perf: count universes and works without materializing every record (#2729)

[atomantic]

Summary

The Wordsmith skill scored Create/Writers Room engagement with listUniverses().length + listWorks().length — two whole-collection reads for a tally. listUniverses() selects and sanitizes every universe JSONB record and loads the entire universe-run history before sorting; listWorks() runs two queries to rebuild every work's manifest with its drafts[]. That cost is paid on every explicit /character read and every saveCharacter() response, and it grows with the user's library.

Adds countUniverses() / countWorks() across both backends (Postgres + the file escape hatch), threaded through the existing store facades, and points Wordsmith's compute() at them. The skill's value is unchanged for every existing case.

Finding: is universes.deleted authoritative and in sync with the JSONB flag?

Yes — they cannot diverge, so WHERE deleted = FALSE is exact. This was the issue's crux, so stating the verification explicitly:

• listUniverses() filters !u.deleted on the sanitized record, where the flag comes from sanitizeSoftDeleteFields (server/lib/syncWire.js) — computed as raw?.deleted === true.
• The universes table has exactly two writers, and each binds the column in the same statement that writes data, from the same record object, with the identical predicate record.deleted === true:
  • writeRaw (server/services/universeBuilder/db.js)
  • the one-time file import (server/scripts/migrateUniversesToDB.js)
• There is no third writer and no UPDATE that touches deleted without rewriting data (verified by grepping all SQL against the table).

So the column is an exact mirror, and it additionally hits the idx_universes_live partial index that exists for precisely this "live set" scan. No divergence to fix.

Two related correctness points the count had to get right:

• Tombstones are excluded — the specific bug a naive (await listIds()).length would introduce, since listIds() is documented as returning live + ephemeral + tombstones and leaving filtering to the service. Pinned by tests on both backends.
• Ephemeral universes are still counted — listUniverses() filters only on deleted, so filtering ephemeral here would undercount. Pinned by a test.

Known boundary (documented in code, deliberately not replicated in SQL): a row whose data can't survive sanitizeTemplate (blank name, non-string id) is dropped by listUniverses's .filter(Boolean) but still counted. It's unreachable through the service — create/update/sync all sanitize before writing — and only a corrupt legacy record imported verbatim could reach it. Replicating the sanitizer's drop rules in SQL would couple the count to sanitizeTemplate's internals and silently drift when they change; a ±1 on an already-corrupt record isn't worth that. Both backends share these semantics, so they agree with each other.

writers_room_works.deleted needed no such analysis: it's a real column (not a JSONB mirror) and it's the same predicate listWorks() already filters on.

Step 5 (saveCharacter(data, { withSkills })) — evaluated, deliberately not shipped

The issue asked to consider it. It has no beneficiary, so the option would be dead weight:

• Every caller of saveCharacter() (addXP, takeDamage, takeRest, addEvent, syncJiraXP, syncTaskXP, updateCharacterFields, /reset) originates from server/routes/character.js — there are zero background/service callers (verified by grep).
• Every one of those routes returns the character to client/src/pages/CharacterSheet.jsx, which does setChar(result.character). Passing withSkills: false anywhere would blank the skills card, so every call site would pass true.
• The performance motivation also largely evaporates with this PR: skill derivation is now a handful of COUNT(*)/stat reads, not a full library scan.

enrichCharacter already accepts { withSkills } (from #2725), so the seam remains if a non-UI caller ever appears.

Test plan

• server/services/universeBuilder/db.test.js — new: countUniverses counts the live set, excludes the tombstone while listIds() still returns 3, counts ephemeral, agrees with the JSONB-flag filter listRaw() applies, includeDeleted variant, and follows an un-delete. Ran against portos_test via the DB config (verified executed, not skipped).
• server/services/writersRoom/db.test.js — new: countWorks agrees with listWorks().length and excludes a tombstone that listWorkIds({ includeDeleted: true }) still returns. Same, against portos_test.
• server/services/universeBuilder.test.js — new (file backend, service level): countUniverses agrees with listUniverses().length before and after a soft-delete, and includeDeleted mirrors listUniverses({ includeDeleted: true }).
• server/services/writersRoom/local.test.js — new (file backend, service level): countWorks agrees with listWorks().length across a soft-delete.
• server/services/characterSkills.test.js — mocks updated to the count helpers; all 27 pass, Wordsmith's values unchanged.
• Full server suite: 20092 passed / 200 skipped. The 2 failures in services/installState.test.js (install-root resolution) are pre-existing on origin/main and untouched by this diff — confirmed by stashing this branch's changes and re-running them red on a clean tree.

No DB guard was weakened; the DB-backed suites ran only via the portos_test config.

Closes #2729