fix: models.dev last-segment match + official-vendor disambiguation

[aurorax-neo]

Summary

• Fix models.dev enrichment for routed/gateway wire IDs (e.g. route/model-leaf): match by case-insensitive exact last / segment only.
• When multiple catalog hits share that leaf, prefer a unique official/source provider (anthropic / openai / google* / xai/x-ai) whose provider key agrees with the model’s source vendor (so a gateway listing like google-vertex/xai/grok-* does not count as xAI).
• If no unique official hit, enrich only when all hits share identical capabilities/thinking; otherwise leave unmatched.
• UI/identity always keep the full wire ID; unmatched models expose all thinking levels and default to off.
• Also harden plugin-mcp child teardown (process group + SIGKILL + stdio destroy) so desktop tests no longer hang.

Test plan

• Shared 1030/1030 + typecheck
• models.dev targeted 45/45; composer/thinking 42/42
• Desktop full suite 2861 pass / 0 fail / 1 skip
• Acceptance against a live OpenAI-compatible gateway catalog: official enrichments for common sonnet/gpt/grok leaves; custom agent-style aliases stay unmatched (no false collapse onto *-latest)

[vastsa]

Thanks — I reviewed commit 9069108 and ran the targeted suites (desktop 116/116; shared 22/22). The core last-segment matching and official/shared-capability disambiguation look implemented, and the PR CI is green.

Before merge, I found a few completeness/regression concerns:

1. initialThinkingLevelForBinding() in packages/shared/src/thinking-levels.ts now returns "off" whenever no default is stored. This helper is also used for normal new-session materialization, not only unmatched models. The current UX spec says a reasoning binding with no stored default falls back to its highest enabled level. Please scope this behavior to unmatched models, or update the product contract and add a user-path regression test.

2. In thinkingProviderForModel(), an unmatched model with an existing binding whose thinkingLevels is [] still gets an empty supportedThinkingLevels array; the new “all levels selectable, default off” behavior only occurs when there is no binding. Please clarify/fix the empty-binding case or adjust the PR summary.

3. The model-list row now uses the wire ID, but Composer.tsx still derives the main model chip label from selectedModelInfo.displayName, so the full wire ID is not preserved in every UI surface.

4. The matching/spec change is not reflected in docs/spec/03-runtime/13-model-catalog-and-selection.md and the related decision/spec docs; they still describe the older suffix/route behavior.

Could you address these before merge?

[vastsa]

Follow-up on my earlier review: I verified that the alias chip is intentional and the exact wire ID remains the model identity and list-row value. The empty thinkingLevels: [] case was a real gap for the PR's unmatched-model contract, and the global default-off change also affected known catalog models. I pushed a minimal landing fix on top of the contributor commits: known catalog matches retain the highest-enabled fallback; unmatched models, including empty generic bindings, start at off while exposing the full thinking ladder; explicit non-empty bindings remain authoritative. Specs and regression tests are synchronized. Relevant candidate checks pass: shared tests 23/23, desktop typecheck, and targeted model/Composer/MCP tests 135/135. The branch also includes the latest origin/main.

[vastsa]

The first post-fix CI run exposed one unrelated latest-main regression: the attachment-history source-contract test still expected the old local name after main changed it to . I synchronized that test in 328eab2 without changing runtime behavior. Local full desktop suite now passes 2971/2971 with the repository loader; the PR branch remains based on the latest origin/main.

[vastsa]

Correction to the previous note: the stale attachment-history source-contract test expected the old local name 'shouldInline' after latest main changed it to 'canInline'. I synchronized that test in 328eab2 without changing runtime behavior. Local full desktop suite now passes 2971/2971 with the repository loader, and the PR branch remains based on the latest origin/main.

[vastsa]

The second CI failure was the native Node test runner unable to resolve extensionless TypeScript imports inside composer model.ts, reached by the new regression test. I added explicit .ts extensions in 5067a90; this is test-loading hygiene only and leaves runtime behavior unchanged. The exact CI command now passes locally: 2971/2971 desktop tests.