Migrate /try chat to Anthropic /v1/messages + table rendering and UX polish

[KillerQueen-Z]

What

Free NVIDIA-hosted models (DeepSeek V4 Flash, Qwen3 Coder, Llama 4 Maverick) leaked raw `<tool_call>{...}</tool_call>` text into the chat — the OpenAI-compat gateway path was lossy when translating tool schemas for them. Switching the chat endpoint to Anthropic-native `/v1/messages` — the same path Franklin CLI uses — preserves the schema end-to-end so every model (free or paid) actually calls tools.

Verified with Qwen3 Coder 480B in the live dev server: tool calls fire as native `tool_use` blocks, no `<tool_call>` text leak.

Bundles a few related reliability + rendering + UX fixes that surfaced during the migration.

Protocol migration

• `CHAT_ENDPOINT` → `/v1/messages` with `anthropic-version: 2023-06-01` header
• Tool schemas rewritten: OpenAI `{ type:"function", function:{ name, description, parameters } }` → Anthropic `{ name, description, input_schema }`
• Request body: `system` lifted to top level, `max_tokens` required, `tool_choice` reshaped to `{ type:"tool", name }`
• Response parsing: read `content[]` blocks (`tool_use` / `text`) instead of `choices[0].message.tool_calls`
• Tool results: Anthropic `tool_result` content blocks batched into a single user message instead of OpenAI `role:"tool"` messages
• Vision: OpenAI `image_url` → Anthropic `image.source`, auto-detecting `data:` URLs (base64 source) vs http(s) URLs (url source)
• Proxy: `messages` added to path allowlist; `anthropic-version` added to forwarded request headers

Chat reliability

• Payment rejection rollback — when a wallet signature is cancelled, the orphan user message is removed from history instead of persisting into the conversation list and titling a stray "New chat" that leaks to other windows on next hydrate
• Empty conversations auto-delete — `useChatHistory.setMessages` drops a conversation (sidebar + GCS DELETE) when its message list goes to zero, instead of keeping an empty shell
• Error path cleanup — catch block now clears `steps` + `activeTool`, so ` · thinking` no longer keeps spinning after a wallet rejection error toast

GFM table rendering

Wide assistant replies (e.g. smart-money comparison grids, vocab grids) used to fall through as raw `| col | col |` text. Now parsed and rendered.

• `parseTableAt` detects header + `|---|---|` separator + body rows; tolerant to leading/trailing pipes and `:---` / `---:` / `:---:` alignment
• `TableBlock` renders the parsed table inside a card with a header bar carrying Copy as Markdown / Download CSV (RFC 4180 quoted) / Expand to fullscreen overlay — matches the existing `CodeBlock` pattern
• Horizontal scroll wrapper so wide tables don't break the chat column on narrow viewports

UI polish

• Wallet pill restyle — single-row layout (network chip + balance + address inline), no dead zone between content and disconnect icon; quieter Base chip; ghost disconnect button with hover affordance
• "OpenAI O3" label (was bare "O3") — aligns with same-group naming (`GPT-5.3 Codex`, `DeepSeek V4 Pro`, `Grok 4.1 Fast`)
• 3 new prediction-market starter prompts replace the heavy-to-implement selfie / cold-call / character-consistency cases — picked to be broad, evergreen, always-returns-data queries:
  • Movers — 24h biggest odds movers leaderboard
  • Sports — hottest sports markets right now
  • Tech — most-active AI / tech markets right now

Cleanup

• Drop the dead streaming chat path in `runChat` (`supportsTools()` was hardcoded to `true` so the streaming fallback was unreachable; it also referenced OpenAI SSE shape that wouldn't have worked anymore against the new endpoint)
• Remove the unused `streaming` state + `setStreaming` calls
• Drop the `casePhone` i18n key (no longer rendered)

Verification

• `tsc --noEmit` ✅
• `eslint` — clean on touched files; the one pre-existing `FranklinChat.tsx:66` `react-hooks/set-state-in-effect` error is unrelated to this PR (file unchanged in that area)
• Tool-calls verified live in dev against Qwen3 Coder 480B (free NVIDIA model) — tool flow works end-to-end via Anthropic protocol
• Table rendering verified live against assistant output with `| col |` markdown
• Wallet rejection / orphan-msg-rollback verified in dev by cancelling MetaMask signature

Follow-ups (not in this PR)

• Friendlier 429 error display — currently NVIDIA rate-limit responses surface as raw `{"status":429,...}` JSON to the user; `upstreamErrorMessage` already has the translation but only runs on 200-with-error-content, not on actual HTTP 429
• Streaming over Anthropic SSE if slow models start exceeding Cloud Run / Cloudflare timeouts (the rewrite is non-streaming for now — fine for the tool-loop pattern but won't give a typewriter effect)