fix: pass embedding dimension to provider and surface sqlite-vec probe failures

[happy-v587]

Which issue or RFC does this PR close?

Closes #1385.

Rationale for this change

The configured EMBEDDING_DIMENSION was never sent to the embedding provider, so any configured dimension different from the provider default failed validation. The failure was invisible to users: the readiness probe swallowed the underlying error and only reported degraded / inference.embedding: unavailable, and if the failure happened during vector index initialization it was uniformly wrapped as CapabilityNotSupportedError: vector (sqlite-vec probe failed), which misleadingly suggests a missing sqlite-vec extension. Additionally, an interrupted startup could leave a vector_id=-1 probe row behind, which poisoned every subsequent startup with the same masked error until the database file was deleted manually.

What changes are included in this PR?

• composition.py: pass the configured embedding profile dimension to the Embedder via EmbeddingSettings(dimensions=...), so providers actually receive the requested dimension instead of returning their default.
• sqlite/memory_index.py: clear a leftover vector_id=-1 probe row before the vector index probe, making startup idempotent after an interrupted run.
• sqlite/memory_index.py: include the underlying error in the CapabilityNotSupportedError detail, with a dedicated hint when the existing vec table dimension does not match the configured profile (message contains both dimension values).
• composition.py: log embedding readiness probe failures via logger.exception before marking inference.embedding unavailable.

Are there any user-facing changes?

• Behavior: an explicitly configured embedding dimension is now honored by providers that support the dimensions parameter; providers that reject the parameter will now surface a visible error instead of silently mismatching dimensions.
• Error messages: vector index probe failures now carry the underlying cause text instead of a bare sqlite-vec probe failed; a stale vec table with a different dimension produces an actionable message naming both dimensions.
• No API, schema, or migration changes.

How was this change tested?

• 5 new unit tests (tests/builtin/persistence/test_sqlite_memory_vector_index.py, tests/builtin/runtime/test_composition_embedding.py): leftover-row self-heal, dimension-mismatch message contains both dimensions plus chained cause, generic cause surfacing, dimensions settings reaching the Embedder, and no-embedding-config passthrough.
• make test: no new failures vs the base commit (16 vs 17 pre-existing environment-dependent failures); make check clean.
• Manual end-to-end validation against a real OpenAI-compatible endpoint: fresh DB with dimension 1536 now reaches ready with vector/hybrid search enabled; a stale 1024-dim table now fails with a message containing both dimensions; an injected vector_id=-1 row is cleaned up on the next start.

AI usage statement

Built with opencode (GLM, Z.ai) under human direction and review; all code, tests, and validation steps were reviewed and confirmed by the author.

[happy-v587]

Both review points are addressed in 6229cb9:

1. Rejected dimensions now surfaces a stable redacted reason. _map_error attaches the structured HTTP status to provider-rejected failures, and dependency_readiness_probe() now exposes the stable code/detail carried by InferenceConfigurationError instead of collapsing everything to misconfigured: the Server and doctor now report inference.embedding: misconfigured: provider-rejected (HTTP 400). The raw provider response body is never included (the reason is built only from our stable code plus the status code), and plain configuration errors without a code stay redacted as bare misconfigured. Covered by new unit tests plus an end-to-end server test with a mock OpenAI-compatible endpoint returning 404, asserting the reason is stable across cached refreshes and the response body never leaks.

2. The dimension-mismatch hint is only emitted for a confirmed pre-existing table. _probe_failure_detail now reads the dimension of a pre-existing pc_memory_entry_vec table from sqlite_master and emits the migration hint only when that table exists with a different dimension; otherwise the actual probe error is preserved. A fresh database with dimension 65536 now reports the real sqlite-vec error (its 8192 maximum) instead of telling the user to migrate a nonexistent table. Covered by a new oversized-fresh-dimension regression test.

Validation: make test (950 passed, 9 skipped), make check clean.

[Teingi]

LGTM