ENG-664: expose compaction result for host persistence + skip low-yield re-summarization

[ea-rus]

Summary

• Adds ChatSession.last_compaction — a read-only result of the most recent
  history compaction this session ran ({"summary": str, "covered_through": int}
  or None). summary is the exact compacted message content (including the
  _COMPACTED_MARKER prefix) so a host can seed it back verbatim next turn and
  have _summarize_history recognize + extend it in place. covered_through
  is how many of the session's seed (initial_history) messages are now
  folded into it, letting a host map that count onto its own message list to
  find the cutoff.
• _summarize_history now skips the LLM round-trip when there's little new
  material to fold in (excluding any already-carried-forward prior summary
  from the measurement) — avoids paying for a summarization call that would
  barely extend an already-large summary.
• No change to the compaction algorithm itself, and no new required
  constructor parameters — existing hosts are unaffected.

This is the anton-side half of ENG-664 (cowork-server persists last_compaction
across turns instead of resending full history every turn). The once-per-turn
compaction guard (_compacted_this_turn) already existed and needed no change.

Test plan

• New tests/test_history_compaction.py: last_compaction is None
  before any compaction, populated correctly after one, and the
  low-yield-skip guard (including the "prior summary doesn't count as
  new material" case).

Fixes https://linear.app/mindsdb/issue/ENG-664/conversations-resend-the-full-history-every-turn-and-throw-away-the

[Dav1dF]

Read this alongside the prompt-caching work (the gateway's cache alignment from mindsdb/mindshub_inference — ENG-601/ENG-586) to check for a conflict, and wanted to note the opposite on the record: this design is exactly what prefix caching needs.

• Replaying the persisted summary verbatim keeps the seed byte-stable between compactions, and the tail is append-only — so a cached prefix survives across turns.
• The low-yield re-summarization skip is also the cache-friendly behavior: every re-summarization rewrites the front of history and resets the cached prefix, so avoiding marginal ones saves twice.

One expectation to write down somewhere (here is fine): when cache alignment turns on in prod, cache-hit dips will correlate with compaction events. That is inherent to any history trimming, and last_compaction / the compaction event give us the attribution — no change needed in this PR, just flagging so nobody reads the dips as a caching bug later.

Linear: ENG-664; caching context on ENG-586.