fix(tui-gateway): expose live cache token totals

[worlldz]

Summary

• expose the agent's existing session cache-read and cache-write counters through the shared TUI gateway usage snapshot
• add the corresponding optional fields to the Desktop UsageStats contract
• cover the gateway mapping with a regression test

Why

The agent already accumulates session_cache_read_tokens and session_cache_write_tokens, but _get_usage() drops both values. Every live consumer of that shared snapshot—including message.complete, session.usage, session.info, and periodic usage ticks—therefore loses cache visibility.

This addresses the cache-token portion of #87769. It deliberately does not add cost or duration:

• user-facing provider cost estimates were previously removed as a product-policy decision and should not be reintroduced incidentally
• Desktop already measures turn duration from message.start to message.complete

Impact

This is an additive wire-format change. Existing clients can ignore the new keys, while Desktop and other gateway consumers can now calculate cache-hit metrics from canonical session totals.

Root cause

_get_usage() maps the agent's input/output/reasoning counters but omitted the adjacent cache counters, even though both are maintained by the same canonical usage-accounting path.

Checks

• scripts/run_tests.sh tests/test_tui_gateway_server.py — 586 passed
• scripts/run_tests.sh tests/test_tui_gateway_server.py -k get_usage — 7 passed
• npm run typecheck in apps/desktop
• npx eslint src/types/hermes.ts
• git diff --check

Related work checked

No PR currently references #87769. Older #43050 concerns persistence/restoration of counters rather than exposing cache totals from the current live TUI gateway snapshot; #67834 is a placeholder for a cost UI and does not implement this data path.

[Enough1122]

AI code review — automated review; please use your judgment.

Small, correct addition: _get_usage now surfaces cache_read_tokens/cache_write_tokens through the same g() fallback helper as the neighboring counters, the TS UsageStats type gains both as optional fields (so older agents without the attributes simply omit them), and a focused test pins the mapping on a bare agent namespace.

No blocking issues found.

— reviewer-a · automated agent review (Hermes week-review)