feat(deepfall): live 6-agent playtest + playtest-driven fixes

Ran an observed 6-player game with persona AI players (one per
Lineage/Calling) via a Workflow (72 agents): avgFun 7.3/10, 6/6 would
replay, tight 84-82-75 finish. PLAYTEST.md captures the session +
designer's assessment.

Findings the Monte-Carlo bots structurally missed:
- the careful/safe line feels UNFAIR, not just weak (banker finished
  last and resented it) — three human angles confirm the sim's
  Cash-vs-Worthy substitution finding
- humans hit the no-Last-Stand feel-bad 6x where rational bots hit it 0x
- deep-collapses-first geometry confuses newcomers (component fix)

Fixes (re-validated, 1000 seeds): WorthyEnd strategy 46.7 -> 42.5%,
safe lines lifted, 6p margin to 9.5%.
- Desperate Strike: a no-Last-Stand Take salvages 1/3 unbanked loot
- Extraction bonus: +3 glory per crew retired alive
- RULEBOOK: "Maw eats NEXT" indicator as a component rule

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

deepfall/PLAYTEST.md

@@ -0,0 +1,99 @@
+# DEEPFALL — Live 6-Player Playtest (2026-06-19)
+
+A complement to the Monte-Carlo balance sweeps (`FINDINGS.md`): six **persona AI
+players** — each a different *kind* of human, each piloting a different one of the six
+Lineages and six Callings — sat at a virtual table and played a full game, making real
+in-character decisions refereed by a faithful rules engine, then reflected on the fun.
+
+This is the thing the bots can't measure: **how it *feels*.**
+
+## The table
+
+| Player | Archetype | Company | Final |
+|---|---|---|---|
+| **Vex** | the spoiler (competitive, denial) | Dvergar Hierophant | 🥇 84 |
+| **Rook** | the gambler (chaos, push-luck) | Sylphid Wardancer | 🥈 82 |
+| **Mina** | the min-maxer (cold optimizer) | Wisp Prospector | 🥉 75 |
+| **Pip** | the first-timer | Gnoll Reaver | 59 |
+| **Tav** | the roleplayer (story over winning) | Revenant Lorekeeper | 50 |
+| **Greta** | the cautious banker | GolemBorn Warden | 30 |
+
+11 rounds, clean ending on Floor 1. Top three within **9 glory** (84-82-75), 2 lead
+changes, 5 contested collapses, 2 Worthy Ends.
+
+## The verdict
+
+**avgFun 7.3/10 · 6 of 6 would play again.** The game is *fun* — its problems are
+balance and clarity, not fun. The drama landed exactly where it was designed to:
+
+- **Rook's round-8 Last Stand** — Maw one tick away, zero loot, stands up and declares
+  anyway, rides the Hall collapse to 22 glory while a rival is swallowed for nothing:
+  *"The Maw blinked first."* Their favorite moment of the night. **The three-exit triangle
+  produces real table theatre.**
+- **Pip (the newbie) literally screamed** when a Den he'd half-forgotten he was holding
+  paid out 10 glory at collapse. *"I did not understand what happened. Everyone looked at
+  me. Worth it."* **The collapse-scoring surprise is delightful even when you don't fully
+  get the rules.**
+- **Tav (the roleplayer)** stood in a burning Library as the floor fell and walked out
+  with 24 glory of "ancient marginalia." The theme carries.
+
+## What it caught (and how it maps to the bot data)
+
+**1. The careful line is under-rewarded — and it feels *unfair*, not just weak.**
+Greta never lost a crew and finished *last*: *"The scoring punishes the careful and
+rewards the reckless. That is not a game, that is a bad lesson."* Mina played the
+provably-optimal cash-out line, found it **joyless** (*"that's not daring, that's
+invoicing"*), and still lost to riskier players. Even **the winner, Vex**, flagged it:
+*"Rook got eaten twice and still finished two points behind me — that gap is not
+satisfying."* → This is exactly the sim's `CashOut`-strategy weakness (2.8%) and the
+open Cash-vs-Worthy *substitution* finding — now confirmed from **three independent human
+angles**, with the emotional weight the win-rate number can't convey.
+
+**2. Humans hit the feel-bad 6× where the bots hit it 0×.** The Monte-Carlo bots cash out
+defensively and almost never get Taken without a Last Stand (0.00/game). But the
+*gambler*, the *roleplayer*, and the *newbie* walked into the silent death **six times**
+this game. Rook and Tav independently demanded the same thing: *"Every death deserves a
+punchline."* **This is the headline meta-finding: the live-agent playtest caught a problem
+the perfect-play simulator structurally hid.**
+
+**3. The deep-collapses-first geometry confuses newcomers.** Pip, round 5: *"Nobody told
+me the floors count DOWN. I thought it went up. Why does it go down. WHO DESIGNED THIS."*
+Got caught thinking he had five turns when he had one. → The fun-critic predicted exactly
+this ("richest = doomed first" fights loss-aversion intuition); the playtest makes it a
+hard **component requirement**.
+
+**4. Contests feel flat.** Mina and Vex both noted that piling markers into a room is a
+free hedge and ties split evenly: *"two markers beats four markers because the split is
+flat — that is irrational"* (Mina); *"flooding a room should be a genuine gamble"* (Vex).
+
+## Changes made (and re-validated in the sim)
+
+| Change | Why | Knob |
+|---|---|---|
+| **Desperate Strike** — a no-Last-Stand Take salvages ⅓ of unbanked loot | Rook & Tav: a silent death shouldn't be a humiliating zero | `taken_salvage_frac = 0.33` |
+| **Extraction bonus** — flat +3 glory per crew retired alive | Greta: reward surviving the delve, not just dying deep | `cashout_extraction_bonus = 3.0` |
+
+Re-sweep (1000 seeds): the strategy band tightened — WorthyEnd **46.7 → 42.5%**, the
+safe/engine lines lifting (PressOn 25.8 → 31.9, CashOut 2.8 → 3.7); 6-player victory
+margin fell to **9.5%** (even tighter top-of-table). Pacing, combo spread, and the
+0.00-feel-bad-for-bots all held.
+
+## Recommended next (not yet built)
+
+- **Marker-weighted contest shares** (Mina) — split a contested room's Glory in proportion
+  to marker strength, and hand every contestant a Dread token (Vex) so piling in is a real
+  gamble. The cleanest fix for finding #4, and it further differentiates Warden.
+- **The "Maw eats NEXT" indicator** (Pip) — a big, unmissable physical marker on the next
+  floor to collapse. A *component* fix for the onboarding cliff; folded into the rulebook.
+- **The deeper Cash-vs-Worthy fix** (finding #1) — the flat extraction bonus helps but
+  doesn't fully solve it. The real lever (per `FINDINGS.md`) is making *denial* score for
+  you, or paying Worthy End on *cumulative* depth dived. The next prototype.
+
+## The bottom line for 6 humans
+
+It **scales** — a clean 11-round game with a tight 3-way finish and real lead changes —
+and it **delivers the fantasy**: the gambler got their blaze of glory, the newbie got their
+scream, the roleplayer got their story, the spoiler got their win by playing the table.
+With the survival incentive, the Desperate Strike, and a big "Maw-eats-next" marker, this
+is a table-ready 6-player game. The one thing still worth a design pass before a print run
+is making the *careful* player feel as clever as the reckless one feels alive.

deepfall/RULEBOOK.md

@@ -239,6 +239,11 @@ fuel. This is your in-game engine (Quarriors' dice-building).
 
 ## 9. The Maw & player count
 
+> **Component rule (playtest-critical):** the Maw marker must **loudly flag the next floor
+> to collapse.** New players reliably forget the dungeon collapses *bottom-up* and think
+> they have more time than they do — give the Maw a big, unmissable "EATS THIS FLOOR NEXT"
+> pointer. (See `PLAYTEST.md`, finding #3.)
+
 The Maw rises **1 floor per round**, **+1** whenever the shared **Dread track** fills (then
 it resets). Players add Dread to the track every time they **Press On** and whenever
 uncanceled ☠ is rolled — so the table's collective greed literally pulls the doom down

deepfall/configs/default.json

@@ -19,6 +19,8 @@
 
   "cashout_depth_dividend": 2.0,
   "cashout_min_bank": 12.0,
+  "cashout_extraction_bonus": 3.0,
+  "taken_salvage_frac": 0.33,
   "worthy_factor": 2.5,
   "worthy_held_bonus": 0.0,
 

deepfall/deepfall-core/src/config.rs

@@ -47,6 +47,12 @@ pub struct Config {
     pub cashout_depth_dividend: f64,
     /// The Cautious bot Cashes Out once its bankable value clears this bar.
     pub cashout_min_bank: f64,
+    /// Cash-Out flat bonus per crew retired "with boots on" — a survival incentive the
+    /// 6-human playtest demanded (the careful banker finished last and felt cheated).
+    pub cashout_extraction_bonus: f64,
+    /// Fraction of unbanked loot salvaged when Taken with **no** Last Stand — the
+    /// "Desperate Strike" so a silent death isn't a humiliating zero (playtest ask).
+    pub taken_salvage_frac: f64,
     /// Worthy-End spike = `worthy_factor * deepest_floor` (+ optional held bonus).
     pub worthy_factor: f64,
     /// Extra Worthy-End credit per point of held room value. **0 by default** to
@@ -97,6 +103,10 @@ impl Default for Config {
             // shallow parties) and cratered to a 1.8% win rate; banking richer hauls
             // before retiring makes the safe line viable.
             cashout_min_bank: 12.0,
+            // Playtest-driven (2026-06-19): reward surviving the delve, and give a silent
+            // death a small consolation. See PLAYTEST.md.
+            cashout_extraction_bonus: 3.0,
+            taken_salvage_frac: 0.33,
             // v2 (iter 3-4): 4 -> 2.5. At 4 the generic Worthy End (any deep party on a
             // Throne) made the WorthyEnd strategy a 45% king; lowering the base while
             // raising the Wardancer multiplier (now 2.2x) shifts the payoff onto the

deepfall/deepfall-core/src/game.rs

@@ -542,7 +542,9 @@ impl Game {
     /// Cash Out: bank loot + depth dividend, convert held markers to Fallen, retire.
     fn cash_out(&mut self, pi: usize) {
         let party = self.players[pi].party.take().unwrap();
-        let raw = party.unbanked + self.cfg.cashout_depth_dividend * party.deepest as f64;
+        let raw = party.unbanked
+            + self.cfg.cashout_depth_dividend * party.deepest as f64
+            + self.cfg.cashout_extraction_bonus;
         let banked = raw * self.comeback_mult(pi);
         self.players[pi].glory += banked;
         self.players[pi].cashouts += 1;
@@ -598,7 +600,12 @@ impl Game {
                 }
             }
         } else {
-            // The one true feel-bad: lose unbanked loot and all active territory.
+            // Taken with no Last Stand. Still the worst exit (territory lost, no depth
+            // dividend), but a "Desperate Strike" salvages a fraction of unbanked loot so
+            // it isn't a humiliating zero — the playtest's most-requested change.
+            let salvage = party.unbanked * self.cfg.taken_salvage_frac;
+            let cm = self.comeback_mult(pi);
+            self.players[pi].glory += salvage * cm;
             self.players[pi].takes_no_laststand += 1;
             for (d, ri) in &party.held {
                 let fi = Self::floor_idx(*d);