fix: retry transient failures in jina (TEI) and ollama embedding providers

[GoodbyePlanet]

Closes #71.

Problem

server/embeddings/jina.py (self-hosted TEI) and server/embeddings/ollama.py had no retry at all — a single transient 5xx or connection reset aborted the file's embedding. server/indexer/pipeline.py:252-254 catches that, logs, and continues, so the file keeps its stale index entries until a later run.

Meanwhile the 429 backoff + batch loop + malformed-response check was duplicated near-verbatim across voyage.py, openai.py and jina_api.py. None of those retried 5xx or transport errors either — only 429.

Change

New server/embeddings/http_batch.py:

• post_with_retry() — retries 429 (honouring Retry-After), 5xx, and httpx transport/timeout errors on the shared [10, 20, 30, 40] schedule, then raise_for_status().
• embed_in_batches() — owns the chunk loop and the vector-count check, parameterized by a per-provider make_body / extract.

All five providers now route through it. Each keeps its own batch size (32 self-hosted, 128 hosted), request shape, and dimension resolution — those are genuinely per-provider. Net −60 lines across the providers.

Tests

• New tests/embeddings/test_http_batch.py — 12 cases: backoff sequence, Retry-After precedence, 5xx retry + exhaustion, transport-error retry + exhaustion, 4xx not retried, chunking, empty input, length mismatch.
• Retry and length-mismatch tests added to test_jina.py and test_ollama.py — new coverage, this is the actual fix.
• The 61 pre-existing provider tests pass unchanged, which is the regression gate on the refactor. Full suite: 362 passed.

🤖 Generated with Claude Code