fix(pflash): drafter resumes after a tool output

[Graffioh]

Summary

In an agent session, the drafter scores the whole prompt again from token 0 on each request after a compressed tool output. These requests should resume from the drafter's stored session instead.

Cause. A tool session's prompt ends with the tool output. Because of that, the window that represents the latest user question as a history query is never scored, and its rows are never stored. This window is the last token of the header of the question's reply. When the next request arrives, it needs that window from before the resume point and finds no rows, so it re-scores from 0.

Fix. The server passes that window as a keep span (CompressRequest::keep_query_spans). The drafter computes the window's rows from hidden states it already has and stores them with the scoring session. Nothing is scored from the keep span, so selections are unchanged.

The session's stored query windows use the same exact-row sets as the query step's token sets. A window is dropped once a later prompt diverges at or before it, so no request scores with rows from another prompt. The windows the next request needs are stored last, so the 8-entry cap evicts this request's windows first.

Measurements

Setup: one R9700 with the release profile and the unified view. The replay is the agent-paste replay, sessions A and B.

Turn	Before	After	Drafter resume
a8: question after a 57K-char tool output	12.8 s	4.5 s	0 → 98304
a9	11.6 s	3.3 s	0 → 98304
b3	10.2 s	5.5 s	0 → 65536

Compressed ids and selected chunks are identical before and after on all 10 compressed requests.

Tests

• New test test_pflash_tool_session_keeps_next_history_query. It checks that the keep span of a tool session equals the history span that the next request scores, and that the token prefix is the same.
• The existing multi-turn test now also checks that a chat whose prompt ends at the reply has no keep span.
• New test query_windows_are_position_sets: a window that reaches the divergence point is dropped, and one before it is kept.
• test_server_unit: 1001 passed, 0 failed.

🤖 Generated with Claude Code

[davide221]

Code review

No issues found. Checked for bugs and logic errors in the changed code.

_{Reviewed d8b0b1ae · 1 candidate issue checked, kept at ≥70% confidence · add the code-review label to review again}