feat: add context compaction to the rlm harness

[mikasenghaas]

Summary

• add the same optional CompactionConfig to the RLM harness as the bash harness carries
• cross ACP with a flat policy: a compaction toggle plus summarize_at_tokens
• pin nano-rlm 0e97b0b, which implements the matching engine-side design

Stacked on #2454 (bash compaction + the interception /v1/models relay) so that PR can merge first.
Companion: nano-rlm #147 — merge before this one.

Breaking

• RLMHarnessConfig.summarize_at_tokens moves to RLMHarnessConfig.compaction.summarize_at_tokens and no longer accepts a (lo, hi) range.
• Leave compaction unset to disable proactive and reactive compaction.

Verification

• uv run pytest -q tests/v1 — passed; live E2E tests skipped without PRIME_API_KEY

Terminal-Bench 2 e2e: 8 tasks, local vLLM poolside/Laguna-XS-2.1 at 32k (glm45 reasoning + glm47 tool parsers), compaction = {} so the engine discovers the threshold itself (32768 − 16384 = 16384). Trace analysis of the pinned engine:

• Threshold discovery and the proactive trigger work through ACP: compaction fired on the 3 episodes whose context crossed ~17k; episodes that stayed below (0.7k-6.7k peaks) never compacted; 20KB tool truncation visible where tool output was large.
• The runs surfaced and the pin now fixes three real integration bugs, each verified against the failing trace: /models discovery crashing on Python 3.10 containers (raw cast_to parse; now models.list()), and two interactions with main's new semantic-edges bookkeeping - a failed checkpoint attempt and a resampled unusable reply each left the compaction's summary-request claim held, killing the retry with "compaction already has a summary request".
• Laguna answers checkpoint prompts entirely in the reasoning channel, so under the summaries-are-content-only rule its compactions exercise the resample-then-end-cleanly path; summary carry-over across branches was demonstrated on content-channel models (Qwen3-0.6B, deepseek-v4-flash).
• A final combined verification run on a content-channel model is pending before merge.

🤖 Generated with Claude Code