Thanks for the Windows memory-retention fix for #832. Triage: high-priority stability/performance PR.

Review will check the mimalloc options and worker-count change separately: we need post-index RSS to return to sane levels, but we also need to avoid an over-broad throughput regression or conflicting with the explicit memory-budget work in #752/#685. Please keep the before/after memory evidence current in the PR.

Yes, still active and waiting for review. All CI checks pass. Happy to adjust if there are concerns about the worker-count change — the mimalloc arena purge (mem.c) is the core fix, and the worker cap (system_info.c) can be dropped if it conflicts with #752/#685.

Quick status note: this PR is one of four open memory/RAM-policy changes (#833, #752, #586, #685) that we've reviewed individually and found genuinely complementary — so rather than merging them piecemeal, we're doing a combined design pass over the whole memory policy (explicit override, host-tiered defaults, retention bounds, post-index release, and the Windows auto-sync driver in #841) and will respond here with a concrete direction shortly. Your work is very much part of that plan — thanks for your patience!

Thank you — your diagnosis found the right lever. The core of the #832 fix is mi_option_page_reclaim_on_free=1: mimalloc v3 flipped the reclaim default so a long-lived thread no longer reclaims pages abandoned by exited worker threads → the RSS ratchet you reported. That exact option is folded into the keystone memory fix, merged as 7884ccb (PR #853), with you credited. The other hunks were dropped after analysis: arena_purge_mult=1 is inert at our purge_delay=0, the DEFAULT_RAM_FRACTION change was dead code (all call sites pass an explicit fraction), and the worker-cap belonged to a different layer (host-tiered budgets, since landed as #752).

Being straight with you: #832 isn't fully closed yet. The keystone also routes the two background index paths through a subprocess (so the kernel returns 100% of each cycle's RSS on exit), and #854 added budget-derived retention — so 'RSS won't come back' is fixed. But the trigger you hit on Windows (the watcher re-indexing on every poll even when nothing changed) is tracked separately as #841 and still to come. So I'll leave #832 open until that lands, but your page_reclaim finding is shipped. Really appreciate the sharp report — closing this PR in favor of the folded fix.