feat(author): drop thin edition clusters from catalogue sync (#2235)

[gchahcg]

Summary

Metadata profiles gain an opt-in minimum edition count filter for author catalogue syncs: works whose title cluster reports fewer known editions than the profile's floor are skipped before they become book rows. Closes #2235.

This is the smaller, safer shape of the request from the issue thread — a per-profile number, not a global catalogue-wide policy — and it leaves the closed PR #2628's cluster-signalling approach behind.

Implementation notes

Aspect	Behaviour
Cluster key	Normalised (lowercased, trimmed) work title
Cluster count	Max of the known EditionCount across the works in the cluster — one well-editioned work saves the title's thinner siblings
Unknown count	Passes the filter ("unknown is not zero", same rule as MinPages). Critical for Hardcover-primary authors, whose records never carry an edition count; a sum-based rule would have emptied their catalogues
Already-tracked books	Maintained in place, never dropped, never counted as skipped
Single-work adds	Exempt, matching the #1612 rule that explicit user picks are not vetoed by catalogue heuristics
Default	0 = filter off; the cluster map is only built when the profile sets a floor, so the default is zero-cost

• Migration 094_metadata_profile_min_edition_count.sql adds min_edition_count INTEGER NOT NULL DEFAULT 0 to metadata_profiles.
• The skip is reported in the sync summary (SkippedThinCluster + sample, capped at 5) and surfaced as its own line in the author page's sync notice.

Where it's mounted

• fetchAuthorBooks (internal/api/authors.go): cluster map built in the candidate loop, filter applied in the create loop after existing-book resolution.
• Profile CRUD: internal/db/metadata_profiles.go, internal/models/metadata_profile.go.
• Web: profile editor form + profile-list chip (MetadataTab.tsx), sync notice breakdown (AuthorSyncNotice.tsx), en i18n keys (other locales fall back to en).

Follow-ups (not in this PR)

• The issue's larger ask — a global, catalogue-wide thin-title policy and the two-stage cluster signal from feat(filterengine): opt-in cluster-based noise filtering preset (#2235) #2628 — can land separately once this shape is proven in practice.

Checklist

• Tests added or updated — 4 new fetchAuthorBooks tests (drop, opt-in default, owned-book exemption, single-work exemption) plus the twin-title fixture fix in EditAuthorModal.test.tsx
• Doc-update gate cleared — docs/User-Guide-Wiki.md profile-filter list
• Wiki pages updated if user-facing behaviour changed

Test plan

• make test (full Go suite, CI gate)
• gofmt -l ., go vet ./..., golangci-lint v2.11.4, govulncheck
• make smoke (real binary, exercises the new migration)
• cd web && npm run typecheck && npm run build && npm test (1002 tests) and npm run lint (0 errors; remaining warnings pre-existing)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[vavallee]

Heads up on a migration number collision, not a review.

This PR adds a migration numbered 086, and so does #2673 or #2633, whichever this is not. Both are open and both claim it. assertUniqueMigrationVersions (internal/db/db.go:685) refuses to boot when two files share a numeric prefix, so whichever merges second breaks startup for anyone with both.

086 is a genuine gap on main, which goes 085 then 087, and it has never been filled by a merged commit. Three separate branches have each created their own 086 at some point, which is how the gap became attractive to all of them at once.

Main tops out at 089. Current claim map across the open PRs:

• 086 → feat(author): drop thin edition clusters from catalogue sync (#2235) #2673, Feature/download client media eligibility #2633
• 090 → feat(api): auto-approve requests per requester account (#2718) #2761, feat(indexer): normalise the audiobook size bonus by runtime and cap grabs (#2740) #2750, feat(downloader): remove the torrent after a successful import (#2046) #2713
• 091 → fix(scheduler): stop a failed queue row blocking an automatic re-grab (#2710) #2714
• 092 → feat(quality): rank releases by profile order, best first (#2733) #2737
• 093 is the first genuinely free number

Suggest renaming to 093 or above rather than reusing the gap, and picking the number by checking open PRs rather than main. The reason this keeps happening is that each branch computes "next free" against a main that excludes everyone else's in flight work.

[gchahcg]

Renamed to 093_metadata_profile_min_edition_count.sql — checked all 33 open PRs, 093 is genuinely free (086 was also claimed by #2633, and 090/091/092 by #2761/#2750/#2713/#2714/#2737). No other file referenced the old number. Force-pushed.

[vavallee]

My fault, and a correction to my own advice on this PR.

I told four PRs that 093 was the first free number. Two of you took it, so #2673 and #2761 now both claim 093, which is the same collision I was flagging. Telling several branches one free number just moves the crash.

Assigning distinct numbers instead, so nobody has to guess:

Migration	PR
086	#2633
090	#2713
091	#2714
092	#2737
093	#2761
094	#2673, this one
095	#2750

So please rename to 094_metadata_profile_min_edition_count.sql. I picked 093 for #2761 only because it renamed first; there is nothing better about its claim.

One other thing on this PR: the body says 086 while the file says 093. Those need to agree whichever number it ends on.

Note for anyone reading this later: 086 is a real gap on main, which runs 085 then 087, and it was never filled by a merged commit. The apply loop keys on the filename number rather than on "greater than last applied", so a back filled 086 does apply cleanly to existing installs. That is why #2633 can keep it.

[gchahcg]

Renamed to 094_metadata_profile_min_edition_count.sql and fixed the PR body to match (it still said 086). Force-pushed.

[vavallee]

This is the shape we agreed on #2235, and the unknown passes rule is the part that matters most: Hardcover primary authors never carry an edition count, so a sum or zero based rule would have emptied their catalogues. I checked it's pinned by turning a missing count into a zero, and `TestFetchAuthorBooks_SkipsThinCluster` fails. Merged onto current main it builds, vets, and `db`, `models` and the catalogue sync tests pass. 094 is unique across main and every open PR.

Two small things before merge:

• Changelog fragment. Every PR needs one under `changelog.d/` (the release notes are assembled from them). Something like `changelog.d/2235-min-edition-count.md` with an `### Added` bullet. House style is no dashes in the text.
• One line in the docs about series fill. Every metadata profile filter, this one included, is skipped by the series fill path today. That's #2208 and not yours to fix, but someone who sets a floor and then fills a series will see thin works arrive and think the filter is broken. A sentence in the User Guide section you added saves that support question.

Tiny nit, take or leave: the cluster key is normalised inline in two places (building the map and reading it). They match today; a small helper would keep them from drifting.

[vavallee]

Pushed two small follow ups so this can make the next release: a changelog fragment (changelog.d/2235-min-edition-count.md), and a sentence in the user guide that series fill skips every metadata profile filter today, this one included (#2208). No code changes. Thanks gchahcg!