Combat stats round 3 from arena round 2 margins

Round 2 (with the new EHP-margin logging): mirrors healthy and the
round-1 side bias was noise; drones-vs-cruisers (10%) and
mixed-vs-battlecruisers (14%) near parity; battleship over at +33%
while dreadnought under at -37%; glass-vs-armored narrowed to 11%
(watch); repair escort fair at 11%; two-to-one decisive; the 3x swarm
now cracks the station set at 13%.

Round-3 knobs: weapon_stabilizer range multiplier 1.5 -> 1.3 (range is
the strongest stat under the orbit AI, and the stabilized battleship
was the one loadout still overperforming), plus per-hull HP trims with
the 15 HP/threat anchor demoted to a prior and per-hull HP documented
as the empirical trim knob: battlecruiser 2700 -> 2500, battleship
7500 -> 7000, dreadnought/carrier 15500 -> 19000.

Action items: mirror-bias and EHP-logging items resolved (the latter
implemented upstream); new item to verify the arena end condition seen
idling in round 1 with all ships dead.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DyCu8vwChKMbLJQ3xosYEN

docs/content_design.md

@@ -455,8 +455,12 @@ verifies it empirically. Change the anchors first, re-derive stats second.
   survivability, range, no losses-to-attrition of own DPS) is worth
   paying for, so bigger guns get less raw DPS per threat.
 - **Hull HP = 15 per threat of hull contribution** (incl. ship base
-  time): drone 60, frigate 300, destroyer 550, cruiser 1500,
-  battlecruiser 2700, battleship 7500, dreadnought/carrier 15500.
+  time) as the *prior*; per-hull HP is the empirical trim knob for
+  arena results (guns are shared across many hulls, so gun changes move
+  many matchups at once — hull HP moves exactly one). Current values
+  after round-2 trims: drone 60, frigate 300, destroyer 550,
+  cruiser 1500, battlecruiser 2500, battleship 7000,
+  dreadnought/carrier 19000.
 - **Armor HP ≈ 31 per threat** — a strong premium over hull HP (15)
   because armor is pure HP with no capability, and fights snowball:
   killing enemies removes their DPS, while surviving merely delays —
@@ -482,7 +486,10 @@ verifies it empirically. Change the anchors first, re-derive stats second.
 - **Weapon modifiers are capital economy:** ×1.2 damage at 22.5 threat
   beats adding a gun once a ship carries more than ~68 threat of weapons
   — so modifiers pay off on gun-heavy big hulls and are a waste on small
-  ones. Kept at ×1.2/×1.2/(×1.5 range, ×0.8 rate) deliberately.
+  ones. Damage/rate kept at ×1.2; the stabilizer's range multiplier was
+  cut ×1.5 → ×1.3 after round 2 — range is the strongest stat in the
+  orbit-AI's hands (free approach fire), and the stabilized battleship
+  was the one loadout still overperforming.
 - **Stations:** a fresh player station (3000 HP, 25 dmg × 1 Hz,
   range 120) holds one early parity wave unaided; the enemy station at
   level 0 matches it exactly and scales per push level
@@ -507,6 +514,21 @@ DPS (17→14, 70→52 damage), armor 640→1000, repair 25→12, station range
 still win flawlessly in round 2, the range ladder (kiting) is the next
 suspect, ahead of further DPS cuts.
 
+**Arena round 2 results (2026-07-04, with EHP-margin logging):** the
+mirror "bias" was noise — repeated runs split randomly, and winners
+finish at 5–7% EHP (near-mutual annihilation, healthy). Drones vs
+cruisers 10% and mixed vs battlecruisers 14% — near parity. The two
+capital fights diverged: battleship beat frigates at +33% while the
+dreadnought lost to destroyers at −37% — same guns, so the difference
+is the destroyers carrying the buffed armor plus the battleship's
+stabilizer putting all its guns at 75–150 m against 50 m opponents.
+Glass vs armored narrowed to 11% (watch; if it persists, armor
+1000→1200). Repair escort 11% — fair. Two-to-one decisive at 80%;
+station assault cracked by the 3× swarm at 13% (stations strong but
+breakable). Round-3 knob changes: stabilizer range ×1.5→×1.3, and
+per-hull HP trims — battlecruiser 2700→2500, battleship 7500→7000,
+dreadnought/carrier 15500→19000.
+
 ## Balancing targets (first pass, July 2026)
 
 The six root numbers for the balancing pass. Every derived value (threat