[bong-water-water-bong]

Follow-up to #579, which added the manifest producer. It matched nothing and wrote mode=live with zero entries, so the models row would have passed vacuously — "0 onebit HRX entries and 0 NPU entries listed". A vacuous pass is worse than a failure: it hides the gap.

The evidence was in plain sight and I read past it twice:

• candidate 6's build log: onebit-models: 0 hrx + 0 npu
• candidate 7's gate R: SKIP models : … (0 HRX + 0 NPU entries are registered …)

Cause. The vendored server_models.json is a dict keyed by model id:

{"Qwen2.5-0.5B-Instruct-CPU": {…}, "Llama-3.2-1B-Instruct-CPU": {…}, …}   286 keys

My walk looked for an "id" field, found none at the top level, and descended into the values without ever associating a record with its key. So every onebit record was invisible.

Fix. A dict value that looks like a record (recipe or onebit_backend present) is taken with its key as the id when it has none. Lists and nested dicts are still walked, so the list-of-dicts shape keeps working.

Verified against the real catalogue on the candidate 6/7 build tree:

	hrx	npu
before	0	0
after	26	0

e.g. GLM-4.7-Flash-1bit, LFM2-1.2B-1bit, MiniCPM-V-4.6-1bit, MiniCPM5-1B-1bit, Qwen2.5-Coder-1.5B-Instruct-1bit.

npu stays 0 because the four onebit_backend hits in this catalogue are not "npu" — that is the catalogue's content, not a parsing gap.

bash -n clean.

Worth pairing with #585 in the next candidate: with this fix the models row becomes a real check on hardware (26 entries expected, each looked up in Lemonade's list), and with #585 the hrx-device row stops failing in the GPU-less VM. Together they turn two vacuous/false rows into the two that would actually have caught c5.

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 7ba6204)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 579 - Partially compliant Compliant requirements: The manifest file is now correctly generated from server_models.json The models row will no longer pass vacuously Skip reasons for chat-hrx and chat-npu are now accurate The secureboot row now reads Secure Boot status from firmware variables Non-compliant requirements: The chat-hrx "did not answer" failure is not addressed in this PR (separate issue) Requires further human verification: Verification that the chat-hrx "did not answer" issue is resolved in a separate change 585 - Partially compliant Compliant requirements: The hrx-device row now uses the same device detection logic as gpu Skip reasons now clearly indicate which device is missing Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Incorrect Manifest Generation for Keyed Catalogue The manifest generation logic was not correctly handling the keyed catalogue format where model entries are keyed by their ID. Previously, it only looked for an "id" field at the top level of each record, which caused it to miss all entries in the keyed catalogue. This resulted in a manifest with zero entries, leading to a vacuous pass of the models row. The fix correctly identifies records in the keyed catalogue by checking for the presence of "recipe" or "onebit_backend" fields and uses the dictionary key as the ID when the record lacks an "id" field. def models(node): # The vendored catalogue is a dict KEYED BY MODEL ID ({"Qwen2.5-0.5B-Instruct-CPU": {...}}); earlier # shapes were a list of dicts with an "id" field. Handle both, and take the key as the id when the # record has none - the first version only looked for an "id" field, so it matched nothing in the # keyed catalogue and wrote a manifest with zero entries, which made the models row pass vacuously. if isinstance(node, list): for x in node: yield from models(x) elif isinstance(node, dict): if "id" in node and isinstance(node.get("id"), str): yield node else: for k, v in node.items(): if isinstance(v, dict) and ("recipe" in v or "onebit_backend" in v): rec = dict(v) rec.setdefault("id", k) if isinstance(rec.get("id"), str): yield rec else: yield from models(v)

[github-actions]

Persistent review updated to latest commit 9e97a69

[github-actions]

Persistent review updated to latest commit 7ba6204