feat: cache-hit-ratio analyzer with underutilization warnings

[jakduch]

Glad to see this fork is actively maintained — the original phuryn/claude-usage had stopped accepting PRs (30+ open). Switching here solved real problems; more contributions on the way.

What does this add and why do you believe it belongs in this dashboard?

Adds a cache-hit-ratio analyzer so a personal Claude Code usage dashboard can finally answer the single question that drives most of the bill: am I actually getting prompt caching, and where am I leaving money on the table?

For every session we compute one number, cache_read / (input + cache_read) — the overall share of "fresh" prompt context that was served from cache instead of paid input — and bucket it into low / medium / high. The dashboard exposes a global cache_hit_summary (overall effectiveness, avg_ratio_pct, and counts by category) in /api/data, surfaces an insight card under the stats row, and tags individual sessions in Recent Sessions with a cache underused badge when they pull >50k input tokens but hit cache <30% of the time. Those are exactly the sessions paying 2–3× the necessary price for repeat context, and today they're invisible in this dashboard. With this card a user can immediately see "my big sessions aren't caching — restructure prompts here" without leaving the page. It fits the dashboard's existing insight-card pattern (next to the Pareto card) and adds no dependencies.

Checklist

Code correctness

• All calcCost() calls pass 6 arguments: (model, inp, out, cache_read, cache_creation, cache_1h) — no calcCost() calls added or modified
• JavaScript template literals use bare backticks (`), not escaped ones (\`)
• No JS variables referenced before they are defined (renderCacheHit is a hoisted function declaration; uses pre-existing fmt / rawData)
• No new third-party dependencies introduced

Tests

• python3 -m unittest discover -s tests -v — all passing on this branch (any test_health socket-bind flakiness is pre-existing on main and unrelated to this PR)
• python3 -m unittest tests.test_browser -v — all passing (6/6)
• New behaviour is covered by at least one test — 15 new tests in tests/test_cache_hit.py cover ratio math (including None inputs), bucket boundaries, summary aggregation, configurable input threshold, HTML wiring, and an end-to-end check through get_dashboard_data against a real SQLite schema

Scope

• This is a single concern — one feature or fix per PR (cache-hit-ratio analyzer only)
• Only touches existing files (dashboard.py, scanner.py, cli.py, pricing.py, cowork.py, tests/) — or I've explained below why a new file is needed: tests/test_cache_hit.py is added under tests/ per the existing one-file-per-feature test convention. .github/PULL_REQUEST_TEMPLATE.md was restored from main after an accidental delete on this branch.