feat: complete reports tokens spent + saved (0.24.0)

[Shahinyanm]

What

complete now shows people what it costs and what it compresses, so the price/value is visible:

complete tj-x: retitled "#: 5" → "Voucher refund…"; closed | spent 1.5k tok ($0.0012) · saved ~88k→1.5k tok (59×)
…
Totals across 9 task(s): spent 14k tok ($0.011) · saved ~512k tok

• Spent — exact. Pulled from the backend's own usage report: the claude -p JSON envelope's usage + total_cost_usd, Anthropic/OpenAI usage. Summed across the judge call and any --enrich calls. Cost shown only when the backend reports a non-zero price (it's $0 under a subscription).
• Saved — estimate. Memory compression: the raw transcript size of the task's sessions vs its compact pack (≈ chars/4), as ~raw→pack tok (N×).
• Batch prints a Totals across N task(s): line.

How

• New LlmUsage { input_tokens, output_tokens, cost_usd } and a new LlmBackend::complete_usage method with a default that reports no usage — so the three real backends opt in by parsing their usage, while mocks and any custom backend keep working unchanged (no signature break).
• finalize::judge returns usage; LlmDreamBackend accumulates usage across enrich chunks; the CLI sums judge + enrich into FinalizeOutcome and renders per-task + batch totals.

Tests

• fmt_tokens unit; stats_suffix spent/saved formatting + empty cases.
• E2E (Unix): the fake claude envelope now carries usage/total_cost_usd, and the test asserts spent 1.5k tok ($0.0012) appears.
• Full local gate green: fmt --check, clippy --workspace --all-targets -D warnings, test --workspace, lean --no-default-features build.

🤖 Generated with Claude Code