chore(adapters): bump granitelib-core pin for fixed requirement-check io.yaml

[planetf1]

What this fixes

The requirement-check aLoRA on LocalHFBackend never activated: the published io.yaml instruction started <requirements>: {requirement}, and the Granite tokeniser merges >: into one token, so the declared invocation sequence never appeared in the prompt (#1679). The activation guard added in #1685 correctly left the adapter off in that case, but it meant callers who explicitly registered the aLoRA got base-model answers plus a one-time warning.

ibm-granite/granitelib-core-r1.0 PR #46 (merged 2026-10-05, commit c4b4fc6ec4fa05b6c8f02b25ac8822da085962b5) drops the colon (<requirements> {requirement}), fixing the tokenisation. This PR bumps _CORE_R1_SHA in mellea/backends/adapters/catalog.py to that revision and removes the strict xfail from test_requirement_check_adapter_moves_score. Two other references to the old pin (a docs tutorial's sample output, the uncertainty Ollama build script) are updated to match.

Only the requirement-check files changed between the old pin (d0a2a96a) and the new one — context-attribution and uncertainty, which share _CORE_R1_SHA, are untouched.

Testing

• GPU (BlueVela/LSF): test_requirement_check_adapter_moves_score, test_check_certainty_adapter_moves_score, and test/stdlib/components/intrinsic/test_core.py's qualitative suite — 5 passed, 2 pre-existing/unrelated xfails, 0 failures.
• CPU: test_run_transformers[requirement_check_alora] passes against the new pin — the canned fixture's expected direction still holds.
• uv run pytest test/ -m "not qualitative": 4758 passed, 0 failed, 20 skipped, 1 xfailed, 3 xpassed — all pre-existing/unrelated.
• ruff format --check, ruff check, mypy: clean.

Scope note

This fixes activation only. requirement-check's catalog entry has no adapter_types restriction, so once the aLoRA activates it becomes the default path for requirement_check(). Whether the aLoRA's verdict quality is good enough for that default is tracked separately in #1707 and intentionally out of scope here.

Fixes #1699
Fixes #1679

[planetf1]

GPU verification done on BlueVela (LSF, job 104708, 1×H100, `p5-r32-n3`):

```
test/formatters/granite/base/test_base_alora_activation.py::TestAloraDifferentialEndToEnd::test_requirement_check_adapter_moves_score PASSED
test/formatters/granite/base/test_base_alora_activation.py::TestUncertaintyAloraDifferentialEndToEnd::test_check_certainty_adapter_moves_score PASSED
test/stdlib/components/intrinsic/test_core.py::test_certainty PASSED
test/stdlib/components/intrinsic/test_core.py::test_requirement_check PASSED
test/stdlib/components/intrinsic/test_core.py::test_find_context_attributions XFAIL
test/stdlib/components/intrinsic/test_core.py::test_find_context_attributions_resolve XFAIL
test/stdlib/components/intrinsic/test_core.py::test_certainty_with_tools PASSED

5 passed, 2 xfailed, 0 failed in 65.12s
```

The two find_context_attributions xfails are pre-existing and unrelated to this change (different adapter). The previously strict-xfailed test_requirement_check_adapter_moves_score now genuinely passes — the adapter activates and the score moves as expected. This closes out the last checklist item in #1699.

[AngeloDanducci]

LGTM