fix(warm_registry): record original filters, not post-prefilter-substitution

[JustMaier]

Summary

Warm registry persist had been silently failing on every diverse-shape query batch. server.rs:2787 overwrites query.filters with the post-prefilter-substitution form when a prefilter matches, prepending a FilterClause::BucketBitmap { bitmap: Arc<RoaringBitmap>, .. } clause. Later at server.rs:2817 we recorded &query.filters — the now-substituted version — into the warm registry. When the merge thread tried to serialize, serde rejected BucketBitmap (#[serde(skip)]) and the entire batch was dropped:

warm registry persist failed: the enum variant
FilterClause::BucketBitmap cannot be serialized

Symptom

887,596 queries over 5 min in the post-PR-#224 regression-resolution run → zero successful warm persists, warm.json stayed at the previous session's 1-entry state, auto-warm on next boot restored 1 stale entry instead of a meaningful warm window.

Fix

Capture original_filters_for_warm = query.filters.clone() BEFORE the substitution call, pass that to warm_registry().record(...). The substituted form is still used for execution; only the shape recorded into the persistent registry comes from the user's input.

14 insertions, 2 deletions. No behavior change on the executor or planner.

Verification

Same replay corpus that exposed the bug:

Metric	Pre-fix	Post-fix
warm registry: persisted N shapes events	0 / 5 min	10+ events, fires every merge cycle
warm registry persist failed: BucketBitmap	9+ / 5 min	0
warm.json size	221 B (1 stale entry)	410 KB (187 unique shapes)
Top entry frequency in warm.json	1	6,248

Stacking

• Branched off main directly. No dependency on other open PRs.
• Independent of perf(snapshot): chunk for_each_versioned to bound lock-hold under high cardinality #225 (range_scan/save chunked iteration, still parked).
• Unblocks the time-to-first-warm-restore measurement that was deferred in docs/_in/regression-resolution-2026-04-25.md.

Test plan

• Build clean (cargo build --profile fast --features server,pg-sync)
• Live verification with 273K-query replay → warm.json populated correctly with 187 entries
• Time-to-first-warm-restore measurement post-merge (separate task)

🤖 Generated with Claude Code

[JustMaier]

Reviewed. Approved.

Surgical fix. Capture original_filters_for_warm = query.filters.clone() before prefilter substitution at line 2787, then thread it through the record call at line 2823 instead of the post-substitution query.filters. Two-comment WHY captures the trap (BucketBitmap serde-skip → serde panic → silent batch drop) — exactly the kind of comment that will save someone hours when they look at this code in 6 months.

Verification table is striking — 0 → every cycle persist firing, warm.json 221 B → 410 KB, 1 → 187 unique shapes. Bug was eating every diverse-shape persist for who-knows-how-long.

No correctness risk on the executor side — substitution still happens AFTER capture, so the planner/executor still sees the optimized form. Only the warm registry sees the original.

One observation (not blocker): clone() on the filter Vec is per-query overhead. Filters are usually shallow (few clauses), so cheap. If a future profiler shows it as hot, swap to recording the canonical form before substitution + reuse the canonical (which is already serializable). Out of scope for this fix.

Approved. Same self-approve caveat — treat this comment as approval. Merging next + tagging v1.0.171-jemalloc.