fix(indexer): match a "Series N: Title" book on its own title

[vavallee]

Summary

Reported on Discord: Feral Mage 4: The Mining Company Contract by Chase Kilgore never matched the MyAnonamouse release The Mining Company Contract by Chase Kilgore [ENG / EPUB] [VIP]. The search debug panel showed it dropped at the relevance stage.

The filter tries two readings of a title: the whole name, which needs "feral" and "mage", and `primaryTitle`, the part before the colon, which here is the series and position rather than the book. A release naming the book itself satisfies neither. Reproduced on main with the exact release.

Stacked on #2812 so this lands in the single filter implementation rather than being written into two copies. Retarget to `main` once #2812 merges, and merge #2812 without deleting its branch until this is retargeted, since deleting a base branch auto-closes the PR on top.

Change

`newznab.SeriesPositionTitle` returns the part after the colon when the part before it ends in a series position, and `""` otherwise:

Title	Returns
Feral Mage 4: The Mining Company Contract	The Mining Company Contract
Cradle Book 8: Wintersteel	Wintersteel
Discworld #3: Equal Rites	Equal Rites
12 Rules for Life: An Antidote to Chaos	nothing, no position
1984: A Novel	nothing, a bare number is a title
Rocky IV: The Novel	nothing, roman numerals are ambiguous with words

The relevance filter tries it as a third reading, only after the full and primary readings fail, and with the same guards as the others:

• At least two significant words, so "Wintersteel" can't turn every Will Wight release containing that word into a match.
• The phrase gate, so "The Mining Company Breach of Contract" is still dropped.
• The conflicting author check, so "The Mining Company Contract by Someone Else" is still dropped.
• The volume number guard from fix(indexer): match "Volume" in a title against "Vol" in a release name #2921, the same as the full and primary readings (added when updating onto main).

The query tier gate in the newznab client gets the same reading, so a tier whose results carry only the book's title isn't taken for a canned feed and walked past.

Tests

• `TestSeriesPositionTitle`: the shape rules above.
• `TestFilterRelevant_SeriesPositionTitle`: the reported release and its variants kept; wrong author, different book, split phrase and the Pamela Palmer release from the same screenshot dropped. Runs through both filter wrappers.
• `TestSeriesPositionTitle_ReachesProductionEntrypoints`: the reported release through `SearchBookWithDebug` and `SearchBookWithOutcomes` against a fake indexer.
• `TestTitleHasRelevantResult_SeriesPositionTitle`: the tier gate.
• Fail before: with the helper present but not wired in, every positive case, the entry point test and the gate test fail.
• Mutations, each caught by a named test: dropping the two word floor, dropping the phrase gate on this reading, and dropping the conflicting author check on it.

Update onto main (2026-10-02)

Merged the updated #2812 branch, which itself now carries origin/main (#2863, #2921), and again after #2812 gained its zero padding fix (cc9d92b), which the series reading picks up through the shared matchers. One conflict, in the drop reason: it takes #2812's new wording and still counts a series reading keyword hit. Fragment renamed to changelog.d/2813-series-position-title-relevance.md with (#2813) in the lead. Fail before re-checked on the merged tree: with SeriesPositionTitle unwired in the filter and the tier gate, every positive case in TestFilterRelevant_SeriesPositionTitle (both wrappers), TestSeriesPositionTitle_ReachesProductionEntrypoints and TestTitleHasRelevantResult_SeriesPositionTitle fail; restored, all pass. Still stacked on #2812: retarget to main after #2812 merges, and do not delete #2812's branch before then.

Not in this PR

The reporter's second point, that there's no way to grab a release the filter dropped, is #2334. The debug panel already lists what was dropped and why; making those grabbable is a separate decision.

Checklist

• Tests added or updated
• Doc-update gate cleared (no user-facing setting or doc describes the readings)
• Changelog fragment, crediting the reporter

Test plan

• `go build ./... && go vet ./...`, plus `GOOS=windows` and `GOOS=darwin` builds
• `go test ./internal/indexer/... ./internal/scheduler/ ./internal/api/`

🤖 Generated with Claude Code

https://claude.ai/code/session_016fJcCVbNnKsmj2MAAMwWj9

[codecov]

Codecov Report

❌ Patch coverage is 94.11765% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/indexer/newznab/words.go	86.66%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[github-actions]

Logic and tests look solid. A couple of small points:

Coverage gap in words.go: Codecov flags two lines. The len(head) < 2 branch of SeriesPositionTitle (line ~224) has no test case — something like {"A: Title", ""} (one word before the colon) would cover it. The pos == "" branch (when the last pre-colon word is just "#" or ".") is also missing — {"Series #: Title", ""} would cover it. Neither is a correctness problem since both paths correctly return "", but the gap means a future regression there goes undetected.

seriesOK skips the elided-apostrophe fallback: fullOK and primaryOK both run through identityOK(), which retries with the elided form when ContainsPhrase fails. The new seriesOK path (internal/indexer/searcher.go:991-993) uses ContainsPhrase directly, without that elided retry. In practice a series subtitle with an internal apostrophe is uncommon, and this is already a third-level fallback, so it isn't breaking — just noting the asymmetry.

Everything else is clean: conflictingTitleAuthor is correctly guarded on seriesTitle != "", the two-word floor and phrase gate are applied consistently, and the tier-gate in client.go gets the same reading. The "fail before / mutations" table in the PR body makes the test discipline easy to verify.

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