Method save

server/src/server/disk_prefix_cache.cpp:424–502

    // Update last_used on disk.
    entry.last_used = now_unix();
    entry.hits++;
    // Optionally rewrite header timestamp (non-critical, skip for perf).
    return true;
}

// ─── Save ───────────────────────────────────────────────────────────────

bool DiskPrefixCache::save(int slot, const std::vector<int32_t> & prompt_ids) {
    if (disabled()) return false;

    // Learn layout on first save.
    if (!layout_known_) {
        learn_layout(slot);
        if (!layout_known_) return false;
    }

    // Check minimum token threshold.
    if ((int)prompt_ids.size() < config_.min_tokens) return false;

    auto ref = backend_.snapshot_ref(slot);
    if (!ref.ctx) return false;

    PrefixHash hash = hash_prefix(prompt_ids.data(), (int)prompt_ids.size());

    std::lock_guard<std::mutex> lock(mu_);

    // Skip if already on disk.
    if (find_entry(hash) >= 0) return true;

    // Pre-write budget check: estimate file size and reject if it would
    // exceed budget even after evicting all evictable entries.
    if (config_.budget_bytes > 0) {
        uint64_t payload = 0;
        uint32_t ntens = 0;
        size_t table_est = 0;
        for (ggml_tensor * t = ggml_get_first_tensor(ref.ctx); t;
             t = ggml_get_next_tensor(ref.ctx, t)) {
            ntens++;
            payload += ggml_nbytes(t);
            table_est += 2 + std::strlen(t->name) + 4 + 32 + 8; // name_len + name + type + ne[4] + nbytes
        }
        size_t est_size = DISK_CACHE_HEADER_SIZE + table_est + payload;
        if (est_size > config_.budget_bytes) {
            std::fprintf(stderr, "[disk-cache] skip save: estimated %.1f MB exceeds budget\n",
                         (double)est_size / (1024.0 * 1024.0));
            return false;
        }
    }

    std::string path = make_path(hash);
    std::string tmp_path = path + ".tmp";

    if (!write_file(tmp_path, ref, prompt_ids)) {
        std::remove(tmp_path.c_str());
        return false;
    }