[88plug]

llama-server b11749 keeps two host caches next to the KV cache, and the fit counted neither:

• the host prompt cache, -cram/--cache-ram, default 8192 MiB (common/common.h cache_ram_mib = 8192);
• context checkpoints, -ctxcp/--ctx-checkpoints, default 32 per slot (n_ctx_checkpoints = 32). Each one is a host copy of the slot's partial state (tools/server/server-context.cpp create_checkpoint, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY): the SWA cache (llama_kv_cache::state_write skips a non-SWA cache under that flag) plus any recurrent state (llama_memory_hybrid::state_write). The server makes them only for models with SWA or recurrent layers.

For the Gemma QAT GGUF on the desktop fixture's CPU plan (5 slots at 262144, KV filling the room), the defaults add 8 GiB plus 5 x 32 x 300 MiB, about 47 GiB of checkpoints, on top of a 45Gi pod limit.

Changes:

• console/fit.py: the llama.cpp flags carry --cache-ram 0 -ctxcp 4 (LLAMA_CTX_CHECKPOINTS, with the reason next to it). ckpt_seq gives one slot's checkpoint bytes: 4 x (SWA bytes per token x GGML_PAD(min(n_ctx_seq, n_swa + n_ubatch), 256) cells + recurrent state).
• On the CPU placement the checkpoints share the KV room in size_kv, so the slot count already accounts for them. On the GPU and experts-in-DRAM placements they are added to the host bytes.
• The doc's ckpt_bytes, the pod limit and every guard rung's limit include them.
• docs/fit.md: the flags and the arithmetic.

Effect on the fixtures: the Gemma QAT desktop CPU plan goes from 262144 x 5 with no caches counted to 262144 x 4 with 4.7 GiB of checkpoints, and its pod limit goes from 45Gi to 44Gi. The GPU plans keep their contexts and slots.

Overlap: #214 (fit: size sliding-window KV per engine) adds the same _pad helper and moves llama.cpp's SWA KV to the same 1536-cell count. Whichever lands second needs a rebase; ckpt_seq can then use swa_tokens.

Tests:

• tests/fit_test.py LlamaHostCaches:
  • the Gemma QAT CPU plan on desktop-i9-nogpu emits --cache-ram 0 and -ctxcp K, and counts seqs x K x 300 MiB in host_bytes, the pod limit and each rung's limit;
  • Flash-Next's checkpoints equal seqs x K x its recurrent state;
  • a full-attention model makes no checkpoints.
• tests/fit_test.sh: k3d-gpu fit --json for the Gemma QAT on the desktop shows the flags, and its host bytes and pod limit hold the checkpoints.
• Both suites failed before the change.

Run before opening:

• tests/fit_test.sh: 35 passed
• python3 tests/fit_test.py: 59 tests OK
• tests/model_opt_test.sh: 37 passed
• tests/claims_check.sh and python3 tests/ledger_drift_check.py: clean
• ruff (F, E9) and shellcheck v0.11.0 at warning severity: clean

[88plug]

Superseded by #229 (merged)