fix: stop false emergencies (real context window) + preserve compressed summaries

[ranxianglei]

Root cause of the death-loop (from acp.log analysis)

The configured modelContextLimit (80K) was ~3.4× smaller than the model's real window (272K, derived from realUsage.percent). The nudge computes fullness as tokenCount / modelContextLimit, so it entered emergency at 64K tokens — only 23.5% of the real window — and stayed in emergency continuously, thrashing the model with compression nudges it could not satisfy (protected zone / too-small ranges). At peak the nudge's own pct hit 270% while the provider showed 81%.

Fixes

1. Derive the real window from the provider's usage percent (src/index.ts, src/config.ts)
realUsage.percent = tokens / realWindow ⇒ realWindow = tokens / percent. When the reading is in a trustworthy 5–100% band, use it as modelContextLimit for the nudge decision. The nudge now matches the footer the user sees; a misconfigured limit no longer causes false emergencies.

2. Stop dropping compressed summaries (src/messages.ts)
coreOutToAgentMessages previously continued past every acp_summary_* message — the kernel's prune node replaces covered messages with these placeholders, and dropping them gave the model total amnesia for compressed sections, driving re-fetch feedback loops that inflated context. Now emitted as user-role text.

Not in this PR (blocked)

• emergencyNothingLeft adapter wiring: read nudge.breakdown.emergencyNothingLeft (from acp-kernel release v0.1.16 #53) to withhold the urgent nudge when nothing is legally compressible. Blocked on acp-kernel release v0.1.16 #53 merge + a version bump (field absent in bundled 0.0.17).
• biggest-reclaim-tier-first nudge arbitration: separate acp-kernel change.

Verification

typecheck clean, 156/156 tests pass (+3 new for resolveEffectiveContextLimit; updated the old "filters out summary" test to assert emission). Build 410 KB.