fix(claude): repair the broken test file and add the endpoints it imports

[fig-ai-agent]

What this fixes

tests/test_claude_endpoints.py (added in #241, commit 3702790) is chat prose wrapped around a fenced code block, not Python. It fails to even parse:

SyntaxError: invalid character '→' (U+2192)

Because pytest imports every file under tests/, that one file broke collection for the whole directory — every valid test beside it was blocked. This PR repairs it and supplies the module it imports but that never existed.

New files

File	Purpose
app/services/claude_client.py	async Claude client over httpx — stream_message() yields decoded SSE frames, send_message() returns the full body. No new SDK dependency.
app/api/v1/endpoints/claude.py	POST /claude/chat/stream (SSE), POST /claude/chat/tools (tool-calling), GET /claude/status

Pre-existing breakage fixed (both blocked any import of app.services)

1. app/services/__init__.py did from .users import UserService — UserService does not exist in users.py (it defines UserRepo). A package __init__ runs before any submodule import, so this made app.services.* app.main-blocking. Now resolves lazily (PEP 562), so one broken downstream module can no longer take out the whole package.
2. app/core/config.py — Settings() raised on the repo's own .env: undeclared keys (WHATSAPP_*, BYTEPLUS_*, …) plus a required OAUTH_CLIENT_ID. Added extra = "ignore" and a default.

Verification

$ python -m pytest tests/test_claude_endpoints.py --noconftest -q
4 passed

Collection errors across tests/, vs. untouched origin/main:

	errors	collected
origin/main	10	28
this branch	5	38

No regressions — every remaining error is present on main unchanged (comm diff of the two error sets). Five fixed, ten new tests collected.

Two things worth your eye

• app/.gitignore contains a bare service* rule that ignores anything under app/services/. The new client module would have been silently dropped, so it is force-added (git add -f). That rule looks like an accident — worth deleting separately.
• Two assertions in the recovered test were wrong and I corrected them: lines == 'data: {...}' compared a list to a str (and would still be wrong even if the types matched), and tool_calls["name"] indexed a list. The endpoint also emitted data: "[DONE]" (JSON-quoted) instead of the bare SSE sentinel.

Not included on purpose: app/db/repositories.py is also broken (AsyncSession/Item undefined — 5 of the remaining collection errors). It is unrelated to this deliverable and belongs in its own PR.