fix(engine): report unimplemented workflow capabilities instead of refusing to load

[justinchuby]

Problem

admit_inference_metadata failed the load whenever the package declared a capability this runtime has not implemented. In practice that rejected every newly-exported mobius package, which declares eight workflow capabilities:

bounded_state_recurrence, emit_valid_length, linear_effects, loop_induction_values,
nested_control_flow, serving_service_contract, typed_emit, workflow_ssa

The rule this PR applies

Not "capabilities are advisory" — that would be wrong, and review caught exactly why. The split is on one question:

Does the capability describe the document, or prescribe computation this path must perform?

	derived from	on the bare-decode path	verdict
Descriptive — workflow_ssa, serving_service_contract, bounded_state_recurrence, typed_emit, emit_valid_length, linear_effects, loop_induction_values, nested_control_flow	the workflow spec	this path derives its IO from the graph and never interprets the manifest	warn, continue
Prescriptive — parameter_adapters, heterogeneous_adapter_batching	metadata.adapters or the workflow spec	this path never applies the overlay; decoding anyway emits base-model output	fatal

Structural defects remain fatal throughout.

Evidence for the descriptive half

I relaxed the gate locally and ran a previously-rejected package end-to-end on the native CUDA backend:

profile_native --model hy-mt2-b128 --ep cuda --runs 1 --tokens 24 --prompt "The capital of France is"
throughput: 7.86 tok/s, 127.251 ms/step
generated_token_ids: [13847, 13, 185, 66410, 316, 244, 5069, 280, 9391, 13, 185, 66410, ...]

It loads, builds a native decode session, executes, and produces a coherent stream whose first token (13847) matches the same model exported without a workflow manifest. The gate was refusing work that runs.

Review then confirmed against the code that all eight derive from workflow spec shape (validation.rs:157-330), and that bounded_state_recurrence in particular comes from workflow.state[..].recurrence — not from the model's KV/attention geometry, which the decode path takes from metadata.decoder_io() / graph derivation without ever reading workflow.state.

Why the prescriptive half stays fatal

The two adapter capabilities are a concrete weight overlay, not a description. The only consumers of .adapters are pipeline/adapters.rs and pipeline/mod.rs — the separate PipelineRuntime, not this path. A well-formed LoRA package would otherwise have loaded here and silently produced base-model output: wrong output, not degraded output, with nothing for the caller to notice.

They are keyed on capability name, not on metadata.adapters.is_some(), because workflow_required_capabilities derives the same two capabilities from the workflow spec with no adapter service present (validation.rs:201 for adapter-segment/count/scale input roles, validation.rs:232 for an onnx-genai.parameter-overlay component). An is_some() check would let those workflow-declared overlays through with only a warning — the exact bug this guards against. That case now has its own test.

Scope

Both admit_inference_metadata callers are bare-decode (native load.rs:370, ORT load.rs:1188). The workflow/serving runtime (PipelineRuntime) does not call it, so this does not weaken a serving path.

Complementary to #1723 (native backend for the universal pipeline.workflow runtime): as that lands these capabilities become genuinely implemented and the warnings stop firing on their own. No file overlap — this touches only engine/load.rs.

Per maintainer decision, no kind: discriminator is added: the workflow shape is the direction for all packages, so there is nothing to discriminate between.

Tests

cargo test -p onnx-genai-engine --lib → 435 passed, 0 failed.

• unsupported_declared_capability_is_reported_but_does_not_block_decode
• workflow_manifest_capabilities_do_not_block_the_bare_decoder
• declared_adapters_fail_closed_because_this_path_cannot_apply_them — structurally valid adapter service (real target manifest, selection contract, artifact)
• workflow_declared_parameter_overlay_also_fails_closed_without_an_adapter_service — asserts metadata.adapters.is_none(), pinning the case that makes the explicit capability list correct
• structural_defect_still_fails_closed — asserts the specific unknown-component defect, not just the generic prefix

The two adapter tests each assert the error is not a structural one before asserting the guard message. That guard was not cosmetic: four successive drafts of those fixtures were structurally invalid (wrong field names, component: decoder instead of model, empty artifacts, artifact paths outside adapters/{name}/, unpinned adapter ABI) and would otherwise have "passed" for the wrong reason.

Closes #1767

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.28%. Comparing base (b9d9c48) to head (1f4afb8).
⚠️ Report is 21 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1790      +/-   ##
==========================================
+ Coverage   80.19%   80.28%   +0.09%     
==========================================
  Files         411      411              
  Lines      199146   200160    +1014     
  Branches   199146   200160    +1014     
==========================================
+ Hits       159706   160702     +996     
+ Misses      33996    33994       -2     
- Partials     5444     5464      +20     

Flag	Coverage Δ
cli-ort-linux	72.47% <ø> (ø)
cli-ort-windows	71.97% <ø> (-0.10%)	⬇️
mlas	85.20% <ø> (ø)
offline	80.42% <ø> (+0.09%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 17 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.