perf(exercises): trim search results to what picks an exercise

[wromansky]

Measurement

A ten-result search_exercises call returns 3,913 characters, measured against a live wger instance:

field	chars	share
translations	2,286	58%
uuid	480	12%
name	365	9%
equipment	328	8%
category	208	5%
images	140	3%
id	106	2%

The translation table is every language's name for every hit — sixteen entries for a common lift — and the caller already filtered by language when searching.

Why it matters

That lands in the model's context on every lookup, and building a routine means one search per exercise. On a self-hosted model with a 32k window it accumulates into a hard failure: the agent that prompted this change exhausted its context part-way through a six-day build, after which every inference returned

{"error":{"code":400,"message":"request (N tokens) exceeds the available context
 size (32768 tokens)","type":"exceed_context_size_error","n_ctx":32768}}

until it burned its iteration cap and gave up with a half-built routine. Nothing in that failure names the real cause — it surfaces only as "reached maximum tool iterations".

Change

search_exercises returns id, name, category, equipment. search_exercises_by_filter keeps muscles too. Both drop uuid, images and the translation list. The matched name is still resolved exactly as before, honouring the caller's language.

Nothing is lost — get_exercise still returns the complete record. Search picks an id; the detail call expands it.

_shape_images had no remaining callers and is removed with it.

Measured after the change, same queries against the same instance:

'barbell bench press'   4,843 -> 1,206 chars
'lateral raise'         3,914 -> 1,107 chars

Tests

4 new tests in tests/test_exercise_search_shape.py: the lean key set, that no translation text or uuid survives into the payload, that filter search keeps muscles, and that get_exercise still returns translations, images and uuid in full.

Full suite passes (75) and ruff check is clean.

Note

If you would rather keep the fields and make them opt-in, a verbose: bool = False argument would work equally well — happy to switch. I went with the smaller default because the common caller is an LLM picking an id, and the detail tool already exists for everything else.

Independent of #2, #3 and #4 — off master, different file.