0.7.8: typo-tolerant make lookup, top_models(country:) by the country's own rank, dataset 2026.10.0

[rameerez]

0.7.8 — typo-tolerant make lookup, per-country top_models, dataset 2026.10.0

Do not gem push from this PR — publishing is owner/release-workflow only. This PR bumps the version and ships the code + bundled data.

Why

1. top_models(country:) sorted by GLOBAL popularity after filtering (bug filed by the SEO session: the Ukraine page and ?country= strips were ordered by everybody else's popularity). global_decile is a presence-weighted average across countries (pipeline#205); a UK list must be in UK order.
2. Coordinator request (STRETCH 2026-10-10): production logs show developers typing misspelled makes ("marcedes" twice from one mechanic) and getting nothing. Exact/alias first, then a bounded edit-distance fallback, never guessing between two equally close makes.
3. The bundled snapshot was 2026.07.4, three releases behind.

What

• Typo tolerance (lib/vehicles/fuzzy.rb, Dataset#find_make): after exact name/slug/alias/built-in alias, the closest make by optimal string alignment (Levenshtein + adjacent swap = 1 edit), stdlib only. Budget: none under 4 characters, 1 edit for 4-5, 2 from 6 (within the requested ≤2). Unique best only — two makes tied → nil ("Acra": Acura/Acma; "telsa": Tesla/Temsa). Guards, each added after a measured failure: a candidate may not have fewer words than the query ("Volvo V 60" → Volvo "60" was possible); every word keeps its first letter ("citroen e" → Citroën DS was possible through the web resolver's prefix loop); the configured other_label never fuzzy-matches ("Otra" → Ora). fuzzy: false opts out on make, models, model, model_options, find. Both validators stay exact — a typo never validates as a stored make. Vehicles.find runs its exact pass over every split before a fuzzy retry.
• Per-country ranking: Model#country_ranks / #rank_in(:gb); top_models(country:) orders by that country's own rank (ranks are per kind — pass kind:), models without a rank there follow in the old order, data without the key degrades to the 0.7.7 order. US/CA cars have no ranks (presence-only sources) → global order (documented).
• Data: rake "data:refresh[v2026.10.0]" (new, rakelib/data_refresh.rake, tags only) writes data/vehicles.json = the release's dist/vehicles.json plus only country_ranks, merged from the same tag's open catalog/<kind>/models.json popularity.by_country[cc].rank; stripping the key gives the release bytes exactly. When the projection ships the key natively (companion pipeline PR, emit.rb write_dist_json!), the task asserts equality instead. data/PROVENANCE-vehicles.md records tag/commit.
• README (typo paragraph + guards, per-country line, rank_in, 15 countries / ~15k models), CHANGELOG 0.7.8, MCP top_models note, version 0.7.8.

Evidence

• bundle exec rake test: 213 runs, 189,522 assertions, 0 failures/errors/skips (base: 192). Required tests: marcedes→mercedes-benz, volkswagon→volkswagen, tie→nil (Acra, fixture asserts the tie is real in the data).
• bundle exec rubocop: 0 new offenses (main already carries 111; identical per-file/per-cop tally).
• Parity with 0.7.7 on the same 2026.10.0 data: 0 changed / 0 lost answers on 28,795 "Make Model" + slug queries through the verbatim web resolver (app/models/resolver.rb), on 1,255 DEVQ A1 sweep queries, and (Opus verifier) on 37,050 find + 65,871 web-resolver-loop queries; 3,284 newly resolved, none >2 edits from its model.
• Labelled random-typo sample (verifier): make() right 1,351/1,869 (1-edit) and 714/1,179 (2-edit); wrong-make 9 and 15 — mostly first-letter typos (renley→Riley), the documented trade-off of the first-letter rule. find wrong-make: 0.
• make fuzzy path 0.5 ms/call; exact hits unchanged (0.004 ms).
• Data: country_ranks on 14,880 models, 0 differences from the catalog ranks across all 6 kinds; top-20 per country equals an independent catalog computation for 88/90 kind×country pairs (the 2 = us/ca cars, no ranks).

Reviews (I-11 + codex)

• codex (codex exec review --base origin/main), round 1: [P2] out-of-budget candidate accepted (Vehicles.make("z" * 28) → Saxas Nutzfahrzeuge Werdau) — fixed a802168 + regression test; [P2] fuzzy: not forwarded through models/model/model_options — fixed a802168. Round 2: no actionable regressions.
• Opus I-11 verifier (independent re-derivation): d1263f9 PASS-WITH-FIXES (P2 transposition telsa→Temsa; P2 Other label; P2 web resolver loop [web-side, see below]; 5×P3 docs/rake) — all gem items fixed f9130ff/f36a6a3; re-check at f9130ff PASS (OSA vs textbook over 104,000 comparisons: 0 mismatches; rake refs/tags-only verified incl. v2026.10.0;echo pwned rejected).
• DEVQ sweep found two more before the verifier: word-swallowing (Volvo V 60) 3bca6a8, first-letter (Citroën e-C3) d1263f9.

For the web session (not in this PR)

• GET /api/v1/makes/:make_slug/models?country=gb does not use top_models; makes_controller#models sorts by name, or [global_decile, name] with order=popular (measured ford/gb: Mondeo #46 first, Fiesta Add first-class "Other / not in the list" escape hatch (v0.4.0) #1 absent from the first six). After bumping to 0.7.8: sort by m.rank_in(country) || Float::INFINITY first when country is given.
• resolver.rb rung 3 calls Vehicles.make longest-prefix-first; with 0.7.8 run an exact pass (fuzzy: false) before a fuzzy pass and label typo hits (make_typo:), lower confidence — today's corpus is clean, but by data, not design.
• Routing (Catalog.find_make) should pass fuzzy: false or 301 to the canonical slug, so /makes/marcedes does not serve a 200 duplicate.

🤖 Generated with Claude Code

[rameerez]

Delta c16818d (coordinator additions, STRETCH 2026-10-10):

• Model#mass_decile (dataset popularity.mass_decile, 2026.10.1+, pipeline#255 / ruling R3) — nil-safe on the bundled 2026.10.0; top_models(by: :mass_decile) (records without it excluded → [] on 2026.10.0; default stays :global_decile; other values raise). README separates presence-average vs mass ordering.
• Fuzzy NEVER fires on id/path/slug-shaped input (any /, or all-lowercase hyphenated): marcedes-benz, marcedes-benz/c-class, /makes/marcedes, car/marcedes-benz/c-class → nil (same as 0.7.7). Human names still resolve. MCP search_makes/get_model exact-only. Tests: test/id_shaped_input_test.rb, test/mass_decile_test.rb.
• CI Lint (Ruby 3.4/rails_8.0) failed on main too (111 inherited rubocop offenses, red since 0.7.7): frozen in a generated .rubocop_todo.yml, no code change; new code fully linted (0 new offenses vs main). Ruby 3.3/rails_8.0 failure was a rubygems.org network error in bundle install.
  Local: 221 runs, 0F/0E/0S; bundle exec rubocop rc=0. Opus I-11 re-verification of this delta running.

[rameerez]

REL merge gate (lane REL, owner-authorised stretch 2026-10-10) — independent check of the delta c16818d..22c4807, since the delta verifier's verdict text was not posted here.

• Identity: 9 commits, author [email protected] on every one; no foreign e-mail in any message. CI 7/7 SUCCESS at 22c4807 (Ruby 3.4 / rails_8.0 was red on main since 0.7.7 and is green here via the frozen .rubocop_todo.yml).
• Delta read (lib only): Fuzzy.eligible_input? now also refuses Symbols, any / _ + \, a leading/trailing/double hyphen and single-case hyphenated tokens (marcedes-benz, MARCEDES-BENZ), keeping mixed-case and plain words human; Model#mass_decile reads popularity.mass_decile only when popularity is a Hash (nil-safe on 2026.10.0); a tie-break comment on top_models(country:, by:). Tests added: test/id_shaped_input_test.rb, test/mass_decile_test.rb. Nothing else in the delta touches behaviour. This matches the stated PASS-WITH-FIXES list (P2 machine shapes + P3s) and the web session's requirement that fuzzy never fires on id/slug/path input; MCP tools stay fuzzy: false (search_makes, get_model) or take no make string (top_models, search_models).
• Merged-tree run (trial worktree, branch already up to date with main): bundle exec rake test 223 runs / 189,562 assertions / 0 failures / 0 errors / 0 skips; bundle exec rubocop 52 files, no offenses.
• The gem repo's only workflow is tests.yml; merging publishes nothing to RubyGems (gem push stays owner-only).