Harden wanted-search language behavior for normalization work

[mjc]

Summary

This PR hardens wanted-search and adaptive-search language handling before the wanted-search storage model changes in the performance work.

It keeps the current serialized wanted-state model in place, but makes the existing movie and episode wanted-search paths stricter and more predictable around malformed language tokens, forced/HI flags, audio-language fallbacks, and failed-attempt updates.

What changed

• Rejects malformed wanted-language tokens instead of carrying unknown flags into provider searches.
• Normalizes forced/HI language flags consistently across wanted movie, wanted episode, and adaptive-search paths.
• Handles mixed-case flags, multi-colon input, empty base-language tokens, and legacy wanted rows predictably.
• Handles malformed or missing audio-language rows without crashing wanted-search planning.
• Keeps failed-attempt updates chained correctly across multiple requested languages.
• Preserves missing scene names as None through wanted-search payload construction.
• Moves wanted/adaptive edge-case coverage onto the shared database-backed fixture layer from PR Prepare wanted-search tests and CI for normalization work #3383.

Why this is separate

The normalized wanted-state PR relies on these language and retry semantics being pinned down first. Keeping this as a hardening PR lets the later storage/query changes focus on performance and schema shape instead of mixing in behavior fixes for malformed wanted-language data.

Tests

The added and updated tests cover malformed wanted-language tokens, forced/HI normalization, duplicate and mixed-case flags, multi-language retry updates, audio-language fallback behavior, missing scene names, scheduled wanted-search filtering, provider wrapper behavior, and adaptive-search property cases. The tests use the shared transactional fixtures and mock only provider/search side effects.

[mjc]

Note on the diff shape: this PR is intentionally opened against development as a draft so it can be reviewed before PR #3383 merges. It builds directly on #3383, so GitHub currently shows the combined/flattened diff. Once #3383 lands, I'll rebase this branch onto the updated development branch and the diff should collapse down to just the wanted/adaptive language hardening changes from this PR.

[mjc]

I rebased this branch on the updated shared wanted-search fixture layer from #3383 and kept the wanted/adaptive behavior coverage here.

This branch owns the follow-up behavior tests for:

• malformed or invalid missing_subtitles values failing safe
• language token normalization, deduping, casing, whitespace, and multi-flag parsing
• adaptive retry filtering and failed-attempt stamping order
• malformed audio profile rows resolving to no audio language instead of crashing
• missing media paths skipping generate_subtitles
• non-integer/unknown series season/episode values in scheduled-search progress messages
• scheduled-search empty/throttled/completed paths and provider refresh between items

I also adjusted these tests to use local worker helpers instead of reintroducing the shared wanted_worker fixture, so #3383 can stay focused on the common fixture/bootstrap layer.