fix: default the weight unit to the trainee's wger profile

[wromansky]

The problem

log_set and add_exercise_with_sets both default weight_unit to a hardcoded "kg":

routines.py       weight_unit: str = "kg",
workout_logs.py   weight_unit: str = "kg",

The wger profile already records which unit the trainee works in, and that default ignores it. A trainee whose profile is set to pounds reports "225", the caller omits the unit, and 225 is stored as kilograms — wrong by a factor of 2.2.

The failure is silent and unrecoverable after the fact. 225 is a plausible number in either unit, and the stored row does not say which one the trainee meant, so no later read can tell a mis-unit log from a correct one.

I hit this with an LLM client driving the server: the model omitted weight_unit on a routine build, and every planned entry landed as kg against a pounds profile. A prompt rule telling the client to always pass lb is a workaround for a default that is simply wrong for that user.

The change

Omitting weight_unit now takes the unit from /userprofile/.

• New profile_weight_unit coroutine in tools/common.py. It reads the profile on each call that needs it, and is deliberately not cached and not a per-registration closure — see below.
• Both write tools resolve the unit only when the caller omitted it.
• The server already reads this endpoint for whoami and the nutrition targets, so no new dependency or scope.

Why there is no cache

The first draft of this PR cached the unit in a closure, copying the shape of language_id_resolver. That was wrong, and the fix is in this branch.

register_all runs once per process from build_server, mcp_auth defaults to oidc, and _ProviderAuth resolves the Authorization header per request off a ContextVar — one shared client serves every user. A cache in that closure is therefore process-wide, not per session and not per request, so the first trainee to omit weight_unit would pin their unit onto every other trainee's writes until restart. That is the same silent wrong-unit write this PR exists to prevent, only spread across users.

language_id_resolver caches safely because wger's language table is global and static. A per-user profile field is not, and the earlier claim that "the setting does not change mid-session, so one call covers it" was answering the wrong question — the scope was never a session.

One extra GET, against the roughly six requests add_exercise_with_sets already makes, is noise. A per-user keyed cache would be more code for no measured benefit.

Behaviour

An explicit weight_unit still wins, so callers that pass one are unaffected. Callers that omit one change behaviour by design: a pounds profile now records pounds where it previously recorded kilograms. That is the point of the change.

A profile that cannot be read falls back to "kg" — today's behaviour — so a transient error never fails the write. Userprofile.from_dict routes weight_unit through check_weight_unit_enum, which raises TypeError (not ValueError) for anything outside {kg, lb}, a null value included; TypeError is caught alongside UnexpectedStatus and httpx.HTTPError, so an unparseable profile cannot crash log_set — a tool that never read the profile before this PR.

Tests

tests/test_weight_units_and_rir.py:

• omitted unit follows a kg profile (log_set)
• omitted unit follows an lb profile — 225 stays 225 lb (log_set)
• explicit kg beats an lb profile
• a profile that cannot be read falls back to kg without failing the write, parametrised over an unreachable profile and an unparseable one
• omitted unit follows an lb profile on add_exercise_with_sets too — the other tool this PR changes

The existing test_kilograms_remain_the_default asserted the hardcoded default; it becomes the profile-says-kg case, which is the same assertion for the right reason.

_mock_creation now stands in for /userprofile/ as well. Without that, the two pre-existing routines tests that omit weight_unit (test_weight_may_be_omitted, test_rir_is_optional) fell through the new lookup with userprofile_retrieve unmocked and sent a real request to https://wger.test/api/v2/userprofile/ — CONTRIBUTING requires respx for outbound HTTP, and tests/conftest.py has no autouse guard. They passed only because the ConnectError landed on the kg fallback, so test_rir_is_optional's weight_unit == 1 was asserting the error path while reading as the profile path. Verified with a temporary autouse guard over httpx.AsyncClient.send: 2 of the file's tests attempted outbound HTTP before the fix, none after. Both still assert what they were written to assert.

Full suite: 252 passed. ruff check . clean; ruff format --check clean on the touched files.

Collisions with the other open PRs

Checked with git merge-tree against the current head of each branch.

• feat(routines): take unit names on slot entries, not just wger's ids #18 (feat/repetition-unit-by-name) — one conflict, the ## Unreleased block in CHANGELOG.md, for whichever lands second. common.py, routines.py and workout_logs.py all merge cleanly. In particular this PR does not touch as_weight_unit; only feat(routines): take unit names on slot entries, not just wger's ids #18 changes that function's signature, so there is no collision there.
• feat(routines): let add_exercise_with_sets record a rep range #17 (feat/max-reps-on-add-exercise) — two conflicts: the CHANGELOG.md ## Unreleased block, and src/wger_mcp/tools/routines.py. Both PRs edit the same statement in add_exercise_with_sets: this branch replaces unit = as_weight_unit(weight_unit) with the profile lookup, and feat(routines): let add_exercise_with_sets record a rep range #17 inserts its max_reps < reps guard immediately after it. Resolution is mechanical — keep both — but it is a real code conflict, not just a changelog one. README.md and tests/test_weight_units_and_rir.py merge cleanly.
  Worth noting for whichever lands second: three of feat(routines): let add_exercise_with_sets record a rep range #17's new tests call _mock_creation and omit weight_unit, so they depend on this PR's _mock_creation mocking /userprofile/. Drop that during conflict resolution and those tests start reaching for the network.
• docs(log_set): say that exercise_id is the movement performed #19 (docs/log-set-exercise-id) — merges cleanly, no conflict.