_into_full_anthropic_streaming_usage_merge – covers real Anthropic pattern where message_start has output_tokens=1
• test_into_full_anthropic_streaming_usage_merge_zero_output – covers Vertex AI pattern where message_start has output_tokens=0
• test_usage_merge_anthropic_cumulative – unit test for merge logic
• test_usage_merge_preserves_costs – verifies cost summation in merge

cargo test -p forge_domain --lib
cargo test -p forge_app --lib
cargo test -p forge_repo --lib

All 1,381 lib tests pass.