feat: add PostHog-gated model switching to chat

[burnedinthesky]

Type of changes

• Feature

Purpose

• Add the ability to select between two LLM models in the chat UI, gated by a PostHog feature flag so we can force a single model or open up a selector without a redeploy

The two models:

Send as (id)	Display name
gpt-oss-120b	GPT OSS 120b
gemma-4-31B-it	Gemma 4 31b

IDs match the sciedu-llm ALLOWED_MODELS allowlist exactly (note the capital B in gemma-4-31B-it; a mismatch is rejected with 400).

Additional Information

Feature flag: force-model-selection

Read with useFeatureFlagVariantKey; resolves to one of three modes:

• value is a model id -> force that model, no selector, every request sends that id
• "none" -> show a selector in the composer bar; user switches freely; choice persisted to localStorage
• absent / undefined / other -> hide the selector and omit model, so the backend uses its default (preserves todays behavior)

Changes

• New models.ts registry (single source of truth for ids and display names)
• New ModelSelectionContext / ModelSelectionProvider that resolves the flag, holds the selection, and persists to localStorage; wraps the chat pages in ChatLayout
• New ModelSelector (Radix Select) rendered in the composer bar; self-hides unless in "none" mode
• model threaded through createMessage -> useChat.sendMessage / startChat -> NewChatPage
• useModelSelection degrades to an "off" default outside its provider, so the shared chat components used by the edu-mode course chat keep working (no selector, no model sent)

Notes and follow-ups

• End-to-end gap: the live reply will not actually switch models until sciedu-backend (Go) threads model from its CreateMessageRequest down to the LLM /chat call. The API contract for this is in feat: add optional model parameter to createMessage sciedu-api#21.
• No new env vars; the flag lives in the PostHog dashboard.

Test

pnpm lint and pnpm build pass. Locally, override the flag in the console:

• posthog.featureFlags.override({ "force-model-selection": "none" }) -> selector appears; sending shows the chosen model in the POST /api/chat/... body; survives reload
• posthog.featureFlags.override({ "force-model-selection": "gemma-4-31B-it" }) -> no selector; every send carries "model":"gemma-4-31B-it"
• clear or false -> no selector; no model field in the body

🤖 Generated with Claude Code