fix(orchestrator): gate diffusion behaviour on registry kind, not the xdit name

[sushildubey171]

Description: what and why

hunyuan_image3 is registered kind=scriptable with throughput_unit="img/s" and ships baseline_hunyuan_image3.yaml / profile_hunyuan_image3.yaml — but several call sites decide "is this a diffusion workload?" by comparing the framework name to "xdit". Those sites silently treat it as an LLM workload.

Before this change a hunyuan_image3 session:

	behaviour	should be
SharedState._roofline_throughput_unit	tok/s	img/s (its own registry entry)
bypass analysis report unit	tok/s	img/s
bypass diffusion-roofline sidecar	not emitted	emitted
bypass steady-state default	off	on
steady-window repeat threshold	LLM threshold	diffusion threshold (2)
compute_roofline_breakdown_from_state	LLM decode roofline	diffusion ceiling

All of these now consult framework_registry (is_scriptable / throughput_unit), which is already the canonical predicate in 26 other places. executors/profile.py:632 is the pattern I followed — it calls is_scriptable() for the server-less branch and only then narrows to xdit for the xfuser-specific check, with a comment noting hunyuan_image3 must not trigger it.

The kernel tools keep working standalone: they try the registry and fall back to a name check, mirroring the existing _is_scriptable_framework helper in tracelens_analysis.py.

Deliberately not changed. I went through every "xdit" literal in non-test code; most are correct and are left alone: the xfuser pip name and xDiT git URL, XditAdapter registration, per-framework config paths, the PATH guard for the xdit console script + hipcc, the XDIT_* env blacklist, the xDiT default grid seeds, and the xfuser baked-profiler verification. cli/model_gate.py already calls is_scriptable() with the literal only as an import-failure fallback.

The two health-warning codes keep their bypass_xdit_* spelling for wire stability even though the condition is now generic; only their messages change from "xDiT" to "diffusion". Renaming those codes felt like a separate call for you to make.

Linked issue(s)

Addresses item 3 of #1114 (see my correction comment there — my original "15 hardcoded comparisons" was a raw grep count and overstated it; this PR covers the subset that is genuinely wrong). Does not close #1114, whose main ask is an SGLang diffusion route.

This is also the prerequisite that issue names: with the kind checks registry-driven, adding a new scriptable framework gets closer to the "single-table edit" framework_registry's own docstring promises. I deliberately did not add an sglang-diffusion entry here — without a matching Magpie wrapper script (external repo) it would be dead config, and #1114 is asking you how you'd want that shaped.

Tests: added/updated? commands run?

No test in the tree referenced hunyuan_image3 before this change, which is why the divergence went unnoticed. The new tests assert against every registry entry rather than a hardcoded list, so a future framework cannot regress the same way:

• scriptable detection and throughput unit match FRAMEWORKS[*].kind / .throughput_unit for all entries, in both the tool and the reader
• the standalone fallback path (registry import forced to fail) still classifies xdit and hunyuan_image3 correctly
• compute_roofline_breakdown_from_state routes xdit and hunyuan_image3 to the diffusion ceiling, and sglang/vllm/atom to the decode path
• unknown/empty framework still resolves to tok/s + decode

pytest src/hyperloom/agents/kernel/tests                     ->  1146 passed, 23 skipped
pytest <roofline_ceiling, framework_adapters, diffusion_roofline,
        xdit_integration_units, backend_gating>              ->   264 passed
ruff check                                                   ->  clean
ruff format --check                                          ->  identical before/after (no new debt)

I did not run the full suite to completion locally — it exceeds ~10 min in my environment — so I ran the kernel-agent tree plus the modules consuming these call paths.

Breaking changes: no

Serving frameworks and unknown/empty names are unchanged: the registry falls back to sglang, so they still resolve to tok/s and the decode roofline. The only behaviour that changes is for kind=scriptable frameworks other than xdit — i.e. hunyuan_image3 — which now gets the diffusion treatment its registry entry already declared. Warning code strings are unchanged.