conformance: report valid extension media types as informational

[thegovind]

Summary

ARD §3.3 defines an artifact-agnostic envelope: the type term carries an IANA media type, and artifact schemas are delegated outside the core specification. The conformance CLI contradicts that by warning on every media type outside a fixed nine-entry list, including application/ai-skill+md from the specification's own example in spec/ard.md. This PR reclassifies valid, unrecognized application/* types as informational and keeps warnings for real compatibility risks.

Companion draft: huggingface/OpenEnv#1093. The two are coordinated, but neither depends on the other. This PR adds no OpenEnv type to ARD core.

flowchart LR
  Open["§3.3: 'type' is an open term"] --> Current["Every unlisted type warns,<br/>including the spec's own<br/>application/ai-skill+md example"]
  Current --> Cost["False non-conformance<br/>and central allowlist pressure"]
  Cost --> Fix["Generic media-type classifier"]
  Fix --> Standard["Standard types: silent"]
  Fix --> Extension["Valid application/* types: informational"]
  Fix --> Risk["Malformed, parameter mismatch,<br/>deprecated, other top-level: warning"]

Loading

Why this is needed

The tool tells a publisher that a conforming type "is not one of standard discovery types." Three consequences follow:

1. The specification fails its own tool. Running the current CLI on the application/ai-skill+md entry from spec/ard.md emits a warning. After this PR it emits an informational message.
2. Conforming extensions read as non-conforming, which pressures ARD maintainers to grow a central allowlist. That pattern already drives OKF bundles (#27), install manifests (#43), and ASM (#66).
3. Extension ecosystems cannot evolve independently of the core release cycle, which is the outcome §3.3 was written to permit.

This is not a request to register one more domain type. It fixes the conformance tool's treatment of ARD's existing extension point.

The application/* scope is deliberate. Eight of the nine current standard discovery types use the application top-level type, and the motivating extension requests in #27, #43, and #66 do too. Well-formed non-application types keep their existing warning so this diagnostic-only PR does not broaden its behavior silently.

Broader goal

The broader effort is task-based discovery of portable RL environments across Hubs, using ARD as the outer catalog and search contract. ARD answers which resource matches a task and where its typed card lives, while an environment-specific card describes the domain details.

The companion OpenEnv RFC proposes application/vnd.openenv.environment-card+json and read-only openenv discover boundaries. OpenEnv is one motivating extension type for a generic ARD fix that is independently useful, and ARD does not depend on OpenEnv.

Changes

• Parse RFC 6838 restricted names and RFC 9110 parameters with a bounded re expression and no new dependency. A timeout test covers pathological backtracking on a 10,000-character invalid subtype.
• Keep standard discovery types silent after case and spacing normalization.
• Report valid, unrecognized application/* types as informational.
• Keep warnings for malformed syntax, unknown top-level types, known types with missing or extra parameters, and ADR-0008's deprecated application/mcp-server+json form. The deprecated diagnostic names its replacement.
• Add a neutral eight-entry fixture and 19 focused tests.
• Run the suite before and after installing optional JSON Schema support in CI.
• Document the diagnostic behavior.

Compatibility

• No specification or schema changes.
• No new runtime dependency.
• Exit codes are unchanged.
• Warning counts can only stay the same or decrease.

Test plan

Without optional dependencies:

• python3 -S conformance/tests/test_media_type_diagnostics.py -v
• python3 -S conformance/bin/conformance-test manifest conformance/tests/fixtures/extension-media-types.json

With optional JSON Schema support:

• python3 -m venv /tmp/ard-conformance-venv && /tmp/ard-conformance-venv/bin/python -m pip install jsonschema
• ARD_REQUIRE_JSONSCHEMA=1 /tmp/ard-conformance-venv/bin/python conformance/tests/test_media_type_diagnostics.py -v

Regression check:

• ./conformance/bin/run-conformance-demo

Expected: 19 tests pass (1 skipped without jsonschema, 0 skipped with it). The fixture exits 0 with 3 warnings and 3 informational messages.

This does not add the types from #27, #43, or #66 to the allowlist.

[YE-YI7]

Independent verification against the current live ASM catalog (30 entries; fetched 2026-08-28; SHA-256 1c72a3e4dc4f712a66e2531a67ea7fa6862994fea9e2905cc5d5b62f16351c2b): base aa3e598 PASS, 0 critical / 60 warnings; PR 56e1325 PASS, 0 critical / 30 warnings / 30 informational. The 30 application/asm+json diagnostics moved from warning to informational; the separate 30 representativeQueries warnings remained. All non-media-type diagnostic output was unchanged. I also ran the 19 focused tests with and without site packages, the neutral fixture in both modes, and the conformance demo; all passed. This matches the intended diagnostic-only boundary. I found no blocker in this slice.