fix(server): drop stale prefix-cache entries when a snapshot slot is reused

[davide221]

Problem

Follow-up to #370, fixing the root cause behind its symptom.

The inline and full-compress prefix caches hand out snapshot slots round-robin (next_slot_), and the counter advances in prepare_*_snap even when the snapshot later aborts (degenerate boundary < 512 tokens, failed generation, client disconnect). One burned step is enough: a later confirm wraps onto a slot that a live entry still references. From then on the entry table and the slot contents disagree — the entry's hash describes one token stream, the slot holds a snapshot of another.

A stale entry then misbehaves in two ways:

• follow-up prompt shorter than the slot's snapshot → snapshot_longer_than_prompt, a failed request with an empty completion (the silent agent hang server: fall back to fresh prefill when a cached snapshot is longer than the prompt #370 reported; server: fall back to fresh prefill when a cached snapshot is longer than the prompt #370 downgrades it to a conservative cache miss);
• follow-up prompt longer than the slot's snapshot → the restore path attaches KV from the wrong token stream with no validation: silent context corruption, which server: fall back to fresh prefill when a cached snapshot is longer than the prompt #370 does not cover.

Fix

When confirm_inline_snap / confirm_full_snap commit a snapshot into a slot, erase every other entry still pointing at that slot. A slot holds exactly one snapshot, so at most one entry may describe it. 24 lines, no API change.

Validation (RTX 3090, Qwen3.6-27B Q4_K_M, --prefix-cache-slots 2)

Deterministic repro: short conv (snap@3801 → slot 0) → tiny conv whose snap aborts (burns a slot step) → big distinct conv (8.2K tokens, wraps onto slot 0) → shortened follow-up of the first conv.

• on main pre-server: fall back to fresh prefill when a cached snapshot is longer than the prompt #370 this sequence ends in ok=false out=0 error=snapshot_longer_than_prompt;
• with this fix the wrap logs [pc] dropping stale entry for reused slot=0, the follow-up is a clean miss with correct output, and a longer same-conversation follow-up restores a valid snapshot (correct KV, correct answer) — the corruption window is closed, cache effectiveness preserved.

Greedy outputs across the whole sequence match the no-cache baseline. All 1905 server unit assertions pass.

🧙 Built with WOZCODE

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}