feat(compaction): hoist system messages into a stable cache-friendly prefix

[tishachawla-jg]

Summary

Improves cost compaction by attacking the real cost driver: the prompt-cache discount. Cached prefix tokens bill far cheaper than fresh ones, and a volatile value ahead of the static instructions busts the whole discount.

_compact_messages (in src/tokenops/control/integration.py) already deduped non-system messages but pinned system messages in place. It now hoists all system messages into a stable leading prefix — static instructions first, volatile conversation after — preserving relative order within the system block and within the tail.

This closes the gap between what the context_compaction policy already documents ("move volatile values below the static prefix") and what the live rewrite actually did.

Behavior

• No-op when system messages are already first (the common case), so most callers see no change.
• When a system message was interleaved later, it moves into the prefix. Honest tradeoff, noted in the docstring: the first such reorder busts the old cache once, then the new stable order caches; net win when the system block is stable across calls.
• Dedup of non-system messages unchanged (keeps first occurrence).

Tests

• New: test_compact_hoists_system_into_stable_prefix, test_compact_is_noop_when_system_already_first.
• Full suite: 227 passed, 2 skipped, no regressions.

Not in this PR (follow-ups)

• Cache-aware detection (fire on cached-vs-fresh token ratio, not just size) — the cache_guard idea in Add a policy: the contributor path, and what is worth adding next #66.
• Recording tokens-saved per compaction for the dashboard.
• True volatile/static splitting inside a message, which needs the prompt-assembly hook.

🤖 Generated with Claude Code