perf(planning): build placement date keys directly

docs/benchmark/pr-278-date-keys.md

@@ -0,0 +1,53 @@
+# PR #278: build placement date keys directly
+
+Measured on 2026-10-16, starting from PR #278 at `cbde23f`.
+
+The unchanged benchmark finished in 299.372 seconds before this change, leaving
+only 0.628 seconds before its 300-second timeout. A separate loaded-year CPU
+profile attributed 8.27 of 66.57 seconds to `placementTieKey`, including its
+callees. That function runs when looking up cached calendar observations and
+when comparing final placement dates.
+
+The one-line optimization sorts the placement map's date keys directly, using
+native string sorting. It removes the intermediate placement-to-date array and
+the comparison callback. Every placement map entry is keyed by its placement's
+date, and native string sorting preserves the existing lexical date order, so
+the resulting keys are identical. Search choices and scoring are unchanged.
+
+## Before and after
+
+Both runs used Node 24.18.0, Bun 1.3.14, Vitest 5.0.3, the same local machine,
+and the exact command below. No other test, build or profiling workload ran
+alongside either measurement.
+
+```sh
+NODE_OPTIONS=--max-old-space-size=2048 timeout 300s bunx vitest run src/app/benchmark/leave-planner-benchmark.spec.ts --maxWorkers=1
+```
+
+All seven warm-ups, fourteen measured calculations, the typed-obligation precheck,
+fixtures and assertions remain unchanged. Coverage is disabled for this command.
+
+| Measurement                                    |    Before |     After |
+| ---------------------------------------------- | --------: | --------: |
+| Entire command, external wall time             | 299.372 s | 268.573 s |
+| Benchmark test                                 | 293.620 s | 263.265 s |
+| Loaded year, mean of two measured calculations |  76.649 s |  67.747 s |
+| Headroom before timeout                        |   0.628 s |  31.427 s |
+
+The entire command was 10.3% faster in this before-and-after pair. Both runs
+completed successfully, and every repeated scenario remained deterministic.
+These are local measurements; Gitea runner performance still needs confirmation.
+
+## Correctness checks
+
+Complete outcomes, stable explanations and final-state snapshots match the
+baseline for all seven scenarios. These separate comparisons ran alongside
+functional checks; their diagnostic timings are not used in the table above.
+
+The functional coverage suite passes all 635 tests across 33 files. Application
+and spec type checks, lint, source formatting and `git diff --check` also pass.
+No browser timing run or live CI rerun was performed for this change.
+
+The profiling runner and full protocol logs are retained in
+`/tmp/vac-opt278-measure`. The original diagnostic reproduction is documented in
+[issue-78.md](issue-78.md#reproduction).

src/app/services/leave-planner.core.ts

@@ -3317,10 +3317,7 @@ function preferenceStateKey(state: State): string {
 }
 
 function placementTieKey(state: State): string {
-	return [...state.placements.values()]
-		.map((placement) => placement.date)
-		.sort(compareDate)
-		.join('|');
+	return [...state.placements.keys()].sort().join('|');
 }
 
 function expiryTieKey(state: State): string {