perf(workbook): cache formula-cell volatility with the dependency graph

crates/workbook/benches/baselines.json

@@ -1,127 +1,139 @@
 {
-  "recorded_at": "2026-08-28",
+  "recorded_at": "2026-09-03",
   "recorded_on": "Apple M1 Max (10 core), macOS 14.4, rustc 1.94.1, release profile; best of 5 full bench runs",
-  "note": "ref_units = min(benchmark time) / min(reference time), each the best of 5 full bench runs, not necessarily the same run - best-of-5 minima can (and do) come from different runs. This is a deliberate choice, not an oversight: best-of-5 biases the floor in the safe direction, and the resulting bias in ref_units is small and one-directional (recorded baselines read slightly inflated, so future runs read slightly more negative). The gate compares ratios, not nanoseconds, so a baseline recorded on one machine still means something on another. best_ns_recorded and reference_best_ns_recorded are informational only - nothing is checked against them. The bands are two-sided on purpose: see .github/scripts/check_perf_regression.py. The five multi_sheet entries were recorded 2026-08-29 on the same machine, best of 3 runs. The four incremental_recalc(_cold)/independent_edit_root entries were re-recorded the same day, also best of 3, because this change moves them: Sheet1 was being folded ~16 times per formula cell even in a single-sheet workbook before this change (16.50 folds per extra formula cell, vs 0 after), so the single-sheet incremental family got legitimately faster (paired runs on one machine: base 216,815/212,730/215,348 ns vs head 196,590/196,909/195,420 ns on independent_edit_root/100, about -8%; independent_edit_root/1000 and the _cold variants moved similarly). The previously recorded independent_edit_root/100 baseline (best_ns_recorded 321,871) was already stale relative to both the base and head commits on this machine (both read ~200,000-300,000 ns) - this change added the last several points on top of that pre-existing drift, it did not cause the whole gap. Every other entry is unchanged from the 2026-08-28 recording and remains within run-to-run noise of the base commit. The six chain entries (full_recalc/chain and incremental_recalc/chain_edit_root|chain_edit_leaf, each at 1000 and 5000) were added 2026-08-29 and recorded on the machine named above, best of 5 full bench runs - the count this file's recorded_on documents, not the best-of-3 the two exceptions above used. All five runs measured one unmodified binary, and their reference minimum was 3,196,028 ns - not the reference_best_ns_recorded field above, which belongs to the 2026-08-28 set - so verify these six as best_ns_recorded / 3196028, the same way the 2026-08-29 entries above divide by their own run set's minimum. They are new benchmarks, so no existing entry moved when they were added; every other number here is unchanged from its earlier recording. The incremental_recalc/all_literal_edit/120000cells entry (issue #984) was added 2026-09-03 on the machine named above, best of 5 full bench runs of the two benchmarks alone (not the whole suite) - its reference minimum was 3,335,120 ns, again not the reference_best_ns_recorded field above; verify it as best_ns_recorded / 3335120. It is a new benchmark, so no existing entry moved when it was added.",
+  "note": "Single-run re-record after rebasing #983 (volatile-cell cache) onto main post-#984 (spill-anchor cache) merge \u2014 every entry comes from ONE cargo bench run of the fully-combined code (rustc 1.98.0 (88d9e12ae 2026-08-18), release profile), not a hand-merge of two separate recordings. Includes both PRs new groups: incremental_recalc/row_totals_volatile_seed/{100,10000,100000} (#983) and incremental_recalc/all_literal_edit/120000cells (#984). best_ns_recorded and reference_best_ns_recorded are informational only - nothing is checked against them. The bands are two-sided on purpose: see .github/scripts/check_perf_regression.py.",
   "reference": "calibration/hash_alloc",
-  "reference_best_ns_recorded": 3270207,
+  "reference_best_ns_recorded": 3164388,
   "regression_pct": 150,
   "improvement_pct": 40,
   "benchmarks": {
     "full_recalc/independent/100": {
-      "ref_units": 0.1436,
-      "best_ns_recorded": 469632
+      "ref_units": 0.1391,
+      "best_ns_recorded": 440310
     },
     "full_recalc/independent/1000": {
-      "ref_units": 1.3395,
-      "best_ns_recorded": 4380494
+      "ref_units": 1.2879,
+      "best_ns_recorded": 4075270
     },
     "full_recalc/independent/5000": {
-      "ref_units": 6.9069,
-      "best_ns_recorded": 22587145
+      "ref_units": 7.0781,
+      "best_ns_recorded": 22397889
     },
     "full_recalc/chain/1000": {
-      "ref_units": 1.196,
-      "best_ns_recorded": 3822395
+      "ref_units": 1.2234,
+      "best_ns_recorded": 3871375
     },
     "full_recalc/chain/5000": {
-      "ref_units": 6.5853,
-      "best_ns_recorded": 21046666
+      "ref_units": 6.7395,
+      "best_ns_recorded": 21326255
     },
     "full_recalc/row_totals/500": {
-      "ref_units": 1.3837,
-      "best_ns_recorded": 4524877
+      "ref_units": 1.4669,
+      "best_ns_recorded": 4641820
     },
     "full_recalc/row_totals/2000": {
-      "ref_units": 6.0191,
-      "best_ns_recorded": 19683565
+      "ref_units": 6.3515,
+      "best_ns_recorded": 20098562
     },
     "full_recalc/block_subtotals/2000": {
-      "ref_units": 1.2934,
-      "best_ns_recorded": 4229585
+      "ref_units": 1.3706,
+      "best_ns_recorded": 4337191
     },
     "full_recalc/block_subtotals/10000": {
-      "ref_units": 8.0703,
-      "best_ns_recorded": 26391430
+      "ref_units": 8.2032,
+      "best_ns_recorded": 25957993
     },
     "full_recalc/multi_sheet/200x50": {
-      "ref_units": 13.8314,
-      "best_ns_recorded": 45612155
+      "ref_units": 13.9405,
+      "best_ns_recorded": 44113229
     },
     "full_recalc/multi_sheet_long_names/200x50": {
-      "ref_units": 13.2369,
-      "best_ns_recorded": 43651501
+      "ref_units": 13.3224,
+      "best_ns_recorded": 42157171
     },
     "full_recalc/multi_sheet_cross_refs/200x50": {
-      "ref_units": 13.4762,
-      "best_ns_recorded": 44440765
+      "ref_units": 15.199,
+      "best_ns_recorded": 48095548
     },
     "full_recalc/tall_sparse/20000": {
-      "ref_units": 39.9981,
-      "best_ns_recorded": 130802021
+      "ref_units": 40.8564,
+      "best_ns_recorded": 129285553
     },
     "depgraph_build/tall_sparse/20000": {
-      "ref_units": 9.9993,
-      "best_ns_recorded": 32699704
+      "ref_units": 9.9017,
+      "best_ns_recorded": 31332860
     },
     "depgraph_build/row_totals/2000": {
-      "ref_units": 0.9512,
-      "best_ns_recorded": 3110508
+      "ref_units": 0.9418,
+      "best_ns_recorded": 2980216
     },
     "depgraph_build/multi_sheet_cross_refs/200x50": {
-      "ref_units": 3.2537,
-      "best_ns_recorded": 10729783
+      "ref_units": 3.1996,
+      "best_ns_recorded": 10124723
     },
     "incremental_recalc/independent_edit_root/100": {
-      "ref_units": 0.0605,
-      "best_ns_recorded": 192758
+      "ref_units": 0.0566,
+      "best_ns_recorded": 179180
     },
     "incremental_recalc/independent_edit_root/1000": {
-      "ref_units": 0.3549,
-      "best_ns_recorded": 1130495
+      "ref_units": 0.2876,
+      "best_ns_recorded": 910143
+    },
+    "incremental_recalc_cold/independent_edit_root/100": {
+      "ref_units": 0.0919,
+      "best_ns_recorded": 290789
+    },
+    "incremental_recalc_cold/independent_edit_root/1000": {
+      "ref_units": 0.736,
+      "best_ns_recorded": 2329101
+    },
+    "incremental_recalc/multi_sheet_edit_root/200x50": {
+      "ref_units": 2.5644,
+      "best_ns_recorded": 8114804
     },
     "incremental_recalc/block_subtotals_edit_root/10000": {
-      "ref_units": 1.2095,
-      "best_ns_recorded": 3823493
+      "ref_units": 0.7581,
+      "best_ns_recorded": 2398977
     },
     "incremental_recalc/chain_edit_root/1000": {
-      "ref_units": 1.1638,
-      "best_ns_recorded": 3719666
+      "ref_units": 1.0771,
+      "best_ns_recorded": 3408477
     },
     "incremental_recalc/chain_edit_root/5000": {
-      "ref_units": 6.3891,
-      "best_ns_recorded": 20419766
+      "ref_units": 5.9662,
+      "best_ns_recorded": 18879251
     },
     "incremental_recalc/chain_edit_leaf/1000": {
-      "ref_units": 0.3183,
-      "best_ns_recorded": 1017376
+      "ref_units": 0.2646,
+      "best_ns_recorded": 837166
     },
     "incremental_recalc/chain_edit_leaf/5000": {
-      "ref_units": 1.6274,
-      "best_ns_recorded": 5201330
+      "ref_units": 1.3121,
+      "best_ns_recorded": 4152095
     },
-    "incremental_recalc/multi_sheet_edit_root/200x50": {
-      "ref_units": 3.2604,
-      "best_ns_recorded": 10751809
+    "incremental_recalc/all_literal_edit/120000cells": {
+      "ref_units": 3.8168,
+      "best_ns_recorded": 12077798
     },
-    "from_json/500rows": {
-      "ref_units": 0.2833,
-      "best_ns_recorded": 926457
+    "incremental_recalc/row_totals_volatile_seed/100": {
+      "ref_units": 0.0793,
+      "best_ns_recorded": 250884
     },
-    "to_json/500rows": {
-      "ref_units": 0.2216,
-      "best_ns_recorded": 724738
+    "incremental_recalc/row_totals_volatile_seed/10000": {
+      "ref_units": 7.5273,
+      "best_ns_recorded": 23819279
     },
-    "incremental_recalc_cold/independent_edit_root/100": {
-      "ref_units": 0.0933,
-      "best_ns_recorded": 297138
+    "incremental_recalc/row_totals_volatile_seed/100000": {
+      "ref_units": 96.2568,
+      "best_ns_recorded": 304593843
     },
-    "incremental_recalc_cold/independent_edit_root/1000": {
-      "ref_units": 0.7459,
-      "best_ns_recorded": 2375688
+    "from_json/500rows": {
+      "ref_units": 0.2837,
+      "best_ns_recorded": 897782
     },
-    "incremental_recalc/all_literal_edit/120000cells": {
-      "ref_units": 3.7609,
-      "best_ns_recorded": 12543118
+    "to_json/500rows": {
+      "ref_units": 0.2194,
+      "best_ns_recorded": 694132
     }
   }
 }

crates/workbook/benches/workbook_perf.rs

@@ -499,12 +499,14 @@ fn bench_incremental_recalc(c: &mut Criterion) {
     // Read the two together, not the ratio between them. That ratio is roughly
     // flat (3.7x at n=1000, 3.9x at n=5000 when recorded), because both cases
     // pay the same per-formula-cell fixed cost every incremental recalc owes —
-    // `snapshot_formula_values`, the volatile sweep over `formula_cells()`, and
-    // `seed_spill_sensitive`, each `O(formula cells)` by construction. At
-    // n=5000 the leaf case's time is almost entirely that floor rather than its
-    // one-cell recompute, which is precisely what makes it worth recording: no
-    // other incremental benchmark holds the dirty set fixed while the workbook
-    // grows, so nothing else here can see that floor at all.
+    // `snapshot_formula_values` and `seed_spill_sensitive`, each
+    // `O(formula cells)` by construction (the volatile sweep over
+    // `formula_cells()` was a third such term until issue #983 made it a
+    // cached lookup instead). At n=5000 the leaf case's time is almost
+    // entirely that remaining floor rather than its one-cell recompute, which
+    // is precisely what makes it worth recording: no other incremental
+    // benchmark holds the dirty set fixed while the workbook grows, so
+    // nothing else here can see that floor at all.
     //
     // The regression signal is the *gap*: a leaf closure that silently widened
     // to the whole chain would jump the leaf case onto the root case's number
@@ -579,6 +581,39 @@ fn bench_incremental_recalc(c: &mut Criterion) {
         });
     });
     group.finish();
+
+    // Reproduces the exact shape issue #983 measured: a formula-heavy
+    // document with zero actually-volatile functions, timing a single
+    // unrelated literal write. Before the fix this scaled linearly with
+    // formula count (367ms at n=100,000, dirty closure verified at 1)
+    // because the volatile-seeding loop re-derived every formula cell's
+    // volatility from scratch on every call.
+    //
+    // The fix makes *that term* O(1) in formula count, but the numbers this
+    // group records are not themselves flat: `snapshot_formula_values` and
+    // `seed_spill_sensitive` are two other O(formula cells) passes this same
+    // call still pays on every incremental recalc (see `chain_edit_leaf`'s
+    // comment above, which documents the same floor), and issue #983 does not
+    // touch either. Don't read a still-scaling number here as the fix having
+    // failed — the isolated volatile-rescan term it actually targets is gone;
+    // what remains is the other two terms, tracked separately.
+    let mut group = c.benchmark_group("incremental_recalc/row_totals_volatile_seed");
+    group.sample_size(10);
+    group.measurement_time(Duration::from_secs(5));
+    for n in [100u32, 10_000, 100_000] {
+        let mut template = build_row_totals(n);
+        template.recalc(&ctx);
+        group.bench_with_input(BenchmarkId::from_parameter(n), &n, |b, _| {
+            b.iter(|| {
+                let mut wb = template.clone();
+                let a1 = Address::new(1, 1).unwrap();
+                wb.set("Sheet1", a1, CellInput::Literal(Value::Number(99.0)))
+                    .unwrap();
+                wb.recalc_incremental(&ctx, &[("Sheet1".to_string(), a1)])
+            });
+        });
+    }
+    group.finish();
 }
 
 fn bench_from_json(c: &mut Criterion) {

crates/workbook/src/graph_cache.rs

@@ -85,6 +85,11 @@ pub struct CachedGraph {
     pub(crate) order: Vec<CellRef>,
     /// The cycle set from that same pass.
     pub(crate) cycle: BTreeSet<CellRef>,
+    /// Every formula cell whose formula text names a volatile function
+    /// (`Workbook::is_volatile`), for this exact graph — computed once at
+    /// build time so `recalc_incremental`'s seeding step never re-derives it
+    /// per call (issue #983).
+    pub(crate) volatile: BTreeSet<CellRef>,
 }
 
 impl CachedGraph {

crates/workbook/src/recalc.rs

@@ -305,10 +305,23 @@ impl Workbook {
         // times per incremental recalc (the spill widen loop) while ordering
         // the same graph every time.
         let (order, cycle) = graph.evaluation_order();
+        // Likewise derived from the graph and cached alongside it: a formula
+        // cell's volatility cannot change without the formula text changing,
+        // which already invalidates this cache — so this is the one place
+        // `is_volatile` needs to run, once per formula cell per graph build,
+        // instead of once per formula cell per incremental recalc (issue
+        // #983).
+        let sheets = SheetIndex::build(self);
+        let volatile: BTreeSet<CellRef> = graph
+            .formula_cells()
+            .filter(|cell| self.is_volatile(&sheets, cell))
+            .cloned()
+            .collect();
         let entry = Arc::new(CachedGraph {
             graph,
             order,
             cycle,
+            volatile,
         });
         self.store_cached_graph(entry.clone());
         entry
@@ -412,11 +425,11 @@ impl Workbook {
             }
         }
 
-        // (b) Volatile cells are always dirty (scope ADR Decision 3).
-        for cell in graph.formula_cells() {
-            if self.is_volatile(&sheets, cell) {
-                frontier.insert(cell.clone());
-            }
+        // (b) Volatile cells are always dirty (scope ADR Decision 3). The set
+        // is cached on the graph at build time (issue #983) — no
+        // re-derivation here.
+        for cell in &cached.volatile {
+            frontier.insert(cell.clone());
         }
 
         // Spill-occupancy seeding (issue #591). A cell's spill footprint or