CL-6477: fix embeddinggemma escaping the embedding-model filter

[TheGreatAxios]

Summary

• EMBEDDING_MODEL_NAME_PATTERN in packages/hub-client/src/model-capability.ts required a trailing delimiter after every family it matched, including embed(ding)?. Google's embeddinggemma:300m has none between "embedding" and "gemma", so it survived the filter, then won the alphabetical default-model tiebreak — every Myra-created agent on a box with this model pulled answered every chat turn with "embeddinggemma:300m" does not support chat.
• Fixed by no longer requiring a trailing delimiter for embed(ding)? specifically (the short abbreviations minilm/bge/gte/e5/arctic-embed keep the stricter, delimiter-bound match — they're prone to false positives as bare substrings). Capability data, when a deployment carries any, already takes priority over this name check in isCompletionCapable — this pattern is only ever the belt-and-braces fallback for the uncataloged case (an unprobed local Ollama pull).
• Separately: added gpt-oss:20b as the new curated Ollama default in packages/hub-client/src/catalog-seed-data.ts — an overnight run found it correct, tool-obedient, and 12-25x faster than the prior default qwen3.8:27b (8s vs 100-216s for the same task). Kept in its own commit so it can be reverted independently.

What I verified

• New tests in packages/hub-client/src/model-capability.test.ts cover: embeddinggemma:300m is never selected as a chat default even with no delimiter after "embedding"; the other embedding names (nomic-embed-text, all-minilm, bge-m3, qwen3-embedding) still get filtered (no regression); a chat-capable model is selected when both kinds of candidate are present; and probed capability data (when present) still wins over the name fallback.
• bun test green across packages/hub-client, packages/chat, packages/inference-settings, packages/onboarding, packages/inference-catalog, packages/provider-pricing (1046 pass, 0 fail) — updated one existing onboarding test (complete-credential.test.ts) that hard-coded the old curated-default name.
• tsc --noEmit clean for packages/hub-client and packages/onboarding.
• bun run lint from the repo root: 0 errors (12 pre-existing warnings in unrelated files, not introduced by this change).

What I did not verify

• Did not boot a real stack or an actual Ollama instance against this change — verified via unit tests only, per the no-stack-boot instruction for this lane.
• Did not run the full monorepo bun run check (typecheck across every workspace) — scoped to the affected packages plus repo-root lint, given the timebox.
• Found, but deliberately left out of scope: OLLAMA_CAPABILITY_MAP in packages/hub-client/src/credential-test.ts has no entry for Ollama's own "embedding" capability (only completion/tools/vision are translated), so a genuinely-probed embedding-only Ollama model whose name doesn't match the name-filter at all would still collapse to an empty capability list — indistinguishable from "never probed" — and fall through to the (now-fixed) name heuristic. Closing that gap for real would need a new signal that Interchange's own Capability enum (@intx/types, never to be modified here) has no room for, likely a schema/DB-level change — bigger than this hotfix's scope. Flagging for a follow-up ticket rather than expanding this PR.

Not merging

Per instructions for this lane, not merging — a peer reviewer should merge after checking the above.