fix(cursor): range-aware scan instead of blind 250k ROWID cap (#482)

[iamtoruk]

Fixes #482. Opening for review (not merging).

Problem

On large Cursor DBs, the scan kept only "the most-recent 250k bubbles by ROWID" and warned unconditionally. Because ROWID ≈ insertion order (not strictly the requested window), this could drop in-range older sessions from long-range reports, and it warned even when the requested window fit well under the cap. The cap exists because the bubble timestamp lives inside the JSON value (no index), so the date filter forces a full JSON decode per row — multi-GB DBs were stalling 30s+.

Fix

Replace the blind ROWID cap with a range-aware paged scan for DBs over the budget:

• Page ROWID-descending, keep only rows within the window (createdAt > timeFloor), and stop once a full page falls past the window floor (so short ranges decode far fewer rows).
• Keep a hard budget (default 250k) as a backstop for genuinely enormous in-range scans.
• The "older sessions may be missing" warning now fires only when the budget is actually hit, not on every large DB.

Small DBs (≤ budget) keep the original un-paged query unchanged.

Behavior

• Short ranges: fewer decodes, no spurious warning, no loss.
• Long ranges that fit the budget: full window returned (previously could drop older in-range sessions).
• Truly huge in-range scans: truncated to the budget keeping the newest in-range bubbles, warned only then.
• Also improves the daily-cache backfill (long-range / menu-bar history) the same way.

Tests

New tests/providers/cursor-large-db-cap.test.ts (budget shrunk via CODEBURN_CURSOR_MAX_BUBBLES):

1. In-range sessions with low ROWIDs are kept when the DB is over budget (old code returned 0 here).
2. Over-budget DB but in-range fits the budget → full window returned, no truncation.
3. Over budget → truncates to budget keeping the newest in-range bubbles (warns only then).

Full suite: 1209/1209. Existing cursor tests unchanged.