Add session cost reports, live TUI visibility, and cache economics

Parent

Part of #610. Depends on enriched pricing and complete LLM-call attribution.

User outcome

An operator can obtain a trustworthy, machine-readable cost summary for a session and see the same essential information live in mecatui without needing a separate analytics product.

Scope

Cost-summary query

Expose a session-scoped report containing:

• total estimated cost and currency;
• pricing coverage percentage and explicitly unpriced usage;
• estimate/price provenance and price-card freshness;
• breakdown by functional label;
• breakdown by requested/resolved/served model where meaningful;
• successful, failed, retry, and fallback spend;
• ordinary input, output, reasoning, cache-read, and cache-write usage/cost;
• top deterministic cost contributors.

Provide a stable machine-readable representation suitable for scripts and later external dashboards. Aggregation must be reconstructable from durable usage observations rather than an in-memory-only counter.

Basic cache economics

Report only the initial high-value indicators:

• cache hit rate;
• cache-read cost;
• cache-write cost;
• estimated uncached input baseline;
• estimated net cache savings.

The calculation must document provider-normalized semantics, price provenance, partial coverage, and where an uncached baseline is only an estimate. It must not double-count cache tokens already included in inclusive input totals.

Minimal TUI surface

Add a compact session-cost view/live meter showing:

• current estimated session cost;
• last-call increment when known;
• pricing coverage/unpriced warning;
• top function/model contributors;
• cache hit rate and estimated savings.

A table or expandable pane is sufficient; charts and historical trends are not required.

Acceptance criteria

• A completed or active session has an obtainable machine-readable cost report reconstructed from durable observations.
• The report totals reconcile with the underlying observed calls and do not double-count delegation.
• Unknown pricing is visible in both amount coverage and token/call counts; totals never imply 100% coverage when usage is unpriced.
• Every displayed amount is labeled estimated and includes its price provenance/freshness.
• Failed/retried spend and requested-versus-served model attribution are inspectable.
• Cache economics handle inclusive input/cache subsets correctly and expose the formula through tests/docs.
• The TUI updates after completed model calls and degrades cleanly when no pricing or usage exists.
• Cost-report content contains no prompts, tool payloads, credentials, or sensitive provider configuration.
• User-facing documentation explains list-price estimates, coverage, cache-savings assumptions, and limitations.
• Tests are offline and include mixed priced/unpriced, cache, retry, fallback, and nested-delegation fixtures.

Out of scope

• Organization/repository historical dashboards.
• Detailed cache-miss root-cause diagnosis.
• Cost forecasts.
• Billing exports or invoice reconciliation.
• Cost per PR or business outcome.

Start by reviewing parent issue #610 and tracing the durable usage observations that can support a session-scoped report. Then examine the mecatui session view and existing offline test structure. Done means reconcilable machine-readable totals, documented cache and pricing assumptions, safe TUI updates, and coverage for the listed mixed-usage scenarios.