feat(utils): add cache-aware chat history compression and compaction

[gugu8intel-i9]

Add src/utils/chat-compression.ts for context window compaction and history pruning. When conversations approach the limits of a model's token budget, this utility replaces a contiguous middle slice of message history with a structured synthetic summary, preserving critical conversational invariants.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[codebuff-team]

The instinct here is good — cache-aware compaction that avoids invalidating a hot prefix (via minFreshInputRatio/freshInputRatio) is a real, non-obvious problem, and the tool-pair boundary alignment (alignToPairBoundary) shows you understand why naive slicing breaks tool call/result pairing.

But as submitted this isn't reviewable as a port candidate:

1. Line 8 literally says // ← verify this specifier on the import from @codebuff/common/util/tokens. That means you don't know if this compiles against the real package, and neither do we without checking. A PR shouldn't ship an admission that its own imports are unverified.
2. Nothing calls compressChat/compressChatToFit/shouldCompress anywhere. This is 554 lines of new surface area with no call site, so there's no way to tell whether the token-budget math, the tail-boundary logic, or the interaction with lastUsage/totalTokens actually behaves correctly against real message shapes used elsewhere in the CLI.
3. No tests. Given the amount of arithmetic here (trigger/target ratios, tail budget math, multi-pass convergence in compressChatToFit), this is exactly the kind of code where the repo would expect unit tests on findTailStart, measureContext, and the compress/no-op paths — regressions here would silently blow context budgets or thrash the cache in production.
4. It duplicates functionality that likely already exists somewhere in the private tree for context management — worth checking before building a parallel implementation, since the two need to agree on message shape and usage semantics.

To make this portable: confirm the import, wire it into an actual call site (even a small one), add unit tests for the boundary/threshold logic, and drop the placeholder comment.