perf(prefilter): skip 0-cardinality auto-promoted prefilters

[JustMaier]

Summary

Auto-prefilter promotion was registering every hot filter shape whose compute_filters returned Ok(bitmap) — including bitmaps with 0 slots. A 0-cardinality prefilter can never substitute a real query (substitute path produces empty results, same as running the query directly), so registering it just:

1. Burns a slot in the bounded prefilter registry (default cap = 32). With cap reached, useful candidates get rejected with prefilter registry full.
2. Adds a stale entry to the merge thread's prefilter-refresh loop. Each refresh holds the FilterField read lock for 80–180 ms while compute_filters re-runs.

Impact (when paired with #224)

Pre-#224 the auto-promotion code was dead, so this never fired. After #224 unlocks it, registry fills with garbage within minutes of startup. Measured locally: flush apply mean regressed from 5.3 ms (post-#222 alone) back to ~146 ms (full stack), with top_filter=[userId=…] indicating the apply write lock was waiting on prefilter-refresh read locks.

Filter 0-card shapes at promotion time → registry stops filling with garbage → fewer stale prefilters per cycle → less read-lock-hold per merge cycle.

Change

Single guard:

let card = bitmap.len();
if card == 0 {
    tracing::debug!("auto-prefilter: skipped '{}' — 0 slots match (freq={})", name, hfs.total_frequency);
    continue;
}

before the merge_prefilter_registry.insert(…) call. 16 lines total including the explanatory comment.

Stacking

• Independent of fix(ops_wal): skip unreadable tail on closed gen to break reader stall #221, perf(filter): track dirty values in side set so merge_dirty avoids O(N) walk #222, perf(snapshot): parallelize per-bucket writes in save_snapshot #223, perf(snapshot): chunk for_each_versioned to bound lock-hold under high cardinality #225 — all touch different files.
• Pairs with fix(merge_thread): close while-loop brace correctly so post-cursor work runs every cycle #224: don't merge fix(merge_thread): close while-loop brace correctly so post-cursor work runs every cycle #224 alone (it exposes the regression). fix(merge_thread): close while-loop brace correctly so post-cursor work runs every cycle #224 + perf(prefilter): skip 0-cardinality auto-promoted prefilters #226 + perf(filter): chunk bitmap_bytes iteration to bound read-lock-hold #227 should land together.
• perf(filter): chunk bitmap_bytes iteration to bound read-lock-hold #227 (snapshot-based compute_filters) is the architectural fix for the same surface.

Test plan

• Build clean (cargo build --profile fast --features server,pg-sync)
• Local stacked re-measure with fix(merge_thread): close while-loop brace correctly so post-cursor work runs every cycle #224 + perf(prefilter): skip 0-cardinality auto-promoted prefilters #226 — apply mean expected back near post-perf(filter): track dirty values in side set so merge_dirty avoids O(N) walk #222 numbers
• Cap=32 registry no longer hits "full" within minutes of startup

🤖 Generated with Claude Code

[JustMaier]

Reviewed. Approved.

Surgical 16-line guard: if card == 0 { continue } before the registry insert. tracing::debug! for visibility, captures the freq for triage. The comment captures WHY (post-#224 observation, registry pollution → 80-180 ms refresh lock-holds) — exactly the kind of WHY-not-WHAT comment that earns its keep.

No correctness risk: a 0-cardinality prefilter can never substitute a real query (intersection with anything is still empty), so skipping registration is strictly correct.

Same self-approve caveat — treat this comment as approval. Ready to merge after Aidan tag/CI confirm. Pairs with PR #227 to address the regression #224 exposes.