fix(data): chunk backfill's candle requests too, the last unpaged window site

[eaitbrahim]

Follow-on to #269 / #271. Supersedes #294, which I opened against 29d9b20 before those landed and have now closed as redundant.

What this is

#269 chunked market_feed.poll_once. #271 swept for the same defect and fixed repair.py and history.py. market_feed.backfill was not reached by either, and has the identical shape — _missing_ranges groups absent timestamps into contiguous ranges, and each range was requested in one call, so a hole wider than the venue's ~350-candle cap 400s exactly as the poll path did. On an empty repo the entire history window is one such range.

It is latent, and I'd rather say so than dress it up

backfill has no production caller today — keel fetch goes through history.ensure_history. Nothing is failing on this right now.

Worth closing anyway on #271's own stated grounds, that every candle-request windowing site in the codebase should agree. A public data-layer entry point that 400s the moment anything calls it is a trap left armed for whoever calls it next.

Why it was easy to miss

The (MAX_CANDLES_PER_REQUEST - 1) arithmetic existed only inside _poll_catch_up, so backfill had nothing to be inconsistent with — it just looked like ordinary pre-#269 code. #271 had to correct that exact off-by-one after finding it duplicated in history.py, which is the same bug class the shared MAX_CANDLES_PER_REQUEST import exists to prevent.

So the windowing now has one definition, _request_windows, used by both backfill and _poll_catch_up. _poll_catch_up keeps its own filter and seen dedup and is otherwise unchanged; folding it onto the helper also removed a window_start = window_end + gran_sec line left stranded when its while became a for.

Tests

Written first and confirmed failing for the right reason — both cap tests failed on a single 575-candle request against the unfixed code:

• test_backfill_never_requests_more_than_the_candle_cap
• test_backfill_chunk_windows_are_contiguous_and_cover_the_gap
• test_backfill_still_uses_one_request_for_a_gap_within_the_cap — regression guard, passes before and after

Gates

• pytest — 2762 → 2765 passed, 1 skipped
• ruff check — clean
• mypy keel — clean, 77 source files

🤖 Generated with Claude Code