fix(catalog): accept current live Models.dev schema in parser (#4185)

[Hmbown]

Summary

Fixes #4185 — the foundation of the Models.dev catalog chain (#4184 → #4185 → #4187 → #4186 → #4188).

Live https://models.dev/catalog.json sends the provider-row interleaved field in two shapes:

• a bare boolean — interleaved: true (32 provider rows as of 2026-07-07, e.g. vercel, amazon-bedrock, azure, togetherai), and
• an object — interleaved: { "field": "reasoning_content" } (the majority, 530 rows).

The parser modeled only the object shape (ModelsDevInterleaved was a struct with an optional field), so serde_json::from_str::<ModelsDevCatalog> rejected every boolean row and the live catalog could not be parsed at all.

Change

• Replace the object-only ModelsDevInterleaved struct with an #[serde(untagged)] enum that accepts both the boolean and object shapes.
• Add is_enabled() and field() accessors so downstream lanes (Normalize Models.dev provider IDs onto CodeWhale provider kinds #4186 / Fetch and cache live Models.dev catalog into ProviderLake #4187) can read interleaved metadata without matching the enum.
• The object variant keeps field optional and ignores unknown sibling keys, so empty/partial objects and future upstream keys still parse.

Scope is strictly parser tolerance. No bundled-data changes (that's #4188), no picker wiring (separate lane). Only crates/config/src/models_dev.rs is touched; ModelsDevInterleaved has no other consumers in the workspace and the bundled snapshot never used the field.

Verification

Confirmed against a live catalog.json fetch on 2026-07-07: 151 providers, 32 boolean interleaved rows, 530 object rows, no other shapes.

• cargo fmt --all --check — clean
• cargo clippy --workspace --all-features --locked (with the repo's -D warnings + agreed -A set) — clean
• cargo test --workspace --locked — passes, except one pre-existing, environment-dependent failure unrelated to this change (tui::hotbar::actions::tests::skill_hotbar_action_activates_skill_through_dollar_alias, "skill registered from the startup skill cache"), which fails identically on pristine origin/main with this change stashed out.
• cargo build --release -p codewhale-tui — succeeds.

New tests in crates/config/src/models_dev.rs cover: interleaved: true, interleaved: false, interleaved: { "field": "..." }, empty/unknown-key objects, and a mixed live-ish sample with unmodeled provider/model keys.

Acceptance (from #4185)

• A fixture with interleaved: true parses successfully.
• A fixture with interleaved: false parses successfully.
• A fixture with interleaved: { "field": "..." } parses successfully.
• Existing bundled Models.dev fixture tests still pass.
• A live-ish sample from catalog.json can deserialize without schema errors.
• The parser still ignores unknown upstream fields safely.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[claude]

Claude finished @Hmbown's task in 1m 29s —— View job

---

Claude finished reviewing

• Gather context (diff, file, CLAUDE.md)
• Understand request (code review of the parser fix)
• Review correctness (serde untagged enum, edge cases)
• Review provider/model/route safety implications
• Review reuse/simplification
• Review test coverage
• Review security
• Post final review

Summary

This is a clean, tightly-scoped fix. ModelsDevInterleaved correctly moves from an object-only struct to a #[serde(untagged)] enum with Enabled(bool) and Field { field: Option<String> } variants, which resolves the actual bug: the live catalog.json's bare-boolean interleaved rows (~32 providers) were previously rejected outright by serde_json::from_str::<ModelsDevCatalog>.

Correctness

• The two variants are structurally disjoint (scalar bool vs. object), so there's no #[serde(untagged)] ambiguity risk — deserialization is deterministic.
• field: Option<String> with #[serde(default)] correctly tolerates {}, {"field": "..."}, and objects with unknown sibling keys (untagged + missing deny_unknown_fields means extra keys are ignored, as intended).
• is_enabled() treats any object form as enabled (matches the stated upstream invariant: object form only appears on interleaved-capable rows) and the boolean form reports its literal value. Reasonable and documented.
• Dropping the Default derive from ModelsDevInterleaved is safe — confirmed no callers in the workspace relied on it (the field is Option<ModelsDevInterleaved> at the call site, crates/config/src/models_dev.rs:230, so None covers the "absent" case instead).

Provider/model/route safety (EPIC #2608)

• No interaction with provider/model selection or routing logic here — interleaved is purely reasoning-wire metadata, not used to infer provider identity from prefixes or freeform text. No concern.

Reuse/simplification

• Confirmed via a workspace-wide grep that ModelsDevInterleaved/.interleaved has no other consumers, matching the PR's stated scope. is_enabled()/field() accessors are a reasonable minimal API surface for the downstream lanes (Normalize Models.dev provider IDs onto CodeWhale provider kinds #4186/Fetch and cache live Models.dev catalog into ProviderLake #4187) rather than making callers match the enum directly.

Tests

• Good coverage: bare true/false, object with field, empty object, object with unknown keys, and a mixed "live-ish" sample with an unmodeled provider key (doc) and unmodeled model key (some_future_flag) to prove forward-compatibility. This matches all of the acceptance criteria from Accept current live Models.dev schema in the catalog parser #4185.

Security

• No secret handling, exec, or input-validation concerns — this is pure data-shape parsing from a JSON catalog.

Note on verification: I wasn't able to run cargo test/cargo clippy in this sandboxed environment (cargo invocations require approval that isn't available here), so I relied on static review of the diff and full surrounding file context. The PR description states these all pass locally except a pre-existing, unrelated flaky test (skill_hotbar_action_activates_skill_through_dollar_alias), which is consistent with what I'd expect from this diff.

No blocking issues found. This looks ready to merge pending the actual CI run.