feat: turnkey conversation-history compression (one-call episodic memory)

Summary

Make conversation-history compression a one-call, turnkey feature. Today the episodic store exists (add_episode_sync(episode_id, summary, metadata)), but the caller must produce the summary themselves and manually decide which turns to keep verbatim vs. compress. That friction is why the most universal pattern in real agents — growing chat history — does not actually benefit from contextweaver out of the box.

Why this matters (evidence)

A fleet-wide review of ~12 NOS GenAI agents found the context firewall fits almost none of them, because they fetch fresh retrieval per turn and keep only Q&A text in history (nothing large is re-carried). The pattern that is shared across nearly all conversational agents is unbounded or crudely-capped history:

• dsi-av-chat-conv — history accumulates with no token cap/truncation found.
• dsi-km-agent — hard cap of "last 3 Q&A pairs" (memory.py L133–138): bounds growth but hard-drops older context.
• The same shape recurs in billing/customer conversational agents.

The right tool for this is episodic summarization — but it's currently too manual to be a drop-in. Closing this gap is the single highest-leverage way to make contextweaver useful in more repos.

Proposed API

A turnkey memory facade that owns the keep-recent / compress-old policy and (optionally) generates the summary:

mem = mgr.conversation_memory(
    keep_recent_turns=4,          # verbatim window
    summarize_after_turns=8,      # when to roll older turns into an episode
    summarizer=None,              # None => deterministic extractive summary; or a callable / LLM hook
    token_budget=2000,            # optional cap for the compressed history block
)

mem.add_turn(role="user", content=...)
mem.add_turn(role="assistant", content=...)

# returns recent turns verbatim + a single rolling episodic summary of older turns
history_block = mem.render_sync()

Internally this wraps the existing add_episode_sync / build_sync machinery; no new storage primitives required.

Acceptance criteria

• One call to obtain a bounded history block (recent verbatim + compressed older), with no caller-side summary production required.
• Deterministic default summarizer (no LLM dependency) so it works in deterministic pipelines; pluggable LLM summarizer hook (reuse the #384 path) for higher-fidelity rollups.
• Honors a token_budget for the rendered history block.
• Summarizer is told which turns it is compressing and must preserve entities/IDs referenced by later turns (round-trip test).
• Documented runnable example: a 30-turn conversation showing prompt-token growth flat-lining vs. the naive accumulate baseline.

Out of scope / relationship to other issues

• Not the context firewall (#403–406, #399, #402): those compress tool results; this compresses conversation turns.
• Complements #384 (LLM-assisted summarization) by reusing it as the optional summarizer backend.

Start by reading the existing add_episode_sync and build_sync entry points, then inspect the #384 summarization path for a reusable hook. Define and test the conversation_memory facade, including bounded rendering, deterministic and pluggable summarization, token budgets, and preservation of entities and IDs. Finish with the requested runnable 30-turn example comparing growth with naive accumulation.