Bug: Fix start of week CI

[mtcarlone]

Fix Monday regression: schema collision bug + drop EOL dbt versions from CI

The first weekly Tier 3 regression against main failed broadly. Triage of all 43 matrix jobs found two independent root causes — one real bug, one environmental. This PR fixes both.

---

Root cause 1 — DBT_VERSION schema collision (real bug)

This is why supported versions (BigQuery 1.9, 1.10) failed.

scripts/ci/test.sh read the positional dbt_version arg into a local variable used only to pick the tox env name — but never assigned it to the DBT_VERSION env var that profiles.yml interpolates into the schema/dataset name. _lib.sh's ensure_dbt_version() then defaulted it to empty.

Result: every matrix job within a warehouse wrote to the same schema (dbt_artifacts_test_commit__<sha> — note the empty version segment). Under max-parallel: 8, parallel jobs on the shared cloud warehouses collided:

409 Already Exists: Table ...microbatch_seed; reason: duplicate
429 Exceeded rate limits: too many table update operations for this table
Not found: Table ...microbatch_seed was not found in location US

BigQuery failed hardest (strict table semantics + rate limits); Snowflake survived this run on luck but was equally exposed. Container-isolated warehouses (postgres/trino/sqlserver) were unaffected — they use fixed schemas. Invisible locally because we only ever ran one job at a time.

Fix: test.sh now does export DBT_VERSION="${dbt_version}"; removed the misleading ensure_dbt_version() helper. Each pinned version gets a distinct schema, so parallel jobs no longer collide.

Root cause 2 — EOL versions can't run on the runner

Every failure on 1.3–1.8 is environmental, not a package bug:

• 1.3–1.7 import distutils (removed in Python 3.12) → die at dbt clean.
• 1.3–1.8 predate the microbatch incremental strategy (a dbt 1.9 feature) that the test fixture uses.

Keeping them green would require a legacy Python 3.11 lane plus per-version gates — tech debt for versions dbt Labs has already EOL'd.

Decision: drop dbt 1.3–1.8 from the CI matrix. Matrix trimmed 42 → 15 slots (1.9 / 1.10 / 1.11 / latest).

---

⚠ This is a CI-coverage change, NOT a support drop

• require-dbt-version is unchanged (>=1.3.0) — older-dbt consumers can still install the package; it's simply no longer verified each release.
• tox.ini envs for 1.3–1.8 are kept — a maintainer can still run tox -e integration_snowflake_1_5_0 on a local Python 3.11 to debug a consumer report.
• The formal support-floor bump (require-dbt-version → >=1.9, removing the tox envs, public announcement) remains a separate 3.0.0 decision. This is Phase 1 (CI descope) brought forward, not a pre-emption of that.
• A consumer-facing note was added to README.md ("Supported dbt versions").

---

Files changed

File	Change
scripts/ci/test.sh	Export DBT_VERSION from the version arg (the fix)
scripts/ci/_lib.sh	Remove misleading ensure_dbt_version() helper
.github/workflows/release.yml	Matrix 42 → 15 slots; explanatory comment
README.md	"Supported dbt versions" note
specs/ci-rework/README.md	§12.3.0 records both root causes + decision