Let an early tap be a weak ball, and let the CPU put a sitter away

Step 4 (b), Oliver's pick. v372 measured why the ladder could not climb:
nearly every lost point was the player's own early whiff, the CPU barely
missed, and the auto-run reached everything, so no opponent dial decided
a point.

- The follow-through: a groundstroke swing meets a bounced ball up to
  60 ms after its window at q 0.05. The early whiff edge goes from ~70 ms
  back to ~130 ms, where the volley window had it before step 1.
- A sitter (a player ball under q 0.5) is put away with a chance that is
  the ladder (0.2 at the bottom of the Climb, 0.85 at the Champion, by
  difficulty in classic): clean, at the line, bounded by the rung's width.
- The player is on his heels for 0.5 s after a put-away of his own sitter.
  Without it the auto-run covered the whole court and no put-away won.
- Fixed a step-2 regression this surfaced: the drop's dig-out cap fired on
  every slice drive, so every low contact against the slicer was a sitter.

Measured at level 26: the Climb now slopes from 95% of points at the
Shambler to 28-39% at the Champion. The net rushers are still easier than
their neighbours. The log names put-aways. Three new tests; v373.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

PROJECT.md

@@ -5585,6 +5585,48 @@ sein, dass durch die geänderte Mechanik alles etwas schwerer geworden ist"*.
   - (c) **Widen the early whiff edge only** (a smaller tail), keeping the error race: undoes most of the
     "harder" without changing the model.
 
+### Step 4 (b), Oliver's pick: the follow-through, and a CPU that can WIN a point (2026-10-14, v373)
+Oliver: *"b"* (options above). Two halves, each useless alone (both were measured alone above):
+- **THE FOLLOW-THROUGH** (`TUNE.swingTail` 0.06 s, rally modes, player, bounced balls only). A swing
+  meets a ball arriving up to 60 ms after its window closed at q 0.05. The early whiff edge goes from ~70
+  ms back to ~130 ms (where the volley window had it before step 1); past it, it is still a whiff. The
+  log says `q=0.05 · early 115ms`. Volleys keep their window: a tail on balls in the air raised the
+  back-court volley share.
+- **THE SITTER IS PUNISHED** (`gameplay.punishes`): a player ball under `punishBelow` (0.5) is answered
+  with a put-away at a chance from `punishMin` 0.2 to `punishMax` 0.85 along the Climb's own progress
+  (`matchStep`, so its brisk bottom counts), classic by `DIFFS[].punish` (0.2/0.5/0.8). A put-away is
+  struck at `punishQ` 0.95, skips the style's soft shots, aims at `punishX` 3.5 m (the line; the rung's
+  width dial still bounds it, so the bottom of the ladder punishes without painting lines), and only
+  counts if no cap took it down (a low volley stays a block). The log says `THEM PUT-AWAY of your weak ball`.
+- **...AND HE IS ON HIS HEELS** (`ball.heels`, `punishRead` 0.5 s): the auto-run stands for that long on a
+  put-away of his own sitter. This is the half that makes it work. The auto-run started on the strike's
+  frame with a perfect read, ~7 m/s plus a 1.75 m reach, and covered the whole court: a q 0.95 put-away
+  to the open court won 0 of 21, at the line 0-2 of 20. Read 0.25 s: ~5%; 0.4: ~20%; 0.55: ~70%.
+  **Feel risk:** he visibly stands still for half a second. If that reads as a bug on the phone, the
+  alternatives are a slowed (not frozen) run, or a shorter read with the put-away hit harder.
+- **A step-2 regression it surfaced:** the drop's dig-out cap read `ball.spin < 1`, and the slicer hits
+  EVERY rally ball at spin 0.8, so every low contact against him (both sides) came back at `lowVolleyQ`.
+  Now `ball.soft && spin < 1`. THE SELLER went from 35% to 47-54% and normal rally lengths.
+- **Measured** (Climb, level 26, 8 seeds × 500 s, point win %):
+
+  | rung | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 |
+  |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+  | σ40 | 95 | 88 | 88 | 87 | 75 | 58 | 52 | 45 | 54 | 48 | 47 | 77 | 42 | 46 | 64 | 40 | 65 | 39 |
+  | σ60 | 95 | 88 | 83 | 78 | 62 | 55 | 53 | 43 | 47 | 46 | 36 | 61 | 30 | 43 | 57 | 28 | 55 | 28 |
+
+  The ladder slopes now, down to the Champion at 28-39%. Classic: easy ~80, medium ~57, hard ~30 (σ60, level 1).
+- **Still open:**
+  - (1) **The net rushers** (Umpire, Club Pro, Director) are 15-25 points easier than their neighbours.
+    They win only ~35-55% of their net points, because the bot ALWAYS passes or lobs a net man, which is
+    their tag's answer played perfectly. The phone decides whether a human does that; if so, the dial is
+    their split-step read (`netReact`) or their put-away from the net.
+  - (2) The bottom three rungs are ~90%. That is easy for level 26, but the Climb is a marathon.
+- Tests (167, nine seeds): the follow-through meets an ~80 ms early tap (and the old edge did not); the
+  put-away share climbs the ladder, goes for the line, never answers a good ball; a low slice drive is not
+  a dug-out drop. The baseline-shape test excludes the answer to a put-away (a man on his heels takes it in
+  the air legitimately), and synthetic returns in the suite now carry `pow: 1` (a stale `pow` from the
+  previous test read as a sitter).
+
 # OPEN — leftovers, in rough priority order
 
 Nothing here is broken; these are known gaps and unverified risks.
@@ -5592,9 +5634,10 @@ Nothing here is broken; these are known gaps and unverified risks.
 0. **THE RALLY GOES FLAT (2026-10-13), the top item.** Every match converged on serve, both at the net,
    reflex volleys, won by a wrong-foot. Step 1 (the court gets its length) SHIPPED v361, awaiting the
    phone verdict; the tap timing + leg knee (v362), the soft ball (v363, drop/lob split v364) and the
-   opponent styles (v364) followed; the swing-log fixes ran to v371. **Step 4 (v372) is BLOCKED on a
-   design decision**: the ladder cannot climb while points are decided by the player's whiffs. See "Step
-   4, measured first" above for options (a)/(b)/(c). Fallback tag `rally-baseline-v360`.
+   opponent styles (v364) followed; the swing-log fixes ran to v371. Step 4: v372 measured why the ladder
+   could not climb (points decided by the player's whiffs); Oliver picked (b), shipped v373 (the follow-
+   through, put-aways, on his heels). Awaiting the phone verdict on v373, especially the half-second
+   flat-footed read and the net rushers. See "Step 4 (b)" above. Fallback tag `rally-baseline-v360`.
 
 1. **`RES` 4 is unverified on a real phone under load.** It nearly doubled the fill rate (810 -> 1080
    device px across). Oliver has not reported stutter, but he also has not played a late horde since.