Oliver on the first shell build: *"I don't hear any of the games sounds.. daniel the umpire is
synthesized which is horrible and no music is playing on start"* — two independent causes.
- **Opaque fetch.** Capacitor serves the bundle through a custom WKURLSchemeHandler, and `fetch`
reports those responses as `status: 0, ok: false` with the whole body present and decoding
perfectly. All three loads in `audio.js` guarded with `if (!r.ok) throw`, so inside the app every
recorded sound failed at the door: the umpire fell through to the synthesized `say()`, no bed
started, no sample played. The synthesized sfx kept working, which made it read as a mixing bug
rather than a loader one. Fixed once, in the shared `bytes()` helper every audio fetch now routes
through — a status of 0 is the ABSENCE of an HTTP status, so trust it only when there is one and
a real 404 still falls back. Any future `!r.ok` on a bundle-relative fetch is this same bug.
- **The ring switch.** WKWebView's WebAudio runs in the SoloAmbient session, which the hardware
silent switch mutes outright — a silent game on a phone that lives on silent, with nothing on
screen to explain it. AppDelegate now sets AVAudioSession `.playback`, the games category.
Confirmed in the shell: the bed reaches PLAYING, the context is `running` with no gesture at all
(autoplay really is allowed in WKWebView, as LATER predicted), and `say()` never fires. On the web,
a genuinely missing file still falls back — checked both.
And the swipe flips: **right goes back**, the way iOS's own pop gesture goes (*"that's the iOS
way"*). Left now does nothing. Noted in PROJECT.md that on a platform gesture the platform's
convention beats the literal wording of the brief, and that is worth saying up front.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A Capacitor iOS project (`npm run ios`), and the first three bullets of LATER item 2 built. The web
build is untouched and stays primary: every line here degrades to a no-op in a browser.
- HAPTICS ride the sound map, as the plan said they would: six call sites, all of them places an
sfx already fires — your racket contact, the serve strike, `killZombie` (the one funnel every
weapon routes through), `explode`, `hurtPlayer`, `killPlayer`, the weapon swap. The restraint
rule lives in `haptics.js`, not at the call sites: a minimum gap per kind, so a bomb that kills
twelve bodies is felt twice and the machine gun is a metronome rather than one long rumble.
Pinned by a test. Untested on hardware — the simulator has no Taptic Engine.
- The plugin is reached through the INJECTED bridge, `Capacitor.nativePromise('Haptics', ...)`.
`Capacitor.Plugins.Haptics` reads like the zero-dependency answer and is silently empty without
@capacitor/core's registerPlugin — it would have fallen through to the web path inside the app
and looked exactly like a phone with haptics switched off. No package JS ships.
- PORTRAIT is enforced where it can be: one orientation in Info.plist, so the app never rotates.
That retires landscape as a surface rather than fixing it (OPEN 6). The web half is best-effort —
`orientation` in both manifests and one guarded `screen.orientation.lock()`.
- SWIPE LEFT GOES BACK. The gesture finds the screen's one visible chevron and clicks it, so the
shape of the wizard stays in the handlers that already own it. Not passive on purpose: a swipe
inside a poster card also fires that card's click, and preventDefault is what stops the screen
answering forward in the frame the chevron walked it back.
- The app has no server behind it, so `config.API` and one branch in `net.js` point the shell at the
live host by name. Consequence: the app always talks to PROD, never staging.
`node test.js` green (105) over four seeds; builds, launches and reaches the native plugin on the
iOS 26.5 simulator; swipe-back walks the whole wizard in the browser pane.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A tour is which body you play in, so it belongs in the look message next to the
kit, the skin and the name. Without it both phones drew the friend in whatever
tour THIS phone had picked, and a men's/women's pairing showed each player
someone their friend was not using.
`oTour()` is the one accessor — yours everywhere, theirs in netplay — and it
feeds the three things that must agree about a body: the sprite suffix, how the
art takes a colour (her whites wash, his blue dyes) and the palm the idle ball
bounces under. An older build sends no tour and falls back to ours, which is
what it did before.
The palm knobs were re-checked against the new front sheets rather than assumed
— the poses were copied, so the ball still lands under the drawn hand.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
His colour landed on the shirt while yours lands on the shorts, because he was
drawn on the roster opponent's red sheets and dyed on an arc of his own. Red is
the one hue that art cannot hang a kit on: the men's shorts are neutral grey at
saturation 0.05, so no rotation ever reached them, and the women's red top sits
at her skin's and her hair's hue — a dark pick was painting her arms and her
ponytail, live, on the shipped build.
So he gets his own sheets. `assets/_frontroll.py` is `_famroll`'s trick pointed
at the tennis poses — image 1 is the ten `o_*` frames composed from disk, image 2
is the player from behind — and two calls drew all twenty. He rides the PLAYER's
arcs and the tour's own mode now: his shorts dye, her whites wash. `TINT.oKit` is
gone. The roster CPU keeps the red art; it is not a person.
The setup screen shows that front too — picking a face off the back of a head was
only ever what the art allowed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
'No that's way too difficult now - I am dying constantly of out. revert
it.' aimAnchor is gone — the lane is absolute again, every out invariant
back exactly as tuned. What survives is the answer to his follow-up: the
ball sometimes flew a LITTLE RIGHT out of a left flick, because a
mistimed left flick's target sits only slightly left of CENTRE — which
is right of a player standing in the left service box. The absolute
lane's blind spot, stated the other way round.
The aimed target is now clamped to the player's own x on the flicked
side, bounded at flickWidth so a body parked in the alley clamps to the
corner rather than conjuring an out. From the left box a left flick goes
at least down your own line, from centre it can no longer cross centre,
a perfect corner flick and the crosscourt sweep are untouched, and the
outward scatter tail keeps the whole tuned out-risk. Verified in his
scene: ten rushed weak left flicks land -2.0..-2.7 where the same case
used to target -0.7..-1.3 (a visibly rightward flight); the clean full
flick still lands -3.92.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Third round of the flick report, and this was the core all along: 'when
the opponent and the player are both standing in the left service box it
is virtually impossible to flick the ball into the left out. no matter
how hard/good I flick.' The lane was absolute — a full flick asked for
flickWidth off the court's CENTRE — so the gesture saturated near a
sideline: from x -2 a perfect left flick moved the ball under 2 m and
could never pass the line, and the better the timing the firmer the cap,
because scatter was the only thing that could cross it. 'Left is
impossible, right works' was never left-vs-right: it was
toward-the-near-sideline vs across-the-court.
TUNE.aimAnchor: the anchor slides from the court centre to the player's
own x as the swipe commits (ch.x * |sway| * aimAnchor), so a flat tap is
exactly what it always was and a full flick is a full sweep from where
you stand — including past the line. Out-by-aim is a choice now, per
Oliver's explicit direction; the placement ring already turns red for
it. 0 restores the absolute lane, blendable live in the panel. Player
rally lanes only; the opponent's lanes and the centre-stated pOut
invariants are untouched. Verified in his exact scene: full left flick
from the left box lands -5 (the alley), half swipe -3.07, full right
+1.93.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Second round on the service-line flick report. The crossing-fired swing
alone did not move the feel, and Oliver's sharpened observation — the
flick only lands left when the opponent already stands left — pointed at
a magnet that does not exist: measured with the body teleported at
contact, the aim never tracks it. The real mechanism: at rushed quality
the symmetric scatter (aimError x (1-q)) outweighs the quality-shrunk
aim, so a mistimed left flick landed anywhere in a ~4 m blob around
'slightly left' — usually centre — and the eye anchors that on the
opponent. Its position was a proxy for preparation time all along.
TUNE.aimTrust: an outward scatter roll is untouched — the committed
swipe's chance of sailing keeps its exact tuned shape, flickReach stays
the only out dial — but an inward roll keeps only 0.3 of itself. Bad
timing costs width, never direction. Player lane shots only; the
opponent's symmetric scatter is its width mechanic. Verified at
worst-case q~0.08: 14/14 rushed left flicks landed left, 9/9 right
flicks right, where before it was a coin toss. The suite pins the
direction guarantee and the aimTrust/aimError/flickReach relationship.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Oliver's service-line report: flicking the incoming ball LEFT was near
impossible mid-court at the service line while RIGHT worked, easing toward
the baseline. The lane math was symmetric to the centimetre — the defect
was latency. The swing waited for the RELEASE, so the whole gesture ran as
dead time on top of the windup: ~120 ms of lateness drops contact quality
1.0 -> 0.08, where flickReach halves the width and aimError drowns the
rest. A left swipe is a right thumb's slowest direction, the service line
leaves ~700-870 ms of budget (measured), and any swipe over FLICK_MS was
silently dead. Left + service line was the slowest gesture meeting the
shortest budget.
Now, in rally modes: crossing flickMin fires the swing (once per gesture,
rally phase only) and the vector stays LIVE — trySwing reads G.swipe at
contact, with the released flickAim carrying the aim only when the ball
arrives after the finger lifts. The FLICK_MS deadline goes with it where
the crossing can fire; it survives where it cannot (gore free-aim, the
serve strike). Serve gesture, duel, gore and keys untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The header block in tools/scenes.json is now the whole schema — every key with
its range and default, what each of the EIGHT weapons does (fire, laser and gun
were missing from both the docs and the validator), the five beats, the crowd
roster diversity walks in order, the bosses by chapter, and the traps worth
knowing: a serve tin is spent in about four seconds of autopilot fire, a burst
tin owns its own window, a horde locks the rack, a boss cannot start inside a
match, and `movement` fights the match's own auto-run legs.
Two scenes added from the campaign's own wish list: `epidemic` (the fire ball
lighting a whole wave off the body it hit) and `champion` (chapter ten, the
night court). Both verified on film.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
tools/scenes.json says WHAT to shoot; tools/clips.js only knows how. Every
knob Oliver asked for is a key: seconds, setting (gore/zombie), beat, court,
boss, weapons, movement, density, diversity, gore, games, screen — documented
in the file's own "_" block, validated in capture.py so a typo names itself
instead of coming back as fifteen seconds of empty court. Nine scenes ship in
the default list, across all four courts.
The frame is a PHONE now (393x852 -> 1080x2340), not 9:16: the portrait branch
derives its logical height from the canvas box, so a 9:16 take was never the
framing a player sees. The player has legs — held arrow keys through the real
axisX, so horde and flamer are no longer a statue with a flamethrower. The
crowd tops itself up where the game's own spawner puts it, diversity walks the
roster, and `gore` paints in the run's history a fifteen-second clip has no
time to earn. The pilot is picked by what the court is doing rather than by
config: tennis mode taps to swing, splatter draws the slingshot.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`python3 tools/capture.py` fills promo/clips/ with 1080x1920 h264/aac takes of
the moments the campaign plan (item 4) already named — transition, horde,
flamer, mopboy, death on the gore side, boss and match on the zombie-tennis
side. capture.py serves the repo with tools/clips.js injected, opens a
throwaway-profile Chrome and converts each POSTed webm with ffmpeg; clips.js
stages every scene through the real doors on window.T and drives it with the
reachability soak's own mediocre autopilot. Nothing is faked for the camera.
The game gains four names on window.T (start, intoSplatter, startHorde,
sendBoy) and nothing else: the frame is pinned by the canvas box rather than
the window, RES 4 comes from lowPower 0 plus device-scale-factor 3, and the
audio is tapped by making BaseAudioContext's `destination` a gain node — so
audio.js needs no master bus for a tool that ships to nobody. tools/ and
promo/clips/ are ignored by git and docker alike.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Hidden behind TUNE.aimCloud rather than deleted: the reading it gives — how
much width the swipe is buying, before the ball leaves — is the only teacher
the gesture has, and the panel can flip it back live for the next look at it.
And "a little harder to hit the out" is a number: flickReach 0.62 -> 0.52,
which halves the worst-case rate from about 15% to about 7%, with the test
tightened so it cannot drift back. The shape is untouched — a perfect flick
still never sails, a badly timed full swipe still can — but a mistimed one now
lands nearer the middle than the line, which is the only half of that trade a
player who cannot see his own timing can do anything about.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The live flick readout was a ring on a ray from the player's feet — a HUD
element laid across the court, and read as one: "pretty bare bones... the line
must go away, it is distracting". It is a soft cloud on the ground now, in the
register the dust and the blood already use: it sits where the ball is being
asked to land, grows with the length of the swipe, and warms from pale lime
through gold to red as that ask approaches the width no timing can keep in. The
threshold stopped being a flip, so too hard is something you watch coming.
Three things it took to make it read, kept in the comments so the next pass
does not re-learn them: it is sized in METRES off the projection, or it is an
eight-pixel smudge at the far corner where it actually lives; it is a stain and
not an additive glow, which turned the red end yellow over green grass; and the
ramp starts pale rather than green, because green on green showed nothing at
all across the whole safe half of the range.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
On touch, a drag in zombie tennis and classic still walked the player — and
that was not a harmless second option. A manual axis suppresses the assist, so
holding a drag traded a full-speed run to the intercept point for a 0.35x
shuffle aimed wherever the thumb went instead of where the ball will be.
Walking was strictly worse than standing still, which is most of "i tend to not
be able to reach the balls".
touchAxis is no longer written under rallyFeel. The drag keeps its one rally
meaning — released quickly it is the aimed swing — keys stay a deliberate
desktop override, and gore keeps its drag-walk, where the assist recovers depth
only and holding your own lane is the entire point.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The rally flick was read as three lanes: sideways past 0.38 of its own length
took the whole corner, a hair under it took dead centre, and nothing lived in
between — "sometimes applied very strong, sometimes it barely does anything",
with no way to tell which one you had asked for. swayOf reads the swipe's LENGTH
as a continuous strength, exactly as the serve already did, and doHit spends
that instead of the lane vote. Keys keep the full lane; the opponent's aim has
no length and is untouched.
That is also what hands the OUT back: going out was never a choice, because
every aimed shot took the full width and then the scatter on top. Below about
0.7 of a full swipe — the line minus aimError, over the width timing can earn —
the worst mistiming in the game cannot sail it now, and above it the old risk
curve is exactly as tested.
And the aim is drawn while the thumb is still moving: the placement ring is a
receipt that arrives with the ball already gone, so the same ring is shown live
with a ray from the player's feet, cream while the ask is safe and hot once the
sway crosses that derived threshold. The receipt goes red when the shot really
is out, so the two rings together are the lesson.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
HANDOFF-MATCH.md was always temporary — written when a session ran out of
context mid-slice, and carrying its own instruction to fold into PROJECT.md and
be deleted once the slice was done. The slice is done, so it is.
What crossed over is the contract, not the diary: what the mode is, the
blood-only scoring and the misreading that forced the blood readout, the four
ladder dials, the card of ten, the wound dials, the fall's frame contract, the
flick arithmetic and its lesson, the umpire and the boy, the opponent's serve
games, classic as its own mode, the record shape, the traps, and the do-nots
(now including the scratched body-shot thrower). The reasoning behind the
numbers already lived in the drill/match session logs and was not copied twice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The slid-serve test asked for 6 of 10 sliced serves returned. Measured over 200
trials a seed across eight seeds, the receiver's real rate is 0.85..0.92 — so
the assertion was sound and the SAMPLE was too small: at p ~= 0.88, 10 trials
land at or below 5 with probability ~0.004, one seed in 230, and SEED=80 was it.
n is now 40 at the same 60% bar, which puts the threshold five standard
deviations below the true rate while still failing instantly if the receiver
regresses toward the 0-in-24 the test was born from. It is free: a returned
serve breaks the frame loop early, and the suite still runs in ~0.11 s.
Green on seeds 1..400.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The drill spawner measured its best-of-9 lane spacing only against bodies still
in the back half (drillNearZ 9 m). On a court where lanes are HELD for life and
speed is stamped at spawn, a body 12 m down still owns its lane — handing it to
a faster arrival books an overtake for ten seconds later, mid-court, as two
sprites walking through each other. That is what the seed-sensitive test failure
(SEED=1234) was reporting: the assertion was right, the stream was not.
Spacing now measures against every live drill body; drillNearZ is deleted (no
other reader). Still best-of-9, no retry loop, so the stream cannot stall.
60 seeds x 150 s of drill: before, 7 seeds overlapped with narrowest clearance
0.00 m; after, zero overlap frames, narrowest 0.46 m (the test asks > 0.3).
Suite green on 20260811/1234/777/99/424242/8 and on 299 of seeds 1..300; the
straggler (SEED=80, the slid-serve return rate) is unrelated and logged as open.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Six sessions of asking an image model for a row of running figures, and it
never once drew a gait: three phases repeated with the same leg leading,
"ropeskipping, not running". A frame-by-frame LEFT/RIGHT skeleton spec, a
closing checklist, a loop clause, six figures instead of eight — all lost.
One 5-second image-to-video clip has no such problem, because it moves a
body instead of drawing poses. ~$0.42 a character through the key already in
.env, pose pinned at BOTH ends (pinned only at the front, she rotated into
profile halfway), stride period MEASURED off the frames rather than sampled
by eye — hers repeats in 6 clip frames, his in 23 — and the phases cut and
palette-snapped through _stride.cut. assets/_vidroll.py films and cuts it;
her first frame came from one still roll that took his approved pose as a
picture instead of a paragraph.
Wired: runN probes the sheet through SK() so the women's tour counts HER
frames, the player is no longer clamped to a two-frame pair, runHz 2.4 ->
1.9 ("this looks like roadrunner"), and the bench grew two PLAYER stops —
a cycle only ever seen mid-rally is a cycle nobody has watched.
Also in: the difficulty retune (medium is a real step above the shipped CPU
and is where `smart` first appears; hard stepped back off superhuman; easy
closed its gap) with MEDIUM's dial-for-dial pin to TUNE removed from the
test, and Oliver's own transparency fixes on the male frames.
SPRV 138 — the frames were right on disk and wrong on screen for an hour
because assets are cached for a year unless the ?v= moves. main.js v269.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A URL only opens the installed app on Android; on an iPhone it opened a second
copy of the game in Safari, with its own storage and none of the name picked in
the app. So the share sheet now sends "JOIN ME IN SPLATTER TENNIS — THE CODE IS
ABCD" and nothing else, and the button says SEND THE CODE.
The #net-ABCD door stays — it is exactly what a deep link will carry the day
this is a real app with universal links. Nothing sends it in the meantime.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Multiplayer setup moves out of a canvas modal and a prompt() dialog into one
menu screen: your name, the two ways to host, and the friend's code. The socket
opens from the menu and the game stays parked behind the overlay until the room
fills — netHooks.paired takes the court one line before netStart writes to it —
so the room code can carry a real SHARE button and a wrong code is a typo you
fix rather than a CONNECTION LOST card over a match that never began.
The name is the world board's name (stats.saveName), crosses on the existing
look message, and the scoreboard finally names the person instead of CPU.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Oliver's hunch was right: the tap unlocks audio and starts a 768 KB fetch, but
the source starts seconds later inside a promise with no gesture near it. iOS
suspends an idle context in that gap, and a source started into a suspended
context is silent for good — nothing on the splash ever asks for audio again.
Resume once more when the buffer lands, which the first tap has already
authorised, and again when the app comes back to the foreground.
Intermittent because it only bites when the decode outlasts the browser's
patience: a phone, a cold cache, a slow radio.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every run starts the ambience, and the ambience was un-ducking both court beds
at once — so once you had played the gore mode, its music followed you into
classic tennis for the rest of the session. Leaving the court is one event;
arriving is not. Each bed brings itself back and nothing else.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Splatter gore no longer makes you win a game of tennis first — the mode you
pick is the mode you get, with the man you came to play already turned. The
transition itself is untouched and still reachable; it just isn't a toll booth.
The menu theme could only ever be started, never brought back, so the first
run of a session silenced it for good while the court's ambience followed you
into the menu. Starting a bed now un-ducks it too.
Three world boards instead of one: kills, opponents felled, and games taken
off the CPU, each with its own name prompt and its own row per player.
...and the gore economy stops paying a barrel for every game won past game 16,
while the shamblers, the mix and the ball machines keep escalating past the
point they all used to stop.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Six new opponent families (head coach, stringer, strawberry seller, chair
umpire, spectator, roller man) through the famroll/zsheet route, placed by
hierarchy into an 18-rung ladder; the champion steps up behind them.
THE DRAW: picking the tournament now rolls its own bracket in front of you —
only the final was ever settled — and what lands is what walks on.
Netplay, off Oliver's cellular report: the mid-flight early-decide was missing
its !G.net gate, so the striker judged its own outbound flight and the two
scoreboards drifted a point apart into a court nobody could serve on. The
host's court now rides the room, its scoreboard rides the pretoss stream, and
a called ball rolls to its own halt instead of freezing where the call caught it.
Plus the night court, and somebody in the stand is taking pictures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two retest findings, one root: the death match borrowed the roster's
body. Gone — nothing spawns, no walk-on card, the living opponent is the
whole picture. His damage reads as blood (bleedO: an anchor per point
taken this game, love to game point, thin drops as ever), the game
point bursts him where he stands, and IT RISES AGAIN brings him back
clean. The heal beat became a card in netplay — a modal freezes this
world while the other phone's keeps moving.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three chips under VS A FRIEND — the host's pick rides the room and the
start message rebuilds both courts on pairing. In the death match your
opponent IS the zombie: their lost points tear its wound stages, their
won points draw your blood, and bleeding out crosses the wire as the
other phone's win. The same receiver-is-judge verdicts drive the match
machinery through resolveInterlude, so the fall, the mop beat and the
burst come along unmodified. New: gushProne — the game-winning body
bleeds out on the ground, from prone-height anchors, until it bursts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
VS A FRIEND on the mode screen (host a code, join a code), a ~110-line
RFC6455 relay in server.py that holds no game state, and src/net.js: each
phone plays itself near-side and the opponent's court is this one rotated
half a turn about the net. The phone a ball flies toward makes every call
on it; the striker's copy is picture. Six tests pin the rule from both
seats. No reconnect yet — a dropped radio is a card and a reload.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The female run took the whole evening because the model would not draw her
RUNNING: roll after roll came back with the same leg forward in every frame
— amplitude wiggle, no gait ("she never changes her leg... floating while
strobing"). The set that ships is the FACE TRANSPLANT: her old cycle's poses
(which always alternated properly — swaps at positions 2 and 6, 78%/74%,
where a runner's legs actually cross) redrawn with only the friendly face
and open beds changed. The model returned that roll on a painted CHECKERBOARD
— it has no concept of transparency, the word alone summons the pattern —
and the figures were custom-keyed off it: light checker blanket, dark
checker border-flood only (the dark cell is nearly her skirt colour; the
first blanket pass shredded the pleats), verified whole over a light ground.
Oliver gated every strip by eye. The alternatives lost on their own terms:
hex-only prompts obeyed the background but drew her dainty (max 44%, no
swap); the explicit leg-swap clause got alternation but clustered it in the
back half (step-step-skip). His pick: the transplant.
The lesson is in PROJECT.md now, next to the bed flag: NEVER say
"transparent" to the model — name one flat hex — and never freehand
background prose in one-off scripts; import _regen.py's CANVAS block, which
has said it correctly all along. SPRV 123, main.js v246, all 14 run frames
verified 200.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
One paragraph in the asset pipeline so the next fresh session does not
rediscover it: prompts ask for magenta beds ON PURPOSE, _fine.key's grey
repaint defeats them unless bed=True is passed, and the solid white paddles
were exactly that interaction.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Oliver's concern, after the un-chunking: a fresh session walked straight into
_regen.py and diagnosed from there, twice — first the legacy prompts, then the
model's rolls — while the real bug sat in the cutter, provable by putting any
fam_ sheet next to the sprite cut from it. The family-pipeline section already
demoted _regen.py, but 270 lines below the "Asset pipeline" heading, and the
session memory that routed the last investigation pointed at _regen by name.
So the routing is now the FIRST thing under "## Asset pipeline": a
which-tool-for-which-job table that names _famroll -> _zsheet for families,
marks _regen.py as the single-sheet remaster tool it still is (bosses, players,
crowd, weapons — those jobs keep it alive, so it stays), and states the
diagnostic rule that would have saved two days: when installed art looks worse
than the _refcheck sheet it came from, the suspect is the CUT, and a re-cut is
free. _regen.py's own header says the same, for whoever opens it first anyway.
The "Only ink votes" lesson joins the family-pipeline section with the numbers:
0.0% of ink blocks uniform at b=11 against 99.2% of empty ones is how a
detector proven on single-row sheets quietly broke on 9-row geometry — the
threshold never changed, the data under it did.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The fourth mode: four rounds up the dead hierarchy (CUP_ROUNDS), each round
drawing a rung from its band, the champion always the final. It rides the
match machinery whole — cupDraw feeds the drawn rung through d.opp so every
ladder dial scales untouched; the body's pool spans the round (need-scaled
pool and divisor, so wounds carry between games); cupGameWon holds the same
body mid-round, kills it on the round, and past the final cupWon banks a
NOT-dead match run under CHAMPION OF THE DEAD CLUB. Blood carries; round one
is a single game; deathmatch stays as the arcade run. Full-run state test.
Sidequests: the ghost trail is pace-qualified (TUNE.trailSpeed) so both
tennis modes streak — pow>0.8 was only ever reached by the gore slingshot;
opponent sprites draw ~18% brighter via a load-time baked copy
(render.brighten, TUNE.oppBright) with no PNG touched; the cup scoreboard
line names the round instead of lying with deathmatch arithmetic; the rack
and the apron demo no longer draw under a trophy screen.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Oliver's four picks, each rolled off its own settled BOSS frame through the
family pipeline (rows in _famroll.py) — the rung and the chapter boss are the
same person. matchFoe decoupled from bossArt's cycling on his call: the match
ladder is 12 with the champion capping rung 12; chapters keep their ten.
The champion refused to bulk through four rolls (words lost, then a new
pair-reference in char_ref only half-won), so his mass is stature: OPPONENTS
rows carry an optional world height h read by zombieH — champion 2.1 m, the
Bulk 2.0 m — and the serve aims at the body actually drawn. HEAD_CROP measured
for all four; ramp test extended to 12 rungs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The old run frames were generated one at a time and it showed: 91/110/92/98 px
of body-width drift, duplicated poses drawn differently, no racket. Each
character (player m/f 6 frames, opponent m/f 8) now comes from ONE 4K call
anchored on a composed identity reference of its settled frames — the broken
run art stayed out of the reference on purpose. Stride-gated, cut with _fine,
judged by eye per roll. drawChar probes run8/run6/run4/run2.
Side effect: the male opponent's old run art faced LEFT against the
art-faces-RIGHT convention — he had been moonwalking right since install.
Fixed by drawing the new rows to the convention.
Pipeline lessons recorded in PROJECT.md (the model negotiates the count; a
back view has no direction; the racket clause goes early and loud).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three mechanical causes of 'no serve agency', measured then fixed:
- the 0.12 m perfect snap was ~0.3 s of hang time (now TUNE.apexSnap 0.05,
and the gold band reads the same knob);
- the fixed mid-box target depth capped every good serve at ~36 m/s in the
loft loop — player serve depth now rides quality, flush aims 0.3 m inside
the service line (ladder: 49 / 52 hard / 36 mistimed / 30 bad);
- the strike flick was splatter-only — in tennis and the duel it now sways
placement (horizontal) and buys pace (vertical, serveHard + serveHardSpin,
the extra dive IS the clearance), paid for in scatter against the real
fault machinery. Tap stays the safe second serve.
Phone verdict pending — see the PROJECT.md session log.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The flattened body curve stays (farK 0.25 — 'really good, helped a ton'); the
broadcast lean was 'distracting' and its knobs sit at 0 like the aimCam pair.
The plumbing stays for the T panel. PROJECT.md: farK is now a settled feel
decision.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Oliver: the far-side opponent is barely visible in portrait, and the lens was
already at its measured ceiling. TUNE.farK (0.25) grows far bodies ~20% while
the court keeps the true projection — the 16-bit tennis trick, anchored so the
near player is untouched. TUNE.farCamY/farCamZ add a broadcast lean while the
ball is on the far half, pivoted about the player's feet (fOff) so the pull
room the apron math promised never moves. Both live in the T panel; both
awaiting the phone verdict — see the PROJECT.md session log.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Oliver walked the sprite bench and reported the same two faults on every family:
the healthy walk loses its head, and the wounded walk's feet do not move. Both
were in the ONE FILE every zombie tennis character is drawn from — and the family
pipeline had copied it faithfully, which is the pipeline working.
WHERE IT CAME FROM. That sheet was not composed from the sprites. It was a
GENERATED RE-DRAW of the club pro, made in nine separate 4K calls, one per row
band, then stacked (gen4k.py, in a session scratchpad, untracked and now gone).
Nine calls returned nine scales, and call 1 came back with the shuffle frame
decapitated. Measured against the sprite it was drawn from, that one figure
scaled 2.341 wide and 1.996 tall where all thirty others were proportional to
three decimals: 33 px missing off the top, width untouched. The local input to
that call was exact at 2.000x. The sheet lost the head, not the sprite.
THE FIX IS TO STOP PUTTING THE REFERENCE THROUGH THE MODEL. _famref.py composes
it from the 31 z_r* frames on disk — one scale, applied once, NEAREST — and costs
no tokens. That kills the head and the five scales together, and it explains two
things in _zsheet.py that read like fudges: the fall1 re-anchor and the walk-frame
levelling exist only to undo per-band scale drift. The levelling is also what
turned a headless z_sh2 into a headless AND 13% oversized one, by reading the
missing head as a height error and scaling up to compensate.
THE GATE WAS MISSING ON THE INPUT. _zsheet.py checks what comes BACK from the
model — band count, figure count, stacked bodies, edges, palette, heights — and
nothing had ever checked the sheet handed TO it. _famref.check() re-reads the file
it just wrote with _zsheet's own detection and refuses one where any figure's
width and height scales disagree by 2%, or where a stride pair is frozen.
THE THIRD DEFECT NEEDED NEW ART. z_rw1b had the same legs as z_rw1 (15.1% against
the healthy pair's 51.1%), so the club pro has slid through both wound stages
since the day he shipped, and every family inherited it. _stride.py draws the
missing twin, and what made it work after words had failed for as long as they
were tried is SHOWING the edit: image 2 is the healthy pair side by side, the same
character performing exactly the change being asked for. The headless twin then
took the ARM-OFF pair as its example — a wounded body is a closer picture of the
edit, the same chaining the wound ladder already uses. Its gate measures three
things and prints all three: legs moved, body above the hips did not, racket kept
its bed. The thresholds are CALIBRATED on pairs that already ship, not chosen —
my first guesses rejected z_e1/z_e2 and z_g1/z_g2, both accepted art. And a twin
FILLS the racket bed instead of keying it out: the first rolls came back with 959
px of enclosed transparency against their base's 0, which would have flickered the
racket face empty/filled once per step.
AFTER, on the installed sprites — leg difference within each stride pair, where
the wounded pairs used to sit between 4.8% and 19.7%:
r 51.1 / 68.1 / 36.7 ru 55.9 / 59.1 / 38.7
br 40.7 / 42.0 / 26.8 gk 41.1 / 44.6 / 30.6
sh 59.4 / 71.4 / 43.9 lj 38.4 / 45.5 / 27.4
bb 56.5 / 67.1 / 43.2 st 51.6 / 66.1 / 35.7
All seven families re-rolled off the rebuilt reference, every play row with its
head, and all 24 family/wound-stage states alternating a real pair — checked by
instrumenting drawImage on the bench, not by eye. br and lj sit lowest on the
headless pair and are worth a re-roll if they ever read as stiff; lj needed one,
its first roll came back frozen on all three pairs with the headless row 31 px
short.
AND THE REFERENCE IS TRACKED NOW. It lived in gitignored _refcheck/, so the thing
the entire pipeline rests on was unreproducible for as long as it existed. It is
assets/_pose_ref_family.png, next to _pose_ref_serve.png, and the scripts that
build and consume it are in the repo instead of a scratchpad.
Also in here: the SHAMBLE BOB, which was a 1 px lift on a wall clock applied to
every body every frame whether it was moving or not — a held pose bobbed forever
and a corpse bobbed while it lay on the ground. It only happens while walking now,
and it steps off ground covered, which is the same rule the strides follow.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
THE FLICK. "when returning a ball with a spin it seems too sensitive, 8/10
times the ball lands out." Arithmetic, not feel. The singles line is 4.115 m;
an aimed shot targeted a flat ax*3.3 and THEN added err() on top — up to
+-2.4 m at quality 0 — so a flicked return landed as wide as 5.7 m, and the
only clamp in the path sat at +-5, outside the court. An UNAIMED return
targets 0 + the same scatter and is therefore always in: aiming was the thing
losing the point, which is backwards for the mode's one act of skill.
TUNE.flickReach scales the aim width by the same quality the scatter shrinks
with, so the two can never sum past the line — worst case 1.49+2.4 = 3.89, and
a perfectly timed flick still gets the full 3.3 with nothing left to scatter.
The test asserts that sum across the whole quality range rather than sampling
rallies. Not named `aimReach`: that key already exists further down TUNE (the
slingshot's 22 m aim ray) and the duplicate silently won until a test caught it.
The placement ring carries `q` now and is the flick's receipt — tight and hot
when your timing earned the full width, wide and dim when you barely got a
racket on it. Without it the rule is invisible: the shot just goes less wide.
THE FALL (handoff item 4). A point that takes a limb knocks the body down and
it drags itself up short of it. The three collapse frames wear the stage it
STILL WAS and the rise wears the stage it has BECOME — that order is the whole
beat, and it is what turns a popping wound stage into something the point did.
Own clock beside the horde's warning, for the reason this file keeps relearning
(a beat that moves the phase under a latch strands it); the 'over' beat it sets
outlives the fall so it cannot expire into setupPoint() mid-collapse. Only
fires when the stage actually moved — a quarter off a full pool is 0.75, so
point one has no limb to explain and a second of collapse for nothing is a beat
the player waits out.
THE OPPONENT SERVES (handoff item 7), and it was mostly switched off rather
than missing. setupPoint already placed a non-player server behind its own
baseline and armed aiServeT; update() already tossed and struck for O. Both
were gated on `G.mode === 'tennis'` and the match runs in splatter under a
permanent interlude — one condition. Genuinely new: service alternates at each
game boundary (the match calls advancePoint directly and the flip lives in
tennis's pointTo, which it bypasses — the same shape as the G.serveNum bug),
the zombie takes O's mark so picture and ballistics agree on where it stands,
and the serve poses are named onto it because O is invisible here.
racketStage() moves to config: drawZombie and the fall must not answer the
same question twice, which is how the boss's skins and its blood drifted apart.
Docs: PROJECT.md gains THE ZOMBIE FAMILY PIPELINE — the measured API contract
(4K is a parameter and the default is 1K; n is 1 so "return both" is not a
thing; aspect is an enum so pad to it), why magenta stays despite green being
the documented key, and the two things _zsheet.py must do globally.
Verified in the browser: brute in the match, z_brfall2 -> z_brw1rise on a
wounding point, and bounce -> toss -> serve off the opponent's own baseline.
58 tests green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A day went into sprite consistency and most of it was spent solving the problem
in the wrong place. Written down so the next session starts from the answer
instead of rediscovering it.
The boundary everything else follows:
the generator, GIVEN REFERENCE IMAGES -> identity, pose, composition
the pipeline, DETERMINISTICALLY -> exact colour
the gates -> layout accidents only
With the proof, because it is the part that is hard to believe until measured:
the brute's serve sheet came back with his build, his garments and his DARK
TONES pixel-identical to the reference and his big pale shirt two shades
brighter. The model was looking straight at the character. Every call
re-renders and re-rendering re-lights. No prompt fixes that.
And the rule underneath all of it — A REFERENCE BEATS A DESCRIPTION, at every
level. The racket zombie's sheet was made from the character as an IMAGE plus
the poses as IMAGES and landed first time; the brute's three sheets each
re-derived the character from the same paragraph and produced three different
brutes. Corollary worth having: every corrective sentence in a prompt is
compensating for a reference you did not supply. When a prompt grows a third
"do NOT" clause, add an image instead.
Also written down: correct-then-check, calibrate-before-you-gate (with the
numbers every threshold was calibrated on), the five gates and what each one
actually measures, six traps that each shipped something silently before they
were caught, and three things still open — including that `changed` passed a
headless brute that still had its head, because it knows something moved and
not what.
The brute is PARKED: out of ZFAMS, 31 frames on disk, nine batches in the JSON,
one line away from live. Its healthy eleven are good; its wound stages are not.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Decided with Oliver 2026-08-12. Records why tennis is the friendly case (the
sync unit is the shot, alternation already enforced by ball.lastHit, each
client authoritative for its own contacts), what the codebase already brings
(deterministic sim, fixed dt, beats.js ownership contract), the zero-dep
WebSocket path (~120 lines of stdlib RFC6455 in server.py, friend-code
rooms), the honest cost (reconnect-and-resume is the real work), and that
step one is the GHOST MATCH over plain HTTP — half the netcode falls out of
it as a byproduct. Plus the hook: score-is-the-body on both players, losing
is zombification.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Fourth recorded session, verdict "better": he beat opponent one — the first
game won in any recorded session — and reached opponent two, which took him
apart with gaps of 1.66-2.13m (60Hz, accurate) while he swung at closest
approach every time. The gradient is real; the step was a wall, because a
rung moves three dials at once. Per-rung increments softened: matchSpeedPer
.12->.08, matchWidePer .4->.25, matchOppPacePer .05->.035. Opponent one
untouched — the bases stay, only the slope flattens.
The contact chain is now verified end to end by recorded play: legs
(matchAssist) -> destination (assistK in intercept + the anyHit short-ball
chase) -> reach (matchReach). Each link was a separate defect found by a
separate recorded session.
49 tests, seeds clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Third recorded session. The previous fixes did their job and exposed the last
gap: he ran onto every ball (2.9-3.9m per point, was 0.2-0.5), the ball
reached him (died at z 22.5-24, was 18-20), he swung every single time — and
the closest approach clustered at a sampled 1.3-1.6m, centimetres outside the
1.3m post-bounce reach, point after point. The assist delivers you NEAR the
ball but not inside a phone-tuned contact radius, and that also explains "not
really better or I played worse": every point suddenly looked makeable and
still died.
matchReach 1.35 widens the contact test in the match only — the gore modes'
"tight contact radius: the tap must be honest" rule stands untouched. A rally
mode lives on contact feeling fair.
Recorded for future sessions: a 20Hz sampler overstates closest approach by
up to ~0.5m on a 20 m/s ball. The recorder samples at 60Hz now.
49 tests, 12 seeds clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Second recorded session: 5 wins and a deuce war to 5-6 — the mode is a
contest — but all seven losses shared one signature: the ball died in court
1.4-2m short of him while he ran 0.2-0.5m. Three defects:
The pace cut created dying rollers. Slow returns bounce and die at z 18-20,
outside intercept's near zone, so nothing chased them, and the rolled-dead
rule hands the point to the zombie. They also arrive at his feet below legal
contact height — the "ball arrives right at the player" feel. intercept keeps
an anyHit candidate now: the earliest reachable point on his half where the
ball is still at hitting height, so the assist runs forward onto the short
ball. Verified: he sprints z 23 -> 16.4 and meets it at gap ~0.
intercept computed reachability with autoAssist — 0 on any fine pointer — so
a desktop match's planner answered "home" every time. He had legs, his
destination was always the baseline. assistK() is shared by mover and
planner now.
And the vignette was a radial gradient in a portrait canvas: the circle
reaches the corners but hits the side edges at fraction zero — "red pulsing
from below" was the bottom edge being the only place it got close. Four
linear edge fades now, hugging the rectangle on any aspect.
48 tests, 12 seeds clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"can you watch me play in the browser pane and see why I loose". A passive
20Hz recorder sampled his real play: seven losses, and in every single one he
SWUNG (his timing was fine) while the ball died 2-3m to his side and his own
x never moved all game. He had no legs: autoAssist keys off pointer:coarse,
the pane reports a fine pointer, so auto-run was 0 — the exact trap the
verification recipe has warned about for weeks, bitten in live play.
matchAssist 1.15: the match brings its own legs on every device. It is a
tap-timing game; footwork is the game's job. Held keys still override. Pinned
by a test that plays with autoAssist 0 and asserts he runs.
The vignette curve is sqrt(damage): the match pool is 8, so hit one was a 10%
alpha nobody can see. sqrt front-loads it — first blood reads at ~0.3.
This also explains three rounds of disagreement between his reports and my
measurements: my autopilot always had legs (it set autoAssist by hand), his
pane play never did.
47 tests, 12 seeds clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"there are still balls which are basically not reachable with auto run..
please recheck" — and, asked twice: "did you write tests for the last issues
we encountered?" The straight answer: no. The unreachable-balls issue only
ever had config assertions, which is why it came back twice.
Pace is the ladder's fourth dial. Width alone never fixed reachability
because the returns left at ~27 m/s, crossing to the baseline in ~0.6s — less
than auto-run needs from the far corner. matchOppPace 0.78 puts opponent one
at ~20 m/s (measured live: twelve returns, 20.0-20.7), a quarter more flight
time, and matchOppPacePer hands it back per rung.
The vignette deepens: dmg x 0.8, cap 0.85, and the clean centre shrinks with
damage instead of only the rim darkening.
And the reachability regression the issue always deserved: a headless 90s
autopilot soak asserting at least 10 returns sampled and at most 15% that
never came within volley reach. Sabotage-verified. Its honest limit is
recorded in PROJECT.md: the autopilot attends perfectly and taps
frame-precisely, so it out-covers a human — it is a floor against
catastrophic regressions, not a proxy for the thumb. Oliver's playtest stays
the real gate.
46 tests, 15 seeds clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three asks. The placement ring: an aimed rally hit stamps a gold ring that
fades where the shot is going — flick, watch it bite the corner, learn the
cause. Not stamped for unaimed hits (marking scatter would teach noise), gone
in 0.8s (a permanent ring reads as a target rather than feedback).
The opponent's width is the ladder's third dial. Its returns scatter off
aimError x (1 - quality), which painted lines from game one. matchWideBase
2.1 keeps opponent one inside what an auto-run player can cover — measured,
25 returns all inside +-2.2m — and matchWidePer earns 0.4m per rung, so
line-painting is the late ladder. Speed, reliability, width: three dials.
The blood vignette replaces the temporary border. The old wash flashed 0.3s
and vanished, so between hits the screen said you were fine. A radial
vignette reads G.hp directly now — its depth is your damage fraction,
permanent until healed, hit pulse riding on top. Global: the hp channel is
global. Same no-HP-bar philosophy one layer out; at 1 hp the frame is
properly blood-rimmed and the court centre stays legible.
44 tests, 12 seeds clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
wuwiandClaude Opus 5