fix(claude): pick models from workspace's serving endpoints (GDS-aware)

[dgokeeffe]

Priority

P0 — Claude Code's Opus tier is broken on every non-US-geo workspace. Workspaces in AU / EU / etc. serve databricks-claude-opus-4-6 but not opus-4-7; setup_claude.py hardcodes the latter, so selecting Opus 404s. App-side workaround for the gateway-side issue tracked at #8.

---

Summary

Fix for #26. Setup scripts had hardcoded model names that don't survive Geo Designated Services restrictions. This PR teaches setup_claude.py to query the workspace's /api/2.0/serving-endpoints, treat that list as the GDS-respecting "what's actually served here" oracle, and pick a working model in each tier.

Changes

utils.py (+52, no logic changes elsewhere):

• discover_serving_endpoints(host, token, timeout=5.0) -> set[str] — returns READY endpoint names. Empty set on any failure (preserves caller's fallback behaviour).
• pick_in_geo_model(preferred, available, fallback) -> str — first preferred entry in available; else fallback.

setup_claude.py (+38/-5):

• Calls discovery once after host+token are known.
• Picks ANTHROPIC_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL from the discovered list, walking a per-tier priority chain.
• Logs when an env-set request is substituted.
• Falls back to original env defaults if discovery returns empty (network failure, auth error, etc.) — behaviour matches main in failure modes.

Test Evidence (verified on the live deployment 2026-05-06)

Smoke test against daveok (AU geo workspace):

$ python3 -c "from utils import discover_serving_endpoints, pick_in_geo_model; \
              avail = discover_serving_endpoints('https://adb-7405613340366915.15...', '$PAT'); \
              print(sorted(avail))"
['databricks-claude-haiku-4-5',
 'databricks-claude-opus-4-6',
 'databricks-claude-sonnet-4-5',
 'databricks-claude-sonnet-4-6',
 'databricks-gpt-oss-120b',
 'databricks-gpt-oss-20b',
 'databricks-qwen3-embedding-0-6b']

# pick_in_geo_model picks correctly:
Active model:  databricks-claude-opus-4-6   (env asked for opus-4-7, not served)
Opus tier:     databricks-claude-opus-4-6   (preferred opus-4-7 unavailable)
Sonnet tier:   databricks-claude-sonnet-4-6 (preferred, served)
Haiku tier:    databricks-claude-haiku-4-5  (only option, served)

Before: /model shows opus-4-7 → 404 on every Opus call. After: opus-4-6 used as Opus default → works.

Test plan

• Deploy on a workspace WITH opus-4-7 (US geo) — confirm opus-4-7 still chosen (preferred winner).
• Deploy on a workspace WITHOUT opus-4-7 (AU/EU) — confirm opus-4-6 substituted, log line emitted.
• Disable network briefly during setup (or revoke the PAT temporarily) — confirm fallback to env-set defaults so setup doesn't fail outright.

Out of scope

• setup_codex.py, setup_hermes.py, setup_gemini.py, setup_opencode.py need the same treatment. Will follow as separate single-agent PRs once this helper lands and the semantics are reviewed. Single-agent scope here keeps the helper introduction reviewable on its own.
• Doesn't fix Test app against Geo Designated Services — AI Gateway Beta surfaces x-geo FMs #8 (the gateway-Beta picker showing x-geo models). That's gateway-side. This is the app-side workaround so users in non-US geos get a working Claude Code today.

Closes #26

This pull request and its description were written by Isaac.

[dgokeeffe]

@datasciencemonkey — flagging for priority review. P0: every non-US-geo workspace breaks Claude Code's Opus tier without this. Helper is small (utils.py +52 lines) and safe (failure ⇒ fallback to existing env defaults, behaviour matches main). Smoke test against the live AU workspace in PR body.