fix(consolidation): expose failed LLM batches

[koriyoshi2041]

Closes #4151.

What changed

Add hindsight.consolidation.batch_failures, a counter for LLM batches that exhaust consolidation's retry policy. It uses a bounded reason label:

• response_validation for malformed JSON, schema validation, and output-length failures
• provider for other exhausted provider failures

The existing failed_consolidation gauge remains unchanged: it counts source memories permanently marked failed after adaptive bisection reaches a singleton. The new counter makes recoverable batch-level failures visible without changing that state-based contract.

Each failed batch increments once after its retry ladder, rather than once per attempt. Successful batches do not increment it.

Verification

• tests/test_metrics.py: 34 passed
• TestBuildResponseModel: 10 passed
• focused Ruff check and format check passed
• git diff --check passed

The repository hook reached the control-plane lint step but could not load @eslint/js because Node dependencies are not installed in this worktree. The changed Python files and tests passed their focused gates.

[koriyoshi2041]

Two CI failures look unrelated to this metrics-only diff:

• verify-generated-files stopped in setup-uv while fetching the latest uv version (fetch failed), before generation ran.
• free-threaded 3.14 completed 6,987 tests and failed only the existing OpenCode session-header fixture after LLMInterface.call() became a named result (result.ok / payload-vs-.content); the changed consolidation tests passed. PR fix(reflect): enforce max_tokens after forced synthesis #4159 hit the same free-threaded job failure on its independent branch.

I cannot rerun upstream jobs from the fork. Could a maintainer rerun the failed jobs?