Registry: reviewed architecture gaps, and a guard against faking support with an alias

[bong-water-water-bong]

This brings the distinct parts of agent-6d3fa8's architecture registry (a port of the 1bit-MONSTER census, never pushed) onto the registry that #94 landed. The #94 registry stays the base: generated from the pinned sources, nothing typed in by hand.

Added

• docs/arch-gaps.md: the reviewed evidence for eleven HF architecture classes that only look like a supported family: DeepSeek V4.1, Qwen3 + Mamba-3, MolmoAct2, Cosmos3-Edge, Qwen-Drive, EnglishBase, GDN2, Haiku, Vapor, Fidel and Zarya. For each: the official config, what it adds over the closest supported architecture, why an alias would advertise support that can't be served (or would silently compute something else), and what real support needs.
  • The review was written for 1bit-MONSTER's own engine. A preface says so, and says which names and paths are that engine's.
  • The per-family evidence is kept as written.
• registry/significant.json: the eleven classes.
• tools/registry_build.py --check-gaps (ctest registry_gaps, no pins or network): fails when one of the classes becomes mapped without a recorded mapped_ok reason, and the build warns the same way. A pin bump that brings real upstream support gets a person's look before the class counts as supported. Tested both ways: it passes on today's registry, and fails on a hand-added VaporForCausalLM mapping.
• docs/registry.md links the page, and the site nav lists it.

Not taken from the port

• the alias table (registry/arch.h, 2,030 strings) and its 100% coverage claim: aliases are exactly what the gap review warns against, and Model registry and the daily HF census (PORTING step 5) #94's counts come from the backends' own code;
• about 1.16 million lines of committed census JSON snapshots;
• its three census workflows (Model registry and the daily HF census (PORTING step 5) #94 has the daily census);
• the old engine's artifact registry.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 366b02b

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 94 - Partially compliant Compliant requirements: Registry generation from pinned sources via tools/registry_build.py registry/architectures.json contains 265 HF architectures with counts per backend Checked models via tools/registry_check.py running tests/serve_e2e.sh over registry/check_models.tsv Documentation updates in docs/registry.md and docs/arch-gaps.md The registry to not be manually edited, only generated The gap review to prevent fake support via aliases Non-compliant requirements: Workflow for daily census runs at 05:17 UTC Requires further human verification: Verification that the daily census workflow has been properly set up and runs as expected

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Gap checking logic The gap_violations function in tools/registry_build.py checks for architectures that are mapped in the registry but do not have a mapped_ok reason recorded in registry/significant.json. This is a good safeguard against accidental aliasing of unsupported architectures. However, the check only considers the presence of mapped_ok in the sig dictionary, and does not validate that the reason provided is sufficient or correct. If a new architecture is added to the registry without a proper mapped_ok reason, the check will fail, but it won't validate the reason itself. def gap_violations(registry): """Classes reviewed as not aliases (registry/significant.json, docs/arch-gaps.md) that the registry maps without a recorded reason ('mapped_ok'): each needs a person's look.""" sig = json.load(open(os.path.join(ROOT, "registry/significant.json")))["classes"] mapped = registry["architectures"] return [(cls, mapped[cls]) for cls, e in sig.items() if cls in mapped and not e.get("mapped_ok")] Exit code handling The main function in tools/registry_build.py returns an exit code of 1 if report_gaps finds any violations, but it also returns 0 if no violations are found. This is correct behavior for a check tool. However, the function also calls report_gaps twice: once when --check-gaps is used and once during the normal build process. This could lead to confusion if the exit code from the second call is not properly handled by the calling process. if a.check_gaps: bad = report_gaps(json.load(open(a.out))) if not bad: n = len(json.load(open(os.path.join(ROOT, "registry/significant.json")))["classes"]) print(f"none of the {n} reviewed significant classes is mapped without a reason") return 1 if bad else 0 text = json.dumps(build(), indent=1, sort_keys=False) + "\n" report_gaps(json.loads(text))