Optimize /authors route and location search queries

next steps

—

key decisions

• Focus analysis on RedditUserIndex (Livewire), not the controller (which only does legacy redirect).
• Location typeahead searches cities Meilisearch index, not raw locations.
• Favorites system uses Maize Markable (polymorphic markable_favorites table) — extendable to Location.
• No user settings page exists; clean slate for default-location presets.

open questions

—

next steps

—

key decisions

• Probe directly on production (forge@23.88.0.74) via SSH for authoritative numbers.
• Use EXPLAIN FORMAT=TREE to capture query plans.
• Focus on Dallas (dense) and a sparse location (5-8 authors) as representative test cases.

open questions

—

next steps

—

key decisions

• Dedicated branch for the plan (no code changes yet).
• Facts doc as the single source of truth — all agents must read it before making claims.
• Use `docs/authors-optimization/` directory for the plan.

open questions

—

next steps

—

key decisions

• The bottleneck is dense-location filtering (338ms/page for Dallas) caused by pivot-driven candidate set forcing a filesort by karma.
• ~94% of 206K locations are junk; a min-author floor of 5-10 removes ~90%.
• Proximity is viable but expensive and must be city-centric, quality-gated, and cached.
• Favorites already exist (Maize Markable) and extend to Location; no user settings page yet.

open questions

—

next steps

• Run a clean adversarial product/UX review to close the gap left by the malfunctioning verifier.

key decisions

• The workflow includes foundations, 6 parallel feature/design agents, 3 adversarial verifiers, and a synthesis agent.
• Two feature agents (quality, favorite-locations/settings) hit structured-output retry caps but wrote their files successfully.
• The product/UX verifier returned a garbage result, leaving a gap.

open questions

—

next steps

• Wait for the product/UX adversarial review agent to complete and fold in its findings.

key decisions

• All three load-bearing corrections from adversarial verification are properly woven into section docs and the README risk register.
• Tool-call leakage in doc 50 was cleaned; no leakage remains.
• Every load-bearing file/symbol reference in the docs exists in the codebase.

open questions

—

next steps

—

key decisions

• Radius option set = {0, 10, 25, 50, 100} miles (no 150/250), resolving B1.
• Coarse-location degrade is a visible disabled state, not a toast, resolving B2.
• Quality control is one global integer threshold + a binary toggle (no per-user granular floor), resolving S1.
• Home-base precedence: users.location_id is the home base; manual chip model supersedes auto-seed, resolving S2.
• Followed × proximity composition: followed_only filters authors, not locations; proximity cache key must include followed_only and user scope, resolving S3.
• Composite location chip: remove manual chip model; typeahead is the single entry point, resolving S4.

open questions

—

next steps

—

key decisions

• followed_only must be in the cache key because it filters authors, not locations, and is not captured by location_ids.
• User scope is added only when followed_only is active, to avoid unnecessary key bloat for the common all-authors case.

open questions

—

next steps

—

key decisions

• Only documentation was committed; no application code changes were made.

open questions

—