fix(sl-toolkit): edits use PATCH, and sql_dialect is snake_case

[cjayyy]

Two bugs in the semantic-layer skill, both verified against a live metastore
(canary-orion, project 431). Either one alone is enough to make the skill fail or do damage.

Found while porting this skill into Kai's in-platform tooling
(AI-3661) —
rather than working around them downstream, fixing them here.

1. sqlDialect is rejected — /sl-build cannot create a model today

The API requires snake_case sql_dialect. camelCase gives:

422  missing property 'sql_dialect'

The skill's POST envelope, sl-build's model-create call, and the test fixture all used
sqlDialect, so the greenfield wizard fails at its very first write.

Both CI guards were asserting the wrong direction. test_sqldialect_is_camelcase and
test_no_dialect_drift required camelCase and forbade the correct key, and
schemas/semantic-model.json did the same in its not.anyOf. That's why the bug shipped and
stayed green. Inverted them, and the consistency check now matches the quoted payload key
instead of the bare word, so the gotchas section can still name sqlDialect in prose to say
it's rejected.

Also corrected the schema enum — only Snowflake and BigQuery are accepted. Redshift and
Postgres were listed; both return 422 value must be one of 'Snowflake', 'BigQuery'.

2. PATCH exists — the documented DELETE + POST edit path is destructive

The skill stated "The metastore has no PATCH — editing is DELETE old + POST updated" and built
rollback machinery around it. PATCH /api/v1/repository/{type}/{id} works:

	before	after
uuid	019fc788-d6cb-7582-95d5-46d96f1bad9f	same
revision	1	2

Verified on both a semantic-metric and a semantic-glossary object, so it isn't type-specific.

Delete-then-post instead mints a new UUID, resets revision history, breaks anything
referencing the old UUID, and leaves the layer missing an object if the POST fails.

Rewrote the edit section around a new api_patch helper — send only the fields that change. The
rename cascade to constraint metrics[] is preserved and is now also in-place. The rollback block
is deleted: nothing is removed, so there is nothing to roll back.

Testing

All 30 tests pass. I also re-introduced the sqlDialect bug in the fixture and confirmed the
corrected guards now catch it (2 failures) rather than demanding it.

30 passed in 0.06s

Release Notes

Fixes the semantic-layer skill's model-create payload (sql_dialect, not sqlDialect), which
was rejected by the metastore with a 422, and replaces the destructive DELETE + POST edit
recipe with in-place PATCH so object UUIDs and revision history survive an edit.

[cjayyy]

CI note: Tier 0 (static lint + case validation) passes — that's the job covering these changes, and it's the one that was previously asserting the wrong casing.

Tier 1 (skill activation routing) fails for an unrelated infra reason — the repo's Anthropic key is out of credit:

anthropic.BadRequestError: Error code: 400 — 'Your credit balance is too low to
access the Anthropic API. Please go to Plans & Billing to upgrade or purchase credits.'

It errors before evaluating any case, so it's not a signal about this diff. Locally: 30 passed.

[linear-code]

AI-3661

[cjayyy]

Re-checked the Tier 1 failure after the CI notification — it is not caused by this PR, and nothing in the diff can fix it. Evidence:

1. It fails on the API call, before evaluating any case. The job runs 3s and dies at:

   anthropic.BadRequestError: Error code: 400 — 'Your credit balance is too low to
   access the Anthropic API. Please go to Plans & Billing to upgrade or purchase credits.'
   

   The secret is present (the secret not available line in the log is the script being echoed, not the branch taken), so this isn't a fork-skip — the repo's Anthropic account is out of credit.

2. Tier 1 routes on skill frontmatter, which this PR does not touch. git diff origin/main..HEAD -- …/SKILL.md contains no name: or description: change.

3. The deterministic half of Tier 1 passes locally — activation/test_activation_cases.py + test_parse_answer.py: 80 passed. Those are the offline case-set and answer-parsing checks that don't need an API key.

4. Tier 0 — the job that actually covers these changes — passes, and it's the one that was previously asserting the wrong sql_dialect casing.

So: needs credit topped up on the repo's Anthropic account to go green. Happy to re-run once that's sorted.