fix(market-adapter): keep Kibana LP fetches alive under proxy connect…

[Impulssi]

…ion resets

The Kibana console proxy at kibana.bitshares.dev now resets connections mid-transfer once a single search page streams enough data (observed with full _source payloads around ~8k documents, and even ~2k documents with _source: true). The LP fetcher queried with size 10000 and _source: true, so every fetch died at the availability probe — and worse, silently:

• doKibanaRequest only listened for request-phase errors; a response stream aborted mid-body never emits 'end', so the surrounding promise stayed pending forever and fetch_lp_data hung with no error output.
• There was no retry, so a single transient reset killed the whole fetch.

Fix

• kibana_client: handle 'aborted'/'error' on the response stream and reject the promise, so callers fail loudly instead of hanging forever.
• kibana_candles: request only the _source branches the field map actually reads (plus the fixed timestamp/sequence fields) instead of full documents. Restricted projections keep each page small and fast (verified live: 2000-doc pages with restricted _source return in ~1.5s where full _source aborts).
• kibana_candles: default kibanaPageSize 10000 -> 2000.
• kibana_candles: retry transient page failures (aborted / connection reset / timeout) with linear backoff — search_after pagination is stateless on the server, so replaying a failed page is safe.
• buildDirectionalDocumentQuery accepts an optional sourceFields argument; omitting it keeps the legacy _source: true behavior for other callers.

Testing Notes

• tests/test_kibana_candles.ts: new coverage for the source projection derivation, the _source wiring in the directional query (plus legacy fallback), and transient-page retry recovery.
• npm run typecheck clean.
• Verified live against kibana.bitshares.dev with fetch_lp_data.js (BTS/XBTSX.USDT pool 1.19.48, 2-month 1h fetch): the availability probe and both chunk fetches complete where they previously hung indefinitely.

---

Note: I'm new to this codebase and worked through this change with an AI
coding assistant (opencode). The bug was hit on my own live instance and
all measurements / reproduction steps are mine — but please flag anything
in the code that looks off and I'll dig in and follow up.