Stream text without exposing a full or partial stop sequence.

Bounded FIFO admission for the engine's independent KV contexts.

Two same-named runtime stubs plus one complete fake backend. Everything is generated and compiled inside a private temporary root whose path

Memoize exact per-layer replays; all implemented policies are layer-local.

A named TRUST-VERIFY-REFUSE failure. `code` matches c/rans.h's rans_err_name strings (E_TRUNCATED, E_TABLE_CRC, ...) plus the container-level

Split GLM's reasoning marker without leaking markers across stream chunks.

A serve-mode inkling, speaking the protocol in docs/serve_protocol.md.

A stub fixture could not be built — distinct from a loader-contract failure. Carrying the compiler command and its complete output means a broken

Strips Inkling's content markers from the visible stream and withholds <|content_thinking|> sections from `content` (they are reasoning, not a

One persistent `SERVE=1` engine process.

A repack-kind tensor (REPACK_KINDS) that CANNOT be repacked: wrong dtype for the fp8 path, or no `_scale_inv` sidecar. Named so the refusal is

Buffers converted tensors and flushes into new model-NNNNN.safetensors shard files once the buffer gets large -- st.h indexes every *.safetensors

Request-scoped K3 TOOL frames with the same stop policy as DATA.

The subset of `coli convert`'s namespace that cmd_convert reads.

Emits whatever `script` says, so a test can drive tool syntax through the parser.

A source shard recorded done in the resume manifest no longer matches its recorded fingerprint (size + mtime_ns): the input changed between ru

A run produced different tokens than the session's first run. The sweep only offers quality-preserving scheduling knobs, so any output change

The resume manifest cannot be trusted as-is -- e.g. its entries predate content-validated resume and carry no size/sha256/fingerprint records to

One persistent SERVE=1 kimi_k3 process; stderr goes to a file so the pipe can never fill and deadlock the engine.

Returns zeros_like(input) regardless of extra args. Replacing a decoder layer's self_attn/mlp with this makes the layer a clean identity: h =

Engine that emits a caller-supplied chunk sequence, to exercise the streaming reasoning splitter across arbitrary chunk boundaries.

Engine that rejects the prompt before ACCEPT — on_accept is never called.

rfile wrapper enforcing a CUMULATIVE deadline on reading one request. SEC: `timeout` below is per socket operation, so it restarts on every byte.

ACCEPTs the prompt (committing the streaming 200), then dies mid-generation.

#597 item 6: the engine's ACCEPT frame gates the HTTP commit, and its invariants.

#597: the DNS-rebinding guard must accept operator-trusted reverse-proxy Host values, while still rejecting everything else by default.

env_for_engine reads whatever flags the family cares about; every one not set here reads as "not given", which is what argparse yields.

Pin the raw-file authority check independent of manifest tampering: a corpus/comparisons/policy file that is not byte-identical to the frozen

The module's real-authority binding (AUTHORITY_SHA256) is exercised end to end only by RealFrozenCorpusOptInTest above, which skips whenever M

The banner's third line must describe the model that is loaded. It said "GLM-5.2 · 744B MoE · int4 · streaming CPU" for every checkpoint, bec

U5 fix round, finding 4: _canonicalize_metadata_order's slot-fit guard was a bare assert -- compiled out under python -O, where an oversized h

#379/#386 r2 (F9): `coli chat` on a non-glm model spawns openai_server as its local server. An explicit --cap must ride along on that command

A client that disconnects mid-response must not print a traceback. `coli chat` polls /health on a 2 s timeout while the model loads and drops

Local CPU must equal cluster-delegated expert sharding, token-exact.