Two ways a cached entry could be wrong.
The key was built by string concatenation in parallel with the URL, so
`search_movies("dune&year=2024", None)` and `search_movies("dune",
Some(2024))` produced the same key for two different requests. Both key
and URL now come from one `Url`, percent-encoded identically, and the
API key is appended at send time so it cannot reach a key or a log.
A successful response was cached before it was decoded, so one malformed
body returned `Error::Decode` from the cache for the whole time-to-live.
Decoding now happens first and only what decoded is stored.
Search and detail lookup, carrying the two fields the policy engine
needs and that are easy to lose at the edge.
`original_language` arrives as bare ISO 639-1, which has one code for
both Portuguese variants, so the dub rule in §5.2 cannot be expressed
against it alone. `Movie` therefore also carries `origin_countries`,
from `origin_country` where TMDB has it and `production_countries`
otherwise, which is what separates a Brazilian film from a Portuguese
one.
Digital release dates come back appended to the detail call rather than
costing a second round trip, reduced to the earliest type-4 entry across
every country: a release existing in one region is a release that exists
on the indexers. `Movie::is_digitally_released` puts the §6.2 gate in
one tested place.
Responses are cached in process for a day, matching the metadata refresh
tick. Failures are never cached, so an outage cannot pin a title into a
bad state until tomorrow.
Tests are wiremock-backed, per §12.
Eleven empty crates per DESIGN.md §11, versions and lints pinned once in the
root manifest, Justfile carrying the §12 gate, nix dev shell.
The frontend steps in `just ci` no-op until web/ exists. arr-core carries one
placeholder test because `cargo nextest` fails a zero-test run, and relaxing
that would let a broken filter go green later.
Closes#1
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>