docs(amiga-ball): criterion 22 measured the burst dying, not the filter closing

Iteration 2's finding, ruled on and applied.  The old wording asked for
the energy centroid to fall by a factor of three; measured over eleven
captures, that fall tracks the 2-10 kHz band's 30 dB point exactly, and a
centroid with voice 1's fundamental excluded moves 1.12x across the whole
window.  So the criterion improved as the demo got worse -- it scored best
at $D40C = $00, where there is no transient for the sweep to act on at
all.  A criterion that rewards deleting the thing it checks is not
measuring that thing.

It now measures the transient's own band: 2-10 kHz must take at least
150 ms to fall 30 dB, so the burst outlives more than half the 267 ms
sweep.  This is a strengthening, and the test of that is that it can fail
and does -- the same measurement on the immediately preceding $D40C = $04
build reads 95 ms.  The shipped build reads 175 ms on both surfaces, and
the wall's centroid is above the floor's at 21 of 21 steps to +200 ms.

Section 13.2 gains the capture-window jitter note: the WAV varies by +/-2
video frames and the impact lands 353-387 ms in, while sid-log.jsonl is
md5-identical run to run.  Anything measured against a fixed grid in the
WAV is partly measuring where the window opened, which is what made
iteration 1's figures irreproducible.

Tally 28 PASS / 0 FAIL.  No third iteration: what remains is the listen
section 13.2 reserves for the maintainer.

README figures corrected against the shipped build -- irq_hwm is 14 on the
gate-on frame and 7 on an ordinary one, and profile reads 450-480 rather
than the 481 quoted from a single run.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WuuVARfoWgx7BzWDyGXH15

Author: nschneir

demos/amiga_ball/AUDIT.md

@@ -909,3 +909,44 @@ Everything else is closed. And the last gate is still §13.2's and not a
 spectrogram's: **the maintainer's listen of `capture.wav`.** This iteration can
 say the transient is 1.84× longer and that the sweep now has a signal under it
 for two thirds of its length. It cannot say whether that is a better boing.
+
+---
+
+## Criterion 22, re-worded and re-scored (maintainer ruling)
+
+Both proposals from iteration 2 were accepted and applied to `SPEC.md`:
+criterion 22 now measures the **transient's own band** rather than the
+whole-signal centroid, and §13.2 records the capture-window jitter.
+
+The re-wording is not a weakening, and the test of that is that the new
+criterion **can fail and does** — the same measurement on the immediately
+preceding `$D40C` = `$04` build reads 95 ms against a floor of 150 ms. The old
+wording had the opposite property: it improved as the demo got worse, scoring
+best at `$D40C` = `$00`, where there is no transient for the sweep to act on at
+all. A criterion that rewards deleting the thing it exists to check is not
+measuring that thing.
+
+| # | Claim (as re-worded) | Verdict | Evidence |
+|---|---|---|---|
+| 22 | broadband column in one frame; 2-10 kHz takes ≥ 150 ms to fall 30 dB; wall centroid above floor at every 10 ms step of the first 200 ms | **PASS** | 2-10 kHz −30 dB at **+175 ms** on both floor and wall (`$04` build: +95 ms, which fails). Wall above floor at **21 of 21** steps to +200 ms. Ordering intact: transient 175 ms under thump 250 ms (floor) / 200 ms (wall) |
+
+**Tally: 28 PASS, 0 FAIL, 0 INCONCLUSIVE.**
+
+Final measured figures on the shipped build (session `ballfin`, 700 ticks):
+`irq_hwm` = **14** raster lines (gate-on frame), `irq_last` = **7** (ordinary
+frame), `c64 profile tick` = **451.6 cycles** mean over 16 arrivals (min 401,
+max 499). `10 + 14 = 24 < 51`, so the tick still finishes in the top border
+before the display begins.
+
+**Is a third iteration needed? No.** Every criterion passes on evidence from
+the running machine, and the two document defects iteration 2 raised are
+applied rather than deferred. What remains is not an iteration but the one gate
+this process cannot close for itself: §13.2 makes the maintainer's listen of
+`evidence/audio/floor/capture.wav` and `wall/capture.wav` the final word on
+whether it sounds like a boing. Everything measurable about that sound has been
+measured and is above.
+
+One item stays open outside this demo and is filed in `docs/todo.md` rather
+than fixed here: `skills/c64-development/references/hardware.md`'s sprite-Y row
+formula measures one high, and correcting it touches `demos/invaders`, which
+reached the opposite conclusion deliberately.

demos/amiga_ball/README.md

@@ -93,10 +93,15 @@ reports pass against scores generated from the impact schedule rather than
 fitted to the recording, and a written audit marking all 28 criteria with the
 command and output that settled each one.
 
-The per-frame job costs **7 raster lines** of a 263-line frame. It is armed at
-line 10 and finishes inside the top border, before the VIC draws a single pixel
-of the display — so there is no tearing to look for, and the margin is not an
-estimate but a `c64 profile tick` reading of 481 cycles.
+The per-frame job costs **7 raster lines** of a 263-line frame on an ordinary
+frame and **14** on the one frame in 64 that gates a new impact — a high-water
+mark the program keeps itself, because a sampler steps over the expensive
+frames and reports a comfortable number that means nothing. It is armed at line
+10 and finishes inside the top border, before the VIC draws a single pixel of
+the display, so there is no tearing to look for. `c64 profile tick` prices the
+same job independently at 450-480 cycles mean, which it can only do because the
+handler is a thin wrapper around an `rts` subroutine — `profile` masks
+interrupts, so it cannot price a handler in situ.
 
 ## The bits worth reading
 

demos/amiga_ball/SPEC.md

@@ -886,6 +886,16 @@ the score and its evidence is the spectrogram — where the noise transient and
 the filter's downward sweep are visible and the piano roll cannot show them.
 Both reports must read `verdict: PASS`.
 
+**The capture window does not open at a fixed offset.** Measured across eleven
+captures: the WAV varies by ±2 video frames in length (165,292 / 166,892 /
+168,492 bytes) and the impact lands anywhere from **353 ms to 387 ms** into the
+file. The register schedule is *not* what moves — `sid-log.jsonl` is
+md5-identical across five captures of the same build — so it is the window
+opening, not the program drifting. Anything measured against a fixed grid in
+the WAV is therefore partly measuring where the window opened, which is exactly
+what made iteration 1's centroid figures irreproducible. **Anchor every
+time-domain measurement on the impact itself**, which is stable to ±5 ms.
+
 The maintainer's listen of `capture.wav` is the final gate on whether it sounds
 like a boing.
 
@@ -949,17 +959,23 @@ machine, never from reading the source.
 21. `evidence/audio/floor/report.md` and `evidence/audio/wall/report.md` both
     read `verdict: PASS` against a score written from the impact schedule.
 22. The spectrogram of each capture shows a broadband transient at the impact
-    frame whose energy centroid falls by at least a factor of three over the
-    following **~100 ms** — the filter sweep, which the piano roll cannot show —
-    and the wall's centroid is higher than the floor's at every point in that
-    window.
-
-    (An earlier draft said ~250 ms, which §8's own instrument cannot deliver:
-    the cutoff ramps for 16 frames = 267 ms, but `$D40C` = `$04` decays voice 2
-    in 114 ms, so the burst is gone before the sweep ends. The second clause is
-    free — measured 4,484 vs 4,204 Hz at onset and 1,881 vs 1,128 Hz at +100 ms
-    — and it makes the criterion test §8's *two surfaces* claim rather than only
-    its one-gesture claim.)
+    frame — a full-height column, arriving in one frame — and **the transient's
+    own band (2-10 kHz, above voice 1's audible harmonics) takes at least
+    150 ms to fall 30 dB**, so the noise burst outlives more than half of the
+    267 ms cutoff sweep and there is something left to hear the sweep in. The
+    wall's spectral centroid is higher than the floor's at every 10 ms step of
+    the first 200 ms.
+
+    (Two earlier drafts asked for the *energy centroid* to fall — by a factor of
+    three over ~250 ms, then over ~100 ms. Measured over eleven captures, that
+    fall is **the noise burst ending, not the filter closing**: its timing
+    tracks the 2-10 kHz band's 30 dB point exactly, and a centroid computed with
+    voice 1's fundamental excluded moves by 1.12× across the whole window. So
+    the old wording rewarded a *shorter* transient — `$D40C` = `$00` would have
+    scored best of all — which is the opposite of what §8 is trying to buy. A
+    criterion that improves as the demo gets worse is not a criterion.
+    `AUDIT.md` iteration 2 has the captures. The replacement above can fail and
+    does: the same measurement on the `$D40C` = `$04` build reads 95 ms.)
 
 **Budget**