train: one training-context derivation; K/V format resolved and recorded (6edd495c slice 1)

[joelteply]

• One owner for whether a K/V format can hold the model: llama_kv_cache_types_check.
  llama_init_from_model refuses by it, and a caller choosing a format asks it first
  instead of reading a null context, which also means out of memory.
• One cache-type parser in common (common_kv_cache_type_from_str / _types / _types_str).
  arg.cpp's file-static copy is deleted.
• common_context_params_for_training holds the training mode's only differences from
  serving: no rollback slots, chunk = batch = ubatch, a 2-sequence unified cache,
  flash attention, every output, no embeddings. server-train's hand overrides are deleted.
• The training K/V format is RESOLVED, not hard-coded: requested (cache_type_k/v on
  /train), else serving's, else the default Q8_0. Each counts only where training has a
  measured run with it (F16 c424b00, Q8_0 a8a235d). Only a requested format is
  refused, by name. Serving at Q4 still trains, at Q8_0, as today.
• The receipt records training_cache {type_k, type_v, source} beside serving_cache.

Where it changes behavior: the core's KvCachePlan serves Q8_0 on CUDA and Metal
(kv_cache_plan.rs:160), so the 5090 and the M5 still train at Q8_0, unchanged. CPU and
Vulkan serving is F16, so those nodes now train at F16 instead of Q8_0: twice the
training K/V bytes, charged to the same lease (budget minus context bytes), and no
quantized dequant on a CPU. Not measured yet; the receipt is a /train beside
responsive serving.

Compiled: clang++ -fsyntax-only on the four changed TUs (IntelMac, 2026-10-13 ~17:00Z).
Not built or run.

Supersedes #59's context_cache_type parser (a copy of kv_cache_type_from_str); #59 should be closed in favour of this.

Not in this slice: shared physical-pool admission (slice 2, after the Metal unified-memory type prerequisite with its MoE-cache measurement), and the chat-input mismatch (serving reviewer).

Receipt needed before adoption: a /train beside responsive serving, on the 5090 or M5 (Q8_0 path, unchanged) and on a CPU node (F16 path, changed), with training_cache and serving_cache from the receipt, plus context_mib, graph_mib, SHA and time.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE

[joelteply]

Closing unmerged. This extends /train as a job in a SECOND context. The design (docs/architecture/ONE-RESIDENT-MODEL-PATIENT-DOCTOR-DREAM.md, S4) is a DREAM stage in the server loop, one optimizer micro-step between update_slots decode batches on the resident context, and Joel restated it on 2026-10-13. I built on the second context without reading that doc first; this PR refines the shape the design deletes.