fix(indexer): run the #2502 relevance guards on the search paths that grab

[vavallee]

Summary

The relevance filter existed as two copies, filterRelevant and filterRelevantDebug. #2502 added its two guards (conflictingTitleAuthor and the title identity phrase gate) to filterRelevant only, and pinned them with tests that call filterRelevant directly.

Production never runs filterRelevant. The scheduler calls SearchBookWithOutcomes, which projects SearchBookWithDebug, and the interactive search calls SearchBookWithDebug. Both run filterRelevantDebug. filterRelevant is reached only by the scheduler's fallback for searchers that cannot report outcomes, which in practice is test stubs.

Since this was drafted, #2863 copied the author guard across by hand (shipped in v1.39.0), and #2921 added a volume number guard to both copies. The title identity gate was never copied, so on current main the production path still keeps:

Release	For	filterRelevant	production path on main
The Power of Writing It Down by Ben Coes	Power Down	dropped	kept
Beyond Order: 12 More Rules for Life	12 Rules for Life	dropped	kept
Rising of the Shield Hero Volume 16	... Volume 17 (same spelling)	dropped	kept

The last row is the divergence #2921's author flagged.

Change

filterRelevantDetailed is now the single implementation. It returns the kept results plus a reason for every drop, and filterRelevant and filterRelevantDebug are thin wrappers over it, so a guard can't land in one path and miss the other again. #2863's author guard, #2921's volumeNumberAgrees, #2763's elision fallback and the identity gate all live in that one function. The drop record tells the interactive panel which check fired: a different author for this title, title words present but with other words or numbers among them, or the plain keyword miss.

Zero padding

Putting the identity gate on the production path at first dropped correct releases that pad a number ("Vol 07" for "Volume 7", "Book 07" for "Book 7"), because the gate compared numbers as literal words. Fixed in this PR (cc9d92b):

• keywordPattern renders an all digit keyword as 0* plus the number without padding, so the phrase, keyword and in order matchers and ContainsPhrase (the identity gate) treat "7", "07" and "007" as one number, both directions. Word boundaries keep "7" from matching "17" or "70". volumeNumberAgrees already ignored padding.
• conflictingTitleAuthor compares whole normalized strings, so foldVolumeMarkers became foldTitleTokens and strips padding too. Without that, "Dungeon Crawler Carl Book 07 - Some Other Writer" slips past the attribution check (a test pins it).

Remaining known gap

A digit and a spelled out number are still different words: "The Seven Habits of Highly Effective People" is dropped for "The 7 Habits of Highly Effective People", which production keeps on main. Pinned with a KNOWN GAP comment in TestFilterRelevantNumberPadding. No word to number equivalence here on purpose.

Comparison against main's production filter

271 distinct (title, author, release) triples: everything the indexer, api and scheduler suites feed the filter, plus 67 hand written realistic release names. Each was run through filterRelevantDebug on origin/main and on this branch. Differences:

Title	Release	main	this PR
Power Down (x2)	The Power of Writing It Down ...	kept	dropped	wrong release
12 Rules for Life (x2)	(Beyond Order:) 12 More Rules for Life ...	kept	dropped	wrong release
Shield Hero Volume 17	... Volume 16	kept	dropped	wrong volume
Shield Hero 7	... 17 / ... 70	kept	dropped	wrong volume
Dungeon Crawler Carl Book 7	... Book 17 / ... Book 017	kept	dropped	wrong book
Dungeon Crawler Carl Book 7	... Book 07 - Some Other Writer	kept	dropped	wrong author
Agent 007	Some Author - Agent 7	dropped	kept	right release
Fahrenheit 451	... Fahrenheit 0451	dropped	kept	right release (contrived)
The 7 Habits of Highly Effective People	... The Seven Habits ...	kept	dropped	right release, known gap

Every other triple, including the three padding rows, is decided the same way as on main.

Tests

• fix(indexer): reject conflicting release title and author evidence #2502's two tables run through both wrappers.
• TestRelevanceGuardsReachEveryProductionEntrypoint drives SearchBook, SearchBookWithDebug and SearchBookWithOutcomes against one fake indexer serving wrong releases beside the right ones; it now includes the same spelling Volume 16 case.
• TestFilterRelevantDebugDropReasons pins the reason text for each check.
• TestFilterRelevantNumberPadding (both wrappers) and TestContainsPhraseNumberPadding: padded and unpadded numbers match both ways; 17, 70, 017 and Volume 16 do not; a padded release naming another author is dropped; 7 versus Seven is the known gap.
• Fail before for the padding commit: with release.go and title_identity.go at the previous commit, the three padding rows, the reverse and triple padded cases, the fix(indexer): match "Volume" in a title against "Vol" in a release name #2921 zero padded case and the matcher test fail. Reverting only title_identity.go fails the other author case.
• TestFilterRelevantVolumeMarkerSpelling (from fix(indexer): match "Volume" in a title against "Vol" in a release name #2921) had separate wantPlain and wantDebug columns documenting the divergence. Changed deliberately: they collapse to one want per case, so the test fails if the two wrappers ever disagree. For production, same spelling Volume 16 for Volume 17 is now dropped (intended); zero padded Vol 07 for Volume 7 stays kept.
• Fail before: with main's debug.go and searcher.go restored (plus a test shim for filterRelevantDetailed), the entry point test fails for Power Down, 12 Rules and Volume 17 on both SearchBookWithDebug and SearchBookWithOutcomes, while SearchBook passes, which is the blind spot. The identity tables, the reason test and the two flipped volume cases fail too.

Checklist

• Tests added or updated
• Doc-update gate cleared (behaviour now matches what docs/Troubleshooting-Wiki.md already describes for fix(indexer): reject conflicting release title and author evidence #2502)
• Changelog fragment (changelog.d/2812-relevance-filter-single-implementation.md)

Test plan

• go build ./... && go vet ./...
• go test ./internal/indexer/... ./internal/api/... ./internal/scheduler/... -count=1
• golangci-lint run ./internal/indexer/...

🤖 Generated with Claude Code

https://claude.ai/code/session_016fJcCVbNnKsmj2MAAMwWj9

[codecov]

Codecov Report

❌ Patch coverage is 93.75000% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/indexer/title_identity.go	75.00%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[vavallee]

Rebased onto main now that #2763 has landed. Its elision fallback is folded into filterRelevantDetailed, #2763's own elision tests pass against the single implementation, and the entry point test still shows 8 wrong releases kept on current main without this change.

[github-actions]

The root-cause analysis is confirmed by the diff: the deleted filterRelevantDebug body computed fullOK = tryMatch && volumeNumberAgrees — no identityOK — while filterRelevant had all three. The unification approach is the right fix; the PR description, test design, and changelog are all solid.

One blocker before merge:

Zero-padding regression — the PR flags it and asks for a decision, so this review is that decision point. titleIdentityWords treats "7" and "07" as different tokens, so "Dungeon Crawler Carl Book 07" is now dropped for "Book 7", "Shield Hero Vol 07" for "Volume 7", etc. Zero-padded volume numbers are extremely common in Usenet release names (scene-style zero padding, SABnzbd defaults). The PR correctly scopes the fix: normalising leading zeros in titleIdentityWords would close it and the test case is already marked KNOWN GAP / want: false. I'd rather see that normalization land here than ship a regression that silently stops grabbing a class of correct releases. The want: false row in volume_marker_test.go:109 is the concrete marker.

Everything else looks good:

• filterRelevantDetailed is the correct single-implementation shape; filterRelevant and filterRelevantDebug are clean thin wrappers.
• continue added at searcher.go after filtered = append(filtered, r) is necessary — without it a kept result would also push a drop record. Correct.
• TestRelevanceGuardsReachEveryProductionEntrypoint is the right kind of test: drives real entry points, not just the filter functions in isolation. This is what was missing from the #2502 test suite.
• TestFilterRelevantDebugDropReasons pinning reason strings is useful for the interactive panel.
• relevanceFilters table in title_identity_test.go / TestFilterRelevantVolumeMarkerSpelling collapsing to a single want is an improvement — the split columns were documenting a bug, not an intended design divergence.
• No security concerns; pure filter logic with no auth/tenancy surface.

— 🤖 Bindery triage bot (automated). Reply to correct me; a human will see it.