live: heartbeat sync 2026-10-19T15:24:41-07:00 (Refs #11)

agent/managed-skills/gguf-lora-adapter-fit-census/SKILL.md

@@ -0,0 +1,93 @@
+---
+name: gguf-lora-adapter-fit-census
+description: "Read-only census before applying a GGUF LoRA/adapter to a llama.cpp (or fork) endpoint: Hub metadata, declared base provenance, runtime flag support, and the behavioural probe that proves the effect."
+---
+
+# GGUF LoRA / adapter fit census (read-only, before you promise a file to a payload)
+
+A 3 MB adapter is a payload change: it edits the compute graph at load. Do the census
+before editing an endpoint's trio, and prove the EFFECT behaviourally — no metadata,
+card table, or health gate can see it.
+
+## 1. Census the Hub repo first (no auth for public repos)
+
+```bash
+curl -sS "https://huggingface.co/api/models/<org>/<repo>?blobs=true" | python3 -c "
+import json,sys; d=json.load(sys.stdin)
+print(d.get('pipeline_tag'), d.get('library_name'))
+print([t for t in d.get('tags',[])][:20])          # base_model / base_model:adapter rows matter
+for s in d['siblings']: print(s['rfilename'], s.get('size'), (s.get('lfs') or {}).get('oid'))"
+curl -sS "https://huggingface.co/<org>/<repo>/raw/main/README.md" > /tmp/card.md   # grep for --lora flags
+```
+
+- `base_model` / `base_model_relation: adapter` name the pack the adapter was built for.
+- **A file with `lfs.oid = None` is a plain git object**: pin the sha256 of the Hub's bytes,
+  not an LFS oid. Fetcher law still applies: `curl -fsSL --fail-with-body`, byte count,
+  sha256, `--verify-only`, no resume.
+
+## 2. Read the adapter's own GGUF KV section (settles provenance in one read)
+
+Small python read of the KV pairs — no torch, no download beyond the adapter itself:
+magic `GGUF`, `u32 ver`, `u64 n_tensors`, `u64 n_kv`; key = `u64 len` + bytes; type `u32`
+(0 u8,1 i8,2 u16,3 i16,4 u32,5 i32,6 f32,7 bool,8 str,9 array,10 u64,11 i64,12 f64);
+array = `u32 elem_type` + `u64 count`.
+
+Keys worth printing: `general.type` (`adapter`), `adapter.type` (`lora`),
+`adapter.lora_alpha`, and whatever provenance the publisher wrote — e.g. PrismML/abliteration
+packs carry `prism.abliterate.base_gguf`, `base_sha256`, `direction_row`, `target_layers`,
+`target_embedding`, `sites`, `basis`.
+
+**Provenance is metadata, not a gate.** The loader matches **tensor names, not hashes**, so an
+adapter fit against a sibling quantization of the same model still applies — which is a
+decision to surface to the operator, not a failure to hide. Compare `base_sha256` against the
+sha the endpoint's own fetcher pins; if they differ, say so in the PR body, the env header and
+the endpoint page, and prove the effect behaviourally instead.
+
+Two publisher facts that decide whether a fork can carry the file at all: the build it was
+verified on, and what it deliberately does NOT touch (an embedding rank is the usual crash —
+`token_embd` left alone is a graph constraint, not an omission).
+
+## 3. Ask the runtime binary, not your memory
+
+```bash
+<llama-server> --help | grep -iE "lora|adapter"
+```
+Expect `--lora FNAME`, `--lora-scaled FNAME:SCALE,...`, `--lora-init-without-apply`
+(+ `POST /lora-adapters` for runtime scale). Prefer `--lora-scaled <file>:<scale>` as ONE
+argv token so a scale can never reach the engine without its file, and follow the endpoint's
+existing opt-out convention (the `MMPROJ_FILE=` shape: `:`-less default so an explicitly empty
+value means "off, on purpose"). Gate the deployer on the file's existence with the adapter's
+OWN fetcher named in the error text.
+
+## 4. Prove the effect: refusal/behaviour probe, before AND after
+
+The probe is not a bench, so the no-bench law does not block it — and the **before** half must
+be captured while the old shape is still live.
+
+- Same prompts, `temperature` low, both before and after; classify with a refusal-marker regex
+  and print only the opening characters, never the generated body.
+- **Budget ≥ 2 048 tokens on a thinking model.** A 600-token request can be swallowed whole by
+  the reasoning channel: `content` comes back empty with `finish_reason: length` and
+  `reasoning_content` full — which reads exactly like a refusal that never happened.
+- A leading `⚠️ Disclaimer`/hedge after the adapter is style, not the circuit: judge the marker
+  sentence, not the tone.
+
+## 5. Post-deploy reads (the ledger is the instrument)
+
+- Boot log shows `llama_adapter_lora_init_impl: loading lora adapter from '<file>'` and then
+  **dumps the adapter's KV block** — the provenance mismatch is printed at every start, so the
+  runbook should point at the journal instead of restating it.
+- Confirm the flag on argv: `tr '\0' ' ' < /proc/$(systemctl show <svc> -p MainPID --value)/cmdline`;
+  confirm `systemctl show <svc> -p DropInPaths --value` is empty (a leftover bench drop-in outranks
+  everything you just deployed).
+- Shape must be untouched by an adapter: `n_slots` / `n_ctx_slot` / KV lines unchanged, and
+  `/metrics` still answering (a llama.cpp endpoint is invisible to Prometheus without `--metrics`).
+- Root units log to the journal; rootless podman storage is read WITHOUT sudo (`sudo podman
+  inspect` sees a different storage namespace and reports "no such object").
+
+## Yields
+
+One sentence per question: does the file exist and at what size/sha; which pack it declares as
+its base and whether that matches the served payload; which flag the runtime actually offers;
+what the before/after probe classified. Anything else is a decision for the operator, and the
+rollback should be one env line on a file the endpoint already loads.