[dny238]

Relates to #1068. Two small, independent search-query changes.

1. Empty-query guard (safety)

A query that sanitizes to an empty string — dash-only or punctuation-only input — must never be dispatched: some indexers treat an empty query as "browse everything" and would return unrelated releases that could then be grabbed. IndexerSearchWorkflow now fails closed on an empty sanitized query (logs and returns no results without calling the provider). Covered by IndexerSearchWorkflowEmptyQueryTests (dash-only, collapsing-dash, punctuation-only, whitespace-only all assert the provider is never called and nothing is returned).

2. Hyphen/dash normalization (Prowlarr/Sonarr-style cleaning)

Hyphens and the Unicode dash variants are normalized to spaces, so a hyphenated title token like Full-Cast is searched as separate words (Full Cast) — consistent with how Prowlarr/Sonarr clean query terms, and defensive against any indexer or proxy that interprets a leading - as an operator. Covered by IndexerQuerySanitizerTests.

Note on #1068's root cause

I ran a read-only comparison against NZBgeek's API. De-hyphenating did not change which releases match — hyphen and space queries returned identical counts in every pair, and a direct -token probe did not exclude. The 0-result reports in #1068 track query specificity and the appended author instead: …Full Cast Edition returned results, but appending J.K. Rowling dropped it to 0 (release names don't carry the author), and adding more title words AND-matched to nothing.

So this PR is a normalization + safety change rather than the fix for #1068's symptom. The actual fix — a title-only / drop-the-author fallback when a query returns nothing — is a sensible follow-up (and matches the reporter's own suggestion). Keeping this open as relates to rather than fixes so #1068 stays open for that work. (Caveat: the report's path is via Prowlarr, which I bypassed by hitting NZBgeek directly, so Prowlarr-side - handling isn't ruled out — but against the indexer's own API the hyphen change is a no-op.)

🤖 Generated with Claude Code

[therobbiedavis]

@dny238 Looks good, just a couple things I'd like you to add to tighten up this PR.

I’d still check the split query against NZBgeek/Prowlarr with a hyphenated title, author and ISBN. The container checks verified the outgoing requests and download handling, but the fixtures returned results regardless of the query, so they don’t tell us whether splitting the words changes which releases match.

I’d also handle queries that become empty after sanitization here, with a negative test confirming they can’t trigger an unrelated download.

[therobbiedavis]

Whoops forgot to leave it as a code review

The two inline comments are in my earlier review.

[dny238]

Thanks — addressed both.

Empty-after-sanitization guard: IndexerSearchWorkflow.SearchIndexerAsync now fails closed when the sanitized query is empty — it logs and returns no results without calling the provider, so a dash- or punctuation-only query can't trigger a browse-everything grab. Added IndexerSearchWorkflowEmptyQueryTests: dash-only, collapsing-dash, punctuation-only and whitespace-only queries all assert the provider is never called and nothing is returned (the mock provider returns an unrelated release that must never surface).

Indexer comparison / ISBN: I can't drive a live NZBgeek/Prowlarr from here, so I can't add an assertion that the split query returns the same releases. A couple of things that bound the risk, though:

• The indexer free-text query is title + author only. ISBNs don't reach this sanitizer — they're used in the Audible metadata path (AudibleAuthorSearchWorkflow.FilterByIsbnAsync), matched against the exact ISBN string, so hyphen-splitting here doesn't touch ISBN matching.
• Space-separation (rather than removal) matches how Prowlarr/Sonarr normalize hyphens, and the reported symptom is exactly the -Cast exclusion this removes — Full Cast is what returned the releases in the issue's own manual tests.

If your container harness can vary fixtures by query, a title/author comparison there would be the definitive check; I just can't reproduce a live indexer in this environment. Happy to adjust if you'd prefer a different handling for any specific token.

[dny238]

Ran the indexer comparison you asked for — read-only against NZBgeek's API — and it changed my read of this PR, so I've reframed it (and updated the title/description).

De-hyphenating made no difference. Every hyphen-vs-space pair returned identical counts:

query	hyphen	space
…Stone Full-Cast Edition	5	5
…Philosophers Stone Full-Cast (reporter's string)	0	0
Harry Potter Full-Cast	16	16

And a direct probe — Harry Potter Cast = 16 vs Harry Potter -Cast = 16 — shows NZBgeek's API isn't treating -token as an exclusion. So the "Newznab exclusion" premise doesn't hold against the indexer directly, exactly as you suspected.

What actually drives the 0 results: the appended author and query specificity. …Full Cast Edition returned 5, but adding J.K. Rowling → 0 (release names don't carry the author); and Harry Potter Full Cast = 16 while Harry Potter Philosophers Stone Full Cast = 0 (the AND of all those words matches nothing).

So I've demoted the hyphen change to Prowlarr/Sonarr-style query normalization (no exclusion claim) and kept it as relates to #1068 rather than fixes. The empty-query guard stands on its own. The real fix is the title-only / drop-the-author fallback you mentioned — I'd like to do that as a follow-up on #1068.

Caveat: the report's path is via Prowlarr and I hit NZBgeek directly, so I can't rule out Prowlarr-side - handling — but against the indexer's own API this is a no-op, so I didn't want to claim it fixes the symptom.

[therobbiedavis]

Looks good to merge. The empty-query guard is verified, and the real indexer comparisons showed no lost matches from splitting hyphens in the tested titles and author.

The updated description also reflects what we found: this is a normalization and safety change. #1068 stays open for the title-plus-author search problem, with the fallback work tracked in #825.

[therobbiedavis]

@dny238 Please fix the test failures.

[dny238]

Fixed — the unit-tests failure was format:backend:check (dotnet format --verify-no-changes) tripping on IDE0005 unused-using directives in the two new test files, not a failing test. Removed the redundant usings (the types resolve through the test project's global usings); format:backend:check now passes clean and the 17 tests across both classes are green. Pushed as 289591c.