fix(provider): salvage partial batches when an endpoint rejects one item

[dearbld]

Salvage partial batches when an endpoint rejects one item

Base: on top of pr-2 · stacked series: pr-1 → pr-2 → this PR → pr-5 → pr-3

Motivation

OpenAICompatProvider.embed sends chunks in batches of 32. When the endpoint answers HTTP 400 with a parameter-error body, the whole embed call throws — but such a 400 is frequently caused by one rejected item (an oversized or otherwise invalid chunk), not the batch shape. The blast radius is worst at boot: the first failed build plus allowFallback permanently pins the index to the local lexical provider, even though 31/32 items were perfectly embeddable.

What changes

On HTTP 400 with a multi-item batch:

1. Retry the batch item by item (input: [one]).
2. Good items keep their embeddings; the dimension is learned from the first good single response when dimension: 0 (inferred).
3. Rejected items become zero-vector Float32Array placeholders so row alignment is preserved — search still ranks them through the lexical channel. Float32Array specifically: the persistence layer calls vec.subarray, which a plain Array does not implement (a plain-Array placeholder crashed persistence and lost the whole build in production).
4. If every item fails, the original batch error is kept (no behavior change for genuinely bad requests).

Per-item retries share the batch's timeoutMs budget via AbortSignal.any — without a per-item cap, a slow endpoint could hang each of the 32 retries indefinitely.

Before / after

// before: one bad chunk kills the batch and (with allowFallback) the index quality forever
if (!response.ok) throw new EmbeddingError(`... HTTP ${response.status} ...`)

// after: 400 on a multi-item batch → per-item salvage, zero-vector placeholders, original error only if all fail

Testing

• tsc --noEmit clean; full suite 82/82 pass, including two new mock-server tests:
  • batch of 3 with 1 rejected item → 3 rows returned, placeholder is [0,0,0], good rows non-zero, request sequence 3 → 1 → 1 → 1;
  • every item rejected → original HTTP 400 error preserved.
• Suggested: replay against a real endpoint that returns code 1210 on token-limit overruns.

[JohnXu22786]

To clarify the intended behavior: after a batch embedding request fails, retry each text individually. Keep vectors for the texts that succeed and use a zero vector for a bad item, so one bad text does not fail the whole batch. Please keep that feature, but use one shared batch deadline: each individual retry should receive only the remaining time, rather than a fresh full timeout. Please add a timeout-boundary test and make sure cancellation is handled correctly.