[plainprogrammer]

Implements spec 037 (0.2.2) per its plan. The collection's filter bar gets a Location select: All locations, No location, or one of the collector's locations. The grid, the table, paging, sorting and bulk mode all honour it, combined with the name search (AND).

What each phase did

• Phase 1, models (ff995c1)
  • CollectionFilter::Place turns location into its canonical form: "", "none", or an id of at most 18 digits. Anything else counts as all. It is then applied as a plain lots.location_id predicate.
  • CollectionFilter reads the account's locations once (places?, the options, the chosen name) and checks the URL's value against them for display. Under a location filter the grid skips its place line.
  • bulk_selections.location_filter is a new string column, default "", null: false, reversible. BulkSelection#context?, #restart!, #lots and #record! use it, never re-resolved, so a removed or another account's location selects nothing (decision 7).
• Phase 2, controllers and helpers (1c57170)
  • The page, collection_listing_path (which leaves out all), the sort headers, Edit many and the bulk form all carry location.
  • Bulk submissions use the rendered rendered_location for their context, pagers, sorts, actions and Done. Only Filter applies the select's value, and it keeps the page only when neither filter changed.
  • BulkSelected's return path carries the stored location; Undo needed no change.
• Phase 3, views (2f4c2bd)
  • The FilterSelect sits in the filter bar with a Filter button, shown only while the account has a location.
  • In bulk mode it joins the bulk form, and the existing default go=filter button becomes visible beside it; it is still the first submit, so Enter still applies the filters.
  • The view-switch and Edit-many forms and both pagers carry location.
  • The empty sentences name the place: "No cards in your collection are in “Blue binder”.", "…are without a location.", and the forms with a query.
• Phase 4, measurement and documents (0690b63)
  • A :timing request spec: 5,000 lots over 300 locations, plus two other accounts' 5,000 lots each. Every page is under 500 ms (largest: the unfiltered table at 163 ms; with a filter, 21–106 ms), so the reserve [account_id, location_id] index is not added.
  • Updated: verification.md, the collection-and-lots area doc, the roadmap, the specs index and CLAUDE.md.
• Review follow-ups (afe0ce8): tests for findings 1–3 (below).
• Wrap-up
  • Screenshots and review record (96c3fbe).
  • Session learnings (46f7ae5): screenshot memory, one lesson.
  • Distillation (1acece1, 5b47f03, 8628b13): the index row is Frozen by collection and lots, plan retirement owed.

Rulings and deviations (each also in its commit body)

• No per-phase subagents. I implemented each phase myself, test first, because the Orca brief forbids helper agents other than the review.
• Empty sentence as a partial. collections/_no_match replaces the plan's no_match_message helper. A helper's String is escaped whole, which turned spec 004's literal quotes into " and broke three existing body assertions. The partial keeps them raw without html_safe.
• Migration rollback. The down uses SQLite's native ALTER TABLE … DROP COLUMN, so the table isn't copied and its marks aren't cascaded.
• No CSS. FilterSelect's existing rules suffice; the 390px system spec checks both modes.
• AC-5.8 refusal shows All locations. The removed location no longer resolves (decision 7).
• System spec written after the views. The request markup specs were the red step for Phase 3.
• Measurement as a timing spec. It runs with SCANNER_TIMING=1 and never gates.
• Commit trailers. ff995c1 and 1c57170 say Co-Authored-By: Claude Sonnet 5.5 by mistake. The work ran on Opus 5.5; later commits say so, and the pushed commits were not rewritten.

Implementation review: SPEC-ALIGNED

An independent read-only sdd-review Mode B on Opus found nothing blocking or major. It reported five minor findings:

1. A selection's ticks shown under a location had no direct test. Fixed (selection_state_spec).
2. The confirm page's Cancel/Back links, and the confirm page for a removed location, weren't asserted. Fixed.
3. The filtered table's pager count was covered only indirectly. Fixed.
4. verification.md sections were missing. Fixed.
5. A local Gemfile.lock platform change. Kept out of every commit.

The review also traced tenancy, the fail-closed paths and every caller; details are in verification.md.

Tests

• New: place_spec (4), collection_location_filter_spec requests (21), collection_location_select_spec markup (14), system spec (6).
• Extended: grid (+8), table (+3), bulk selection (+5), bulk removal (+1), selection state (+1).
• Not in the suite: one timing example, excluded unless SCANNER_TIMING=1.
• Suite: 2,483 examples, 0 failures, 1 pending. Local bin/ci is green, also in the pre-push hook on 8628b13.
• Flake: one earlier pre-push run failed on the known cross-worktree tmpdir race in Upgrade::Harness (Flaky: Upgrade::Harness "removes its temporary directory" races other suites through the shared tmpdir #178), unrelated to this change. It passed on re-run.

Screenshots

Taken with headless Firefox against a dev database seeded with placeholder printings (no card art): one collector, three locations, 14 lots.

Desktop, light	Desktop, dark

Phone 390px, light	Phone 390px, dark

Worth checking by hand

• A real phone: the native picker for the select, and touch targets (coarse pointer gives 40px). Headless Firefox can't test either.
• Bulk filter bar layout: a visible Filter button now sits beside the select in bulk mode for accounts with a location. Accounts without one keep the old screen-reader-only button.
• Measurements: taken in the test environment on the dev machine, as spec 033's were.

Closes #47

🤖 Generated with Claude Code

[plainprogrammer]

Follow-ups filed by the coordinator: #181 (shard 1's Net::ReadTimeout at sign-in, re-run green), #178 (harness tmpdir race). For the maintainer: check the location select on a real phone (native picker, touch target).