New SQLite table `exercise_aliases (alias, canonical, source)` seeded
with ~40 common gym shorthand entries (OHP, RDL, "Bench", "Squat",
plural/singular drifts, slang). Lookups go through this table first,
then fall through to the strict exercise_db matcher — so the strict
matcher's "false negative for ambiguous single tokens" property is
preserved while still resolving every-day vocabulary.
Schema decision: every seed row is tagged `source='seed'` and re-seeded
on every init_db (deleted-then-reinserted), so editing the seed dict
in code is the one source of truth. User-inserted rows are tagged
`source='user'` and never touched by re-seeding. Migration path covers
existing DBs where the `source` column didn't exist (those rows tagged
'seed' on first migration, then refreshed from the current seed).
New helper db.lookup_exercise(name) wraps the alias resolution + the
exercise_db.lookup() call.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the static exercise reference data (~870 entries, public
domain, source: github.com/yuhonas/free-exercise-db) plus a
conservative name matcher. New endpoint:
GET /api/exercises/lookup?name=<name>
→ {"match": {"name", "primary_muscles", "secondary_muscles",
"equipment", "category", "level", ...}}
→ {"match": null} when nothing plausibly matches.
Matcher tiers (priority order):
1. exact (case-insensitive)
2. compressed exact ("Pull-ups" → "Pullups")
3. compressed substring, with a guard: single-token generics
like "Bench"/"Squat" return null instead of misleading the
user — the planned alias table will handle these properly.
4. token-overlap with ≥50% coverage of the user's tokens.
UI integration ("Trains: chest · shoulders") comes in step 2.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
