fix(pflash): batch automatic FlowKV compression

[davide221]

Summary

• route multi-turn --prefill-compression auto requests through FlowKV without requiring the unrelated disk-cache compression flag
• activate from aggregate aged-history tokens and select --prefill-curve from the full context length
• batch eligible aged messages inside one target/draft parking window, while preserving per-request fallback behavior for other backends
• include the keep ratio in the frozen-content cache key and clarify the CLI/docs behavior

Verification

• built dflash_server and test_server_unit from a clean origin/main worktree on RTX 5090
• focused regressions: 2 passed, 0 failed
• Lucebox server smoke: HTTP 200, expected OK, 25,097 prompt tokens reduced to 2,313 effective tokens in 5.43 s
• server logs confirm one target/draft park, one PFlash drafter load, and one restore for eight aged messages

Existing test failure

The full GPU unit binary reports 359/360. The untouched DS4 test fractional_route_quota_rounds_over_the_batch also failed before this patch; the focused PFlash/FlowKV tests pass.

[cubic-dev-ai]

All reported issues were addressed across 8 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>