feat(compact): add forge_compact crate with lossless sliding-window conversation compaction

[tusharmath]

Summary

Add a new forge_compact crate implementing a generic, lossless conversation compaction algorithm using a sliding window approach, and clean up the existing compaction code by removing the token_threshold_percentage feature and over-defensive error handling.

Context

Conversation compaction is critical for long-running agent sessions: as context grows, it must be summarised before exceeding model limits. The existing compaction logic in forge_app was tightly coupled to domain types and contained defensive fallbacks that masked bugs. This PR extracts the core algorithm into a standalone, fully-tested crate (forge_compact) and trims the feature surface of Compact config to what is actually needed.

Changes

• New forge_compact crate: generic Compaction<Item> struct with a compact_conversation sliding-window algorithm that works with any type implementing ContextMessage
• Summary type: wraps original messages alongside their synthesised summary, enabling lossless reconstruction of compacted history
• compact_conversation algorithm: grows a window from size 1 up to the full length, triggers compaction when the caller-supplied threshold is exceeded, and restarts after each successful compaction; respects retain to keep the most-recent messages untouched
• find_compact_range: correctly handles tool-call / tool-result pairs so they are never split across a compaction boundary
• Removed token_threshold_percentage from Compact config — the percentage-based capping logic was removed in favour of simpler absolute thresholds
• Simplified Compactor in forge_app: replaced .get(start..=end) defensive fallbacks with direct slice indexing (panics are now loud, not silent bugs)
• Removed dead tests that were testing the removed token_threshold_percentage feature

Key Implementation Details

Compaction<Item> is fully generic over any Item: ContextMessage + Clone. The caller supplies two closures: summarize (produces a single summary item from a slice) and threshold (decides whether the current window needs compacting). A retain count pins the most-recent messages outside the compaction window, preventing recent context from being summarised away.

The sliding window algorithm is designed to be lossless — each Message::Summary carries a source field with the original messages, so callers can reconstruct the full history if needed.

Testing

The new crate ships with a comprehensive unit test suite covering:

• Basic single-pass compaction
• Compaction restart after a new summary is introduced
• retain window pinning the tail of the conversation
• Correct tool-call / tool-result boundary handling
• Cache-key stability (deterministic output for the same input)

cargo insta test --accept -p forge_compact
cargo insta test --accept -p forge_app

Links

• Related to ongoing context-management work in the forge_app compaction pipeline