Track token usage and cost of every LLM call

[anhnh2002]

Problem

In subscription mode (claude-code / codex), caw already returns cost_usd for every run, but each call overwrote backend.last_usage and only the incremental updater read it. A full codewiki generate never reported its tokens or cost. On top of that, Codex runs were almost always priced at $0 by caw.

Changes

• UsageLedger (backend.py) records one entry per CLI process / agent run: phase, kind, target, model, seconds, tokens and cost. It is thread-safe for caw's MCP worker threads. Nested caw sub-module agents are separate processes that record themselves, so totals add up without double counting.
• Both backends route usage through LLMBackend._record_usage, which still sets last_usage for the updater.
• Phases: clustering, modules, parent_modules, overview, update.
• metadata.json gets a usage block (totals, by_phase, unpriced_models, calls) and generation_info.provider.
• CLI summary prints tokens used and an estimated cost (marked API-equivalent for subscription providers). Models with tokens but no price show as "cost unknown" instead of $0, with the caw pricing set hint for Codex.
• caw pin moved to main, which includes Price codex runs on the default model and newer OpenAI models anhnh2002/caw#2: Codex is priced on the model it actually runs (-m or ~/.codex/config.toml), newer OpenAI models fall back to the opencode table, and gpt-6.1-sol is priced.
• Docs: "Cost tracking" section in guides/providers.md.

Known gap

In API-key mode, pydantic-ai sub-module agents' usage isn't counted. Counting it would mean passing usage=ctx.usage, which also makes the parent and its sub-agents share one request limit.

Testing

• New tests/test_usage_ledger.py (9 tests): summing, phases, unpriced detection, thread safety, CawBackend recording, metadata block, CLI output.
• Full suite: 339 passed. 1 failure (test_clustering_skip) also fails on main.
• End-to-end with claude-code (Haiku) on a 2-file repo: the CLI printed 125,446 tokens and $0.08, and metadata.json recorded the same.
• Live Codex call through CawBackend with no main_model: resolved gpt-6.1-sol and priced it.