fix(agent): recover from provider context overflows

[wsp1911]

Summary

Improve context overflow handling with structured provider error detection, bounded recovery compression, and recent-context preservation.

Overflow recovery now remains within the same logical model round while exposing each recovery attempt through lifecycle events and traces.

Type and Areas

Type:

Bug fix

Areas:

Rust core, AI adapters, agent stream, core contracts

Motivation / Impact

Some providers omit token usage, making pre-request context estimates unreliable. An underestimated request could exceed the provider context window and fail without recovery.

This change:

• detects context overflow using structured HTTP and provider error metadata, with conservative message matching as a fallback
• retries failed requests after progressively trimming and compressing older context
• preserves recent exact context on atomic message and tool-call boundaries
• retains user and Todo anchors when the latest turn is only partially included
• keeps recovery attempts under one logical model round
• distinguishes logical round attempts, local retries, and transport attempts in events, logs, and traces
• preserves the existing local fallback compression budget

Recovery is bounded to avoid unbounded retry loops. Plain HTTP 400 responses without overflow evidence remain classified as invalid requests, and output-token limit errors are excluded from context overflow recovery.

Verification

Passed:

• pnpm run fmt:rs
• cargo test -p bitfun-core-types
• cargo test -p bitfun-ai-adapters
• cargo test -p bitfun-agent-stream
• cargo test -p bitfun-core round_executor::tests
• cargo test -p bitfun-core execution_engine::tests
• focused compressor tests
• cargo check --workspace
• node scripts/check-core-boundaries.mjs
• git diff HEAD --check

cargo check --workspace reported only pre-existing CLI unused-import and dead-code warnings.

Reviewer Notes

The main behavioral changes are:

• Main-request overflow recovery is limited to two recovery cycles.
• Compression overflow handling uses up to four progressively smaller attempts.
• Recent exact context targets approximately 10K tokens by default.
• Structured provider status, code, and error type take precedence over message-based classification.
• Adapter-local transport retries remain separate from logical model-round attempts.

No persistence migration is required.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.