feat: show token usage and rate-limit footer per turn

[aryansk]

Fixes #70

Problem

Free OpenRouter models are rate-limited and users get no feedback on how much context they're consuming. ChatOpenAI streams carry usage metadata, but GCode drops it. Issue #70 proposes capturing usage/rate-limit info from stream chunks and printing a compact footer after each turn: tokens in/out, approximate cost, and remaining rate limit, plus a /usage surface.

Change

• Add gcode/agent.py:_print_usage(response, ui) that extracts usage from response.usage_metadata (LangChain 0.3+) with fallback to response.response_metadata['token_usage'/'usage']
• Fix _stream to carry usage_metadata/response_metadata onto the canonical AIMessage it returns — without this the rebuilt message discarded the provider's token counts and the footer never printed
• Call _print_usage after each _stream in run_turn (initial + each tool-loop iteration, so multi-tool turns show one line per model call); silently omits when provider supplies no usage
• Footer format: Usage: <in> in / <out> out · total <n> · ~$<cost> (only present fields shown; zero counts are preserved via explicit None checks)
• Rate-limit header parsing dropped: ChatOpenAI never exposes raw response headers in response_metadata, so that path was unreachable. Surfacing rate-limit state needs an httpx hook on the underlying client — tracked as a follow-up
• Never breaks the turn on footer errors; unit tests added for metadata preservation, both usage sources, zero-count handling, cost display, and silent omission

Why this approach

Reuses the existing streaming path; the footer is the minimal user-visible fix matching the acceptance criterion "A turn prints usage when the provider supplies it; silently omits otherwise." No new /usage command needed because the footer is automatic; folding into /history would be redundant.

Testing

command: uv run ruff check gcode tests
result: all checks passed

command: uv run mypy gcode
result: success, no issues

command: uv run pytest tests/
result: 173 passed

Unit tests in tests/test_agent.py: _stream preserves usage_metadata; footer renders from usage_metadata, from response_metadata['token_usage'], keeps zero counts, shows cost, and stays silent with no usage.

Manual (with API key): start a session, trigger a turn, and observe footer like Usage: 842 in / 128 out · total 970 when OpenRouter returns usage; no footer when Ollama/local returns no usage.

Documentation and release impact

• User-facing documentation updated: footer is visible in terminal
• Changelog/release note needed: feature
• Migration or compatibility note needed
• No documentation impact

Review notes

• Known limitations: cost is shown only when the provider returns it; rate-limit remaining is not shown (see follow-up)
• Follow-up issue: surface OpenRouter rate-limit headers requires a custom httpx event hook — not reachable through LangChain's ChatOpenAI today
• Security/licensing considerations: none