[joelteply]

Continuum sizes its one launch from the engine's own fit instead of its own formula (continuum card 4a20255d, ONE-BUDGET 3b). Today a launch sized from a plan's transient started at 2,048, 8,448 or 83,968 tokens, and a growth relaunch fixed it afterwards while holding admission for 177-262 s.

What changes

• --fit-json: llama-fit-params prints the fit as one JSON object (status, n_ctx, n_gpu_layers, tensor_split, overrides), built with nlohmann::json so override regexes are escaped by construction. A failed fit reports its status and exits non-zero. A caller reads a typed report, never the CLI text.
• --kv-unified is now registered for LLAMA_EXAMPLE_FIT_PARAMS. Without it, a 5-sequence fit read 1,310,720 tokens (one cache per sequence) for an engine launched with one shared pool.

Measured

M5, Qwen3.6-35B-A3B Q4_K_M, 2026-11-04 01:36Z, with the M5's real launch arguments (--parallel 5 --kv-unified --ubatch-size 2048 -ctk q4_0 -ctv q4_0 -fa on -ngl 999):

• --fit-target 4096 --fit-ctx 20000: {"status":"success","n_ctx":262144,...}
• --fit-target 16384: the same 262,144.
• The server's own --fit on with the same arguments also chose 262,144 (01:29Z), so the tool and the server agree.

--fit-ctx is the smallest the fit shrinks to, not a size it must reach. A refusal comes from a margin the device can't leave, which this run didn't exercise.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo

[joelteply]

PASS at 7233724 (Cormac). --fit-json emits one typed object (status, n_ctx, n_gpu_layers, tensor_split, overrides), and --kv-unified now reaches llama-fit-params, so its n_ctx is the shared pool's.

names[status] is correct: the enum is SUCCESS = 0, FAILURE = 1, ERROR = 2 (common/fit.h:9-11). Non-blocking: a switch over the enum would turn a future fourth status into a compile warning rather than an out-of-bounds read.

Not reproduced here: Fable's M5 run (n_ctx 262,144 shared by 5, matching the server's own --fit on).

[joelteply]

PASS at 18ecc49 (Cormac). The delta from 7233724 is only the status switch: every enum case named, no default, so -Wswitch flags a fourth status. Nothing else changed.