Caching and request deduplication #2594

Parent: #2592 — Reduce Hardcover API usage and support daily quotas

Pre-flight

• Searched existing issues; related reports are referenced by the parent.
• Checked the roadmap; this cache/deduplication scope is not already planned there.

Problem

SearchAuthorsWithOutcome and SearchBooks in internal/metadata/aggregator.go fan out to providers without server-side search caching. Metadata cache get/fetch/set sequences do not combine identical concurrent misses. GetEditions and GetEditionsFromProvider use different keys even when they resolve to the same Hardcover book.

The existing 24-hour cache is bounded but in memory, so restarts discard it. These redundant calls consume a plan-dependent daily allowance: see Hardcover Supporter.

Proposed solution

• Add bounded, short-lived provider-level search caching so ordinary searches and enrichment can reuse equivalent successful responses. Include query options and any relevant account/provider configuration in cache identity.
• Combine identical in-flight requests so concurrent misses perform one upstream fetch. Preserve cancellation semantics and avoid sharing mutable cached results unsafely.
• Resolve the provider before canonicalizing edition cache keys. Unify equivalent routed and explicit-provider lookups without confusing numeric IDs with slugs or mixing providers/accounts.
• Reuse existing metadata caching where practical. Evaluate persisted edition freshness for restart reuse; preserve an explicit refresh path and avoid treating partially populated local editions as a complete upstream snapshot.
• Cache legitimate empty results according to policy, but do not cache rate-limit/auth/network failures as successful empty data. Invalidate affected state on configuration changes.

Acceptance criteria

• Repeated equivalent searches and concurrent identical misses reduce counted upstream requests.
• Equivalent edition lookup routes share cached results; different providers, identities, and account scopes remain isolated.
• Cache bounds, expiry, cancellation, configuration changes, empty results, and failures have focused coverage.
• Existing search ranking, enrichment, edition completeness, and explicit refresh behavior are preserved.

Additional context

Covers review items 3 + 5. Coordinate with daily request accounting so cache hits are not charged locally. Related #2173 concerns skipping unchanged ABS items and is not replaced by this cache work.