feat(engine): unify compaction under one Compactor abstraction; add summarizer + truncation; comprehensive stats

[OsherElhadad]

Summary

Unifies the pipeline under a single Compactor abstraction and adds two new compaction approaches, uniform model/credential handling, a self-caching-client lever, and comprehensive metrics — plus a real 10-task SWE-bench run.

Every compaction approach now implements one interface — engine.Compactor (conversation in → conversation out), name-registered and selected via compactors:. reduce, extract, summarize, truncate, cache are all Compactors; a new approach plugs in the same way (implement → Register → list by name).

What's included

• Summarizer (internal/summarize): ReSum-style prompts ported from CE-Manager; replaces older turns with one <summary>, keeps the last keep_last verbatim, dropped span recoverable. Opt-in.
• Truncation: naive keep-last-N baseline (no model), recoverable. Opt-in.
• Extract fully decoupled from the reduce candidate flow — does its own large-output detection (reduce.SelectLLMCandidates) and runs before reduce.
• Uniform model + credentials for every LLM compactor (config.ModelTransport + engine.ModelSpec): provider/model/auth/base, inline api_key or key_env, and source: config|incoming (reuse the proxied request's own model + credentials).
• --reduce-cached-prefix: run the deterministic reducers + extractor on a self-caching client's cached prefix (e.g. Claude Code) instead of deferring to its cache.
• Comprehensive /stats: per-session breakdown, a recent-per-call ring, and full latency distribution (p50/p95/p99/max/mean) at global + session scope; the per-call slog event carries the same metrics.
• Naming purge: the upstream prototype name is gone from the runtime surface — markers («labcx:…», labcx_expand), /labcx/expand, x-labcx-bypass, LABCX_* env, the eval compose service; provenance comments reworded (only the ../winnow path reference to the prototype dir remains).
• Docs: README / RUNNING / RESULTS + the swe-bench integration doc updated for the new abstraction, flags, stats, and the run below.

Validation

• make fmt lint test build clean; new tests for the Compactor registry, summarize/truncate, model-spec incoming, modelBase prefixing, and the enriched aggregator.
• Live: offline reducers up to −98%; Haiku extractor −57…−80% on structured fixtures; summarize + truncate live on Haiku.
• SWE-bench (10 real tasks, Claude Code → lab-cx → gateway): −35.5% aggregate input tokens (4,405,821 → 2,840,296; 1,565,525 saved over 263 requests; −10.4…−51.8% per task), from the deterministic reducers (1,598 blocks) on the cached prefix. Caveat: arm64 Epoch bases (native Apple-Silicon runs) are best-effort for grading, so resolve status may understate accuracy; token reduction is architecture-independent.

Assisted-By: Claude Opus 4.8