fire-planner: Wave 2 chart-first — flex spending, categorised life

events, interactive Visx Gantt + spending-profile chart

Charts are now the primary editor for life events. The Plan-tab body
re-orders to make charts ~80% of viewport real-estate; legacy form
sections are collapsed into a drawer.

Backend:
- alembic 0004: life_event.category enum (essential / discretionary /
  not_spending). Defaults to essential so existing rows keep their
  full spending impact.
- Simulator gains discretionary_outflows + flex_rules params. Tracks
  per-path running ATH, applies the deepest applicable cut to
  discretionary outflows when portfolio drops vs ATH (PLab-style flex
  spending). Cut amount stays in the portfolio (refund pattern).
- New flex_spending module with FlexRule + applicable_cut +
  cuts_per_year (vectorised). Sortable rules; "deepest cut wins" so
  users specify cumulative cuts at each tier.
- New /scenarios/{id}/spending-profile endpoint returning per-year
  base / essential / discretionary / flex_cut / total breakdown.
- SimulateRequest gains flex_rules + life_event.category roundtrip.
- 8 new tests; 246 total pytest pass; mypy + ruff clean.

Frontend (Visx + ECharts):
- Installed @visx/{scale,shape,group,axis,event,responsive,tooltip}
  for native SVG drag interactions.
- New <SpendingProfileChart> — Visx stacked-area of base/essential/
  discretionary with red flex-cut overlay, hover tooltip, click-to-
  scrub-year.
- New <EventGantt> — interactive Visx Gantt:
    * Click empty space → popover create at that year (default
      essential spending event)
    * Click a bar → inline edit popover (name, kind, range, £/y,
      category) with delete button
    * Drag bar middle → moves the whole event (year-resolution snap)
    * Drag bar edges → resizes year_start / year_end
    * All gestures persist via PATCH /life-events/{id}
- New <FlexRulesEditor> — list of {from_ath_pct, cut} tiers, save-on-
  change to scenario.config_json.flex_rules.
- Plan-tab redesign: NW fan dominant top with floating stat badges
  (Year/Age/NW/Δ NW/Spending/Eff. tax) over the chart; spending-
  profile chart middle; Gantt bottom; flex-rules editor; legacy form
  sections in a collapsed <details> drawer.
- Frontend typecheck + 7 vitest tests + production build all clean.

fire_planner/simulator.py

@@ -33,6 +33,7 @@ from decimal import Decimal
 import numpy as np
 import numpy.typing as npt
 
+from fire_planner.flex_spending import FlexRule, applicable_cut
 from fire_planner.glide_path import GlideFn
 from fire_planner.strategies.base import StrategyState, WithdrawalStrategy
 from fire_planner.tax.base import TaxInputs, TaxRegime
@@ -186,6 +187,8 @@ def simulate(
     bucket_split: _BucketSplit = default_bucket_split,
     income_inflows: npt.NDArray[np.float64] | None = None,
     income_taxable: npt.NDArray[np.float64] | None = None,
+    discretionary_outflows: npt.NDArray[np.float64] | None = None,
+    flex_rules: list[FlexRule] | None = None,
 ) -> SimulationResult:
     """Run the MC simulation. `paths` shape: (n_paths, n_years, 3).
 
@@ -223,6 +226,11 @@ def simulate(
         income_inflows = np.zeros(n_years, dtype=np.float64)
     if income_taxable is None:
         income_taxable = np.zeros(n_years, dtype=np.float64)
+    if discretionary_outflows is None:
+        discretionary_outflows = np.zeros(n_years, dtype=np.float64)
+    rules = list(flex_rules) if flex_rules else []
+    # Track running ATH per path so we can decide flex cuts each year.
+    ath = np.full(n_paths, float(initial_portfolio), dtype=np.float64)
 
     for y in range(n_years):
         alloc = glide(y)
@@ -246,6 +254,19 @@ def simulate(
             income_tax_breakdown = regime_at(y).compute_tax(
                 TaxInputs(earned_income=Decimal(str(round(float(income_taxable[y]), 2)))))
             portfolio = portfolio - float(income_tax_breakdown.total)
+        # Flex spending: per-path, decide the cut from this year's
+        # drawdown-from-ATH and refund the trimmed discretionary
+        # back to the portfolio. The cashflow_adjustments array already
+        # subtracted the *baseline* discretionary, so we add back
+        # `cut_pct * baseline` to leave only the post-cut amount drawn.
+        if rules and discretionary_outflows[y] > 0.0:
+            for p in range(n_paths):
+                if ath[p] <= 0:
+                    continue
+                drawdown = max(0.0, 1.0 - portfolio[p] / ath[p])
+                cut = applicable_cut(drawdown, rules)
+                if cut > 0:
+                    portfolio[p] += cut * float(discretionary_outflows[y])
 
         # Strategy is per-path Python — 600k iterations at 60y × 10k paths.
         # Profiled: ~3 seconds for the full Trinity / GK / VPW set.
@@ -280,6 +301,9 @@ def simulate(
 
         portfolio_history[:, y + 1] = np.clip(portfolio, a_min=0.0, a_max=None)
         portfolio = portfolio_history[:, y + 1]
+        # Update running ATH per path so next year's flex decision uses
+        # the post-close peak.
+        np.maximum(ath, portfolio, out=ath)
 
     # Success = portfolio stayed positive through every interim year.
     # Excludes the very last year-end because VPW deliberately drains