perf(gateway): pool backend conns; loadtest evaluate hot path

The loadtest harness never modelled game.evaluate — the debounced per-tile
play preview a real client fires several times per turn, the hottest gameplay
call. Model it (one evaluate per placed tile + reconsideration re-previews +
draft.save, human-paced; --eval / --eval-recon toggle it).

That realistic load surfaced the real bottleneck: the gateway's backend HTTP
client used the default transport (MaxIdleConnsPerHost=2), so every sync call
to the single backend host churned a fresh TCP connection — ~26500 TIME_WAIT
sockets at 500 players (near the ephemeral-port ceiling), burning ~1.75 gateway
cores while the backend sat near-idle. It was the unfixed root of the residual
transport_error the earlier passes chased on the client side.

Widen the keep-alive pool (backendMaxIdleConns=512, ~2x the observed 225-conn
peak). At 500 players the churn collapses to ~0 and peak gateway CPU drops ~7x
(~1.75 -> ~0.26 cores); postgres (~1.65 cores) becomes the busiest service.
This overturns the earlier "gateway is the binding constraint, scale it
horizontally" sizing — that was sizing around this bug, not a real floor.

Consolidate the loadtest trip reports into one loadtest/REPORT.md (drop the
R2/R7 split) and bake the finding into README / PRERELEASE / ARCHITECTURE /
TESTING.

loadtest/internal/scenario/scenario.go

@@ -42,19 +42,35 @@ type RealisticConfig struct {
 	GamesPerPlayer int           // target concurrent games per player; 0 => random 3..5
 	Tick           time.Duration // per-player operation cadence (keeps a player under the per-user limit)
 	SecondaryProb  float64       // chance per tick of a non-move operation
+	Eval           bool          // model the per-tile evaluate preview (the gameplay hot path); false reproduces the pre-evaluate harness
+	EvalRecon      int           // extra full-composition evaluate re-previews per play, beyond one per placed tile
 }
 
 // DefaultRealistic returns the moderate ramp: 50 -> 200
-// -> 500 concurrent players, ~12 minutes per step, ~1 op/s per player.
+// -> 500 concurrent players, ~12 minutes per step, ~1 op/s per player, with the
+// per-tile evaluate preview modelled (the realistic hot path).
 func DefaultRealistic() RealisticConfig {
 	return RealisticConfig{
 		Steps:         []int{50, 200, 500},
 		StepDur:       12 * time.Minute,
 		Tick:          800 * time.Millisecond,
 		SecondaryProb: 0.08,
+		Eval:          true,
+		EvalRecon:     1,
 	}
 }
 
+// evalGapBase and evalGapSpan bound the modelled pause between successive tile
+// placements: the client's 250 ms debounce coalesces faster drags into a single
+// evaluate, so a thoughtful player's previews are spaced by a gap drawn from
+// [base, base+span] — wide enough that a normal composition stays under the per-user
+// rate limit, the way a real one does (the limiter's cost is measured by the hammer,
+// not by self-inflicted rejections here).
+const (
+	evalGapBase = 250 * time.Millisecond
+	evalGapSpan = 500 * time.Millisecond
+)
+
 // RunRealistic runs the staged ramp. Each step activates more players (drawn from the
 // seeded pool), assembles a cohort of games for them and starts their turn loops; the
 // loops run until the whole ramp ends. Players from earlier steps keep playing, so
@@ -128,7 +144,7 @@ func (d *Driver) playerLoop(ctx context.Context, p seed.Account, games []*Game,
 				d.secondaryOp(ctx, c, p, g, rng)
 				continue
 			}
-			if d.playTurn(ctx, c, p, g, rng) {
+			if d.playTurn(ctx, c, p, g, cfg, rng) {
 				active = slices.DeleteFunc(active, func(x *Game) bool { return x == g })
 				gi = 0
 				if len(active) == 0 {
@@ -161,10 +177,10 @@ func (d *Driver) subscribeLoop(ctx context.Context, c *edge.Client, p seed.Accou
 }
 
 // playTurn plays one turn in g over the player's client when it is the player's
-// move: fetch state, replay history, pick a legal move and submit it (or exchange /
-// pass). It reports whether the game has finished, so the caller can drop it from the
-// rotation.
-func (d *Driver) playTurn(ctx context.Context, c *edge.Client, p seed.Account, g *Game, rng *rand.Rand) (finished bool) {
+// move: fetch state, replay history, pick a legal move, compose it (the per-tile
+// evaluate previews a real client fires) and submit it (or exchange / pass). It reports
+// whether the game has finished, so the caller can drop it from the rotation.
+func (d *Driver) playTurn(ctx context.Context, c *edge.Client, p seed.Account, g *Game, cfg RealisticConfig, rng *rand.Rand) (finished bool) {
 	seat := g.seatOf(p.ID.String())
 	if seat < 0 {
 		return false
@@ -196,6 +212,7 @@ func (d *Driver) playTurn(ctx context.Context, c *edge.Client, p seed.Account, g
 	}
 	switch action.Kind {
 	case "play":
+		d.composePlay(ctx, c, p, g, action.Tiles, cfg, rng)
 		t0 = time.Now()
 		_, code, _ := c.SubmitPlay(ctx, p.Token, g.ID, action.Tiles)
 		d.rec.Record("game.submit_play", code, time.Since(t0))
@@ -211,6 +228,59 @@ func (d *Driver) playTurn(ctx context.Context, c *edge.Client, p seed.Account, g
 	return false
 }
 
+// composePlay models a player arranging the chosen play tile by tile before committing:
+// the debounced evaluate preview the real client fires on each placement (a growing prefix
+// of the tiles), a few full-composition re-previews for reconsideration (recall a tile, try
+// another spot), and the single draft persistence the client debounces out. evaluate is the
+// hottest gameplay request at scale, so omitting it (the pre-evaluate harness) understated
+// the load; cfg.Eval false reproduces that baseline for an A/B comparison. Every step
+// honours ctx, so end-of-run cancellation never blocks on a sleep or an in-flight preview.
+func (d *Driver) composePlay(ctx context.Context, c *edge.Client, p seed.Account, g *Game, tiles []edge.PlayTile, cfg RealisticConfig, rng *rand.Rand) {
+	if !cfg.Eval || len(tiles) == 0 {
+		return
+	}
+	// One evaluate per landed tile: the growing prefix mirrors the client re-previewing
+	// after each placement (an early prefix is often illegal, which is still a successful
+	// "ok" round trip — exactly the backend work a real composition triggers).
+	for n := 1; n <= len(tiles); n++ {
+		if !jitterSleep(ctx, rng, evalGapBase, evalGapSpan) {
+			return
+		}
+		t0 := time.Now()
+		code, _ := c.Evaluate(ctx, p.Token, g.ID, tiles[:n])
+		d.rec.Record("game.evaluate", code, time.Since(t0))
+	}
+	for r := 0; r < cfg.EvalRecon; r++ {
+		if !jitterSleep(ctx, rng, evalGapBase, evalGapSpan) {
+			return
+		}
+		t0 := time.Now()
+		code, _ := c.Evaluate(ctx, p.Token, g.ID, tiles)
+		d.rec.Record("game.evaluate", code, time.Since(t0))
+	}
+	// The client persists the in-progress composition (debounced to one upsert). Its opaque
+	// JSON content does not affect the call's cost, so a minimal valid shape stands in.
+	t0 := time.Now()
+	code, _ := c.DraftSave(ctx, p.Token, g.ID, `{"rack_order":"","board_tiles":[]}`)
+	d.rec.Record("draft.save", code, time.Since(t0))
+}
+
+// jitterSleep pauses for a randomised gap in [base, base+span], modelling the human pause
+// between tile placements that the client's debounce coalesces into one evaluate. It
+// returns false if ctx is cancelled during the wait, so a composition unwinds promptly at
+// end of run.
+func jitterSleep(ctx context.Context, rng *rand.Rand, base, span time.Duration) bool {
+	d := base + time.Duration(rng.Int63n(int64(span)+1))
+	t := time.NewTimer(d)
+	defer t.Stop()
+	select {
+	case <-ctx.Done():
+		return false
+	case <-t.C:
+		return true
+	}
+}
+
 // secondaryOp exercises one of the non-move edge operations the plan calls out, so
 // the run touches nudge / chat / check-word / draft / profile / stats too, over the
 // player's own client.