Reduce allocation churn of events#past

[mroderick]

Summary

Reduce per-request allocation churn of GET /events/past by 15.8% (87,196 -> 73,401 GC objects on the cache-miss render) and 7.3% on the steady-state warm render (46,830 -> 43,426), with SQL statements per request down from 9 to 6. Rendered output is unchanged. Three changes in EventsController:

• Trim redundant eager associations in load_events: root-level :permissions on Workshop/Meeting/Event and :sponsorships on Event are never read on the render path. organisers is a through-permissions association (its join already covers permissions), sponsors goes through sponsorships, and the chapter-organisers fallback reads chapter.permissions, which stays eager-loaded. Dropping them shrinks the cartesian join fan-out per workshop row.
• Collapse latest_model_updated from four MAX(updated_at) queries (Workshop, Meeting, Event, Member) into one UNION ALL subquery with a single MAX. Same nil-when-empty semantics, one round trip instead of four; this also feeds the ETag.
• Skip the fetch pipeline after a performed 304: fresh_when renders a 304 without halting, so the fetch methods previously ran on top of the already-performed response. Mirrors the workshops#show guard.

Measurements

GC.stat(:total_allocated_objects) deltas through the full request stack, median over 25 requests per phase, anonymous GET over a seeded 20-card page backed by ~1,850 records. Run-to-run noise within a process is ±2 objects; the benchmark is sensitive to Postgres plan changes in the eager queries, so each comparison below was taken from back-to-back runs inside a single plan window.

phase	before	after
cold render (fragments cleared)	87,196	73,401 (−15.8%)
warm fragments (steady state)	46,830	43,426 (−7.3%)
SQL statements, cold	9	6

Production drain logs (2026-09-24..30) put the warm path at the site's heaviest controller endpoint per request, so the warm column is the steady-state win.

Verification

• full test suite: 1,593 examples, 0 failures
• RuboCop clean across app/
• rendered HTML verified identical: the page was rendered with the old and new pipeline in one process over the same fixtures (including a workshop with no organisers, exercising the chapter-organisers fallback), byte-for-byte equal

Review notes

• The dropped associations are the riskiest part: anything reading workshop.permissions or event.sponsorships directly would now lazy-load. The render path reads only organisers, sponsors, venue, chapter and the chapter-organisers fallback, all still eager-loaded.
• The UNION query's value feeds fresh_when; semantics (including nil when all tables are empty) match the previous four-query form.
• No schema, data or Gemfile changes; the eager lists on other pages are untouched.