Sylpheed/docs/re/ui-keyframe-time-unit.md

# What a keyframe time is worth, and what shape the ramp has

> ## ✅ 2026-08-29 — the argument on this page about WHICH BLOCK OWNS A TIME is over
>
> It was never a choice between two readings. A placement group is
> `frames` records of `{u32 time; 36-byte pose}` after an 8-byte header, so the
> time word **precedes** its pose; the group's "lead-in word" is pose 0's time,
> and **no** time is missing. `SYLPHEED_KF_TIME_SHIFT` had the association right
> and pose 0 untimed, which is the only reason it looked like it cost build 7
> 13.1 % of its pixels. Decoded disc-wide, with controls, in
> [`ui-keyframe-record-layout.md`](ui-keyframe-record-layout.md); the gate is now
> `SYLPHEED_KF_TIME_LEGACY=1`.
>
> **Everything else on this page stands** — the ramp is linear, the clock advances
> 2 units per submitted frame, and `1 unit = 1/60 s` is measured. Read the
> sections below with that correction applied: where a table pairs a time with a
> pose, the pairing is the corrected one.

**Status:** ✅ `CONFIRMED` for the two things the port is blocked on — the ramp is
**linear**, and the animation clock advances **2 keyframe time units per frame the
game submits**. 🟡 the conversion to *seconds* rests on one further step: the game
was measured presenting **27.6 frames/second**, which reads as a 30 Hz title
running at ~92 % under the emulator, giving **1 unit = 1/60 s**. That last step is
reasoning over a measurement, not a measurement — see *the reach of the negative*
below.

This answers `docs/port/MISSION.md` **Q1**. Everything animated downstream — the
title wordmark zoom, the main menu's staggered buttons, every fade — is scaled by
this number.

## The measurement

The question cannot be asked of a screen that is sitting still, and
`log_ui_draws` armed *at* a screen only ever sees the steady state. So arm it
repeatedly through the boot and keep every log: each F10 opens a new numbered
file and **closes** the previous one, which stays on disk complete. Re-arming
every 3 s tiles the whole approach to a screen, and whichever file straddles the
build contains it. `tools/re-capture/screen_build_capture.sh` does this; it sends
no pad input at all, because Ⓐ during the boot has ended a run on a permanent
black screen (`canary-scripted-input-traps.md`).

The developer-logo splash is the cheap target: it is the first thing the guest
draws, roughly ten seconds in, and its bundle (`GP_TITLE.pak`, `--all --build 11`
/ `14` — the `palogo` group) declares short, unambiguous ramps.

The capture is committed: [`captures/ui-timing/splash-build-draws.log`](captures/ui-timing/splash-build-draws.log)
(the raw draw stream) and [`captures/ui-timing/splash-build-quads.csv`](captures/ui-timing/splash-build-quads.csv)
(`tools/re-capture/kf_time_probe.py`, one row per quad per frame: pixel rect and
the per-vertex colour whose high byte is the element's fade).

Frame numbers are the emulator's **VdSwap count** — frames the guest submitted —
so an emulator running at 80 % of real time does not move them. That is the whole
reason to measure in this unit rather than with a stopwatch.

### The quads are the splash, by position and size

Every sprite in the capture lands on its declared placement:

| capture quad | declared element (build 11) | declared placement |
|---|---|---|
| `666x65 @ (307,331)` | `palogo_sqex.t32` 666×68 | (309,330) |
| `525x90 @ (378,155)` | `palogo_gamearts_eff.t32` 521×91 | (379,154) |
| `262x108 @ (512,306)` | `palogo_seta_eff.t32` 261×110 | (511,305) |
| `499x72 @ (390,162)` | `palogo_gamearts.t32` 500×71 | (390,164) |
| `243x86 @ (518,317)` | `palogo_seta.t32` 240×89 | (521,316) |

(A quad runs a few pixels under its sprite; that offset is already recorded in
`ui-title-paint-order-capture.md` and is not what is being measured here.)

## Result 1 — the ramp is linear, exactly

`palogo_gamearts_eff` and `palogo_seta_eff` declare a fade-in of
`t=15 a=0 → t=30 a=255`: a **15-unit** ramp. Their fade alpha, frame by frame,
straight out of the capture:

| frame | 94 | 95 | 96 | 97 | 98 | 99 | 100 | 101 |
|---|---|---|---|---|---|---|---|---|
| alpha | `0x22` 34 | `0x44` 68 | `0x66` 102 | `0x88` 136 | `0xAA` 170 | `0xCC` 204 | `0xEE` 238 | `0xFF` 255 |
| ⇒ units into the ramp *k* | 2 | 4 | 6 | 8 | 10 | 12 | 14 | ≥15 |

`round(255·k/15) = 17k` reproduces **all seven samples with zero error**. An eased
ramp cannot do that: any ease-in/ease-out would bend the first and last steps, and
these are a constant 34 throughout. **Linear interpolation, refuted-nothing-left.**

The same element's fade-*out* gives the same law with a one-count offset —
`255·(1−k/15) − 1` = 254, 220, 186, 152, 118, 84, 67, 33 for
*k* = 0, 2, 4, 6, 8, 10, 11, 13, which is exactly what the capture holds on frames
108–115.

## Result 2 — 2 time units per submitted frame

Read *k* off the row above: 2, 4, 6, 8, 10, 12, 14 on seven consecutive submitted
frames. **The clock advances 2.000 units per frame, over six consecutive
intervals, with no residual.** The fade-out shows the same 2-per-frame step with
one single-unit frame (*k* goes 10 → 11 → 13), so the quantum underneath is 1 unit
and the normal step is two of them.

Cross-checks in the same capture, all consistent:

* the declared hold `t=30 → t=45` (15 units) is held for frames 101–107, **7
  frames** ≈ 14 units;
* `palogo_sqex`'s fade-out steps in multiples of 17 per frame (−34 with the odd
  −17), the same 255/15 quantum;
* the `gamearts`/`seta` pair holds full alpha for frames 116–198 and fades out
  over frames 199–211.

## Result 3 — the conversion to seconds, and how far it reaches

300 submitted frames took **10.87 s** wall-clock, measured by arming one bounded
capture over the splash and timing it to its own `[UI-CAP] done` line:
**27.6 frames/second**.

With 2 units per frame that is **55 units/second measured**. The two readings that
fit are:

* **30 Hz present, 60-unit-per-second timeline** — the emulator running at 92 % of
  real time. A 30 Hz renderer stepping a 60 Hz timeline is exactly the 2-units-
  per-frame quantum that was measured, and 92 % is an ordinary number for this
  container.
* 60 Hz present at 46 % of real time, giving a 120-unit-per-second timeline. This
  requires the emulator to be running at *less than half speed* while drawing four
  quads over a black screen, and it requires the game's timeline to tick at 120 Hz.

**Taking the first**: `1 unit = 1/60 s`. The title build (`GP_TITLE.pak` build 4,
t = 16 … 269) is then **4.2 s**, and the main menu build (build 5, t = 12 … 80)
**1.1 s**.

**The reach of this negative** *(as written before the test below — kept for the
reasoning)*: the present rate was measured *during the boot splash*, where the
guest is also streaming from the ISO, so it is a lower bound on the emulator's
speed and cannot by itself exclude the 60 Hz reading. What would settle it is the
same 300-frame timing taken on the **idle title screen**, where nothing is
loading — if that also comes out near 28 fps the game is 30 Hz and the unit is
1/60 s; if it doubles to ~55 fps the game is 60 Hz and every duration on this page
halves.

## ✅ That test was run — 2026-08-28. The game is 30 Hz; `1 unit = 1/60 s`

Reached the title with a single Ⓐ to skip the intro
([`movie-binding.md`](movie-binding.md)), let it settle 12 s so nothing was
loading, then armed the bounded 300-frame capture and timed it to its own
`[UI-CAP] done` line:

| | 300 frames in | rate |
|---|---|---|
| idle title, trial 1 | 10.40 s | **28.8 fps** |
| idle title, trial 2 | *arm produced no capture* | — |
| idle title, trial 3 | 10.60 s | **28.3 fps** |
| *(prior)* boot splash | 10.87 s | 27.6 fps |

**~28.5 fps on an idle title — the same rate as the loading splash.** By this
page's own criterion that settles it: the title presents at **30 Hz**, the
timeline ticks at **60 units/second**, and **`1 unit = 1/60 s`**.

**The 60 Hz reading is now excluded, not merely disfavoured.** It would require
the emulator to be running at 47 % of real time *while sitting idle on a static
title* — and the capture says that screen costs **1 526 draws over 300 frames,
about 5 draws per frame**. Nothing there is expensive enough to halve the
emulator's speed, and the splash and the idle title returning the same rate is
exactly what a constant ~95 %-of-real-time emulator looks like.

So the durations on this page stand as written: title build 4 ≈ **4.2 s**, main
menu build 5 ≈ **1.1 s**, and `EXTRAS` build 6's declared fade-in ≈ **0.87 s**.

⚠️ Still **measured, not decoded** — no field on the disc says "sixtieths of a
second". What changed is that the measurement now has an idle-state control and
the competing reading is ruled out.

## What this does not say

* Nothing here is a *decode*: no field on the disc says "sixtieths of a second".
  The number is measured from the running game, and the port is transcribing a
  measurement, not a disc value.
* The splash drew `palogo_sqex`, then `palogo_gamearts` + `palogo_seta` with their
  `_eff` sprites — but **not** `palogo_anima` (declared at 388×136 @ (446,449)) and
  not `palogo_sqex_eff`, in the ~7 s of capture after the pair faded out. The
  `_sqex_eff` absence is explainable (the capture joined the SQUARE ENIX logo
  mid-hold); the missing studio-anima logo is not, and is left as an open
  observation for Q2 rather than a claim.

---

## 🟡 Open: the `palogo_*` LOGO elements do not play their declared timeline

**Added 2026-08-28.** Everything above stands — it rests on the `_eff` glows, and
they reproduce exactly. What follows is a **different element in the same bundle
and the same capture**, and it does not.

Reproduce with
[`tools/re-capture/splash_ramp_check.py`](../../tools/re-capture/splash_ramp_check.py);
output committed at [`data/splash-ramp-check.txt`](data/splash-ramp-check.txt).

### Why this is a test and not a fit

The clock is calibrated on `palogo_gamearts_eff` — its declared 15-unit fade-in
`0@15 → 255@30` against its captured alphas 34, 68, 102, 136, 170, 204, 238, a
constant step of exactly 34. That gives `t = 2f − 171`, and the calibration
checks itself: the glow's declared hold ends at `t=45`, predicted **frame 108.0**,
and the observed last full-alpha frame is **107**.

That calibration is then applied to `palogo_gamearts` — same bundle, same frames,
**no free parameter left**:

| declared | predicted frame | observed alpha |
|---|---|---|
| `a=0` at `t=15` | 93.0 | not drawn |
| `a=0` at `t=30` | 100.5 | not drawn |
| `a=255` at `t=190` | 180.5 | 255 |
| `a=232` at `t=206` | 188.5 | **255** |
| `a=32` at `t=210` | 190.5 | **255** |

The logo is still at full alpha nine frames after it should have been at `a=32`.
Its fade-out actually runs frames **199–211**, some 17 frames late, and its
declared 80-frame fade-in (`t=30→190`) is **not drawn at all** — the element's
first appearance, frame 116, is already at 255.

⚠️ "Not drawn" is not a culling artefact: the same element is submitted all the
way down to `a=7` on the way out, so low-alpha quads plainly do reach the GPU.

### The shape mismatch, which needs no calibration at all

The declared fade-out spends **12 of its 16 units** dropping only **23/255** of
the alpha — a near-flat leg — then **200/255** in the remaining 4. The captured
per-frame drops are `1, 11, 6, 22, 34, 33, 17, 33, 33, 17, 25, 8, 8`. There is no
near-flat leg.

### 🟡 A candidate, offered as one and NOT adopted

If the word at `+36` were the time of the **next** keyframe rather than of its own
block, the logo's fade-out would read `255@190 → 232@194 → 32@206 → 0@210`:
slow, fast, slow — which is the captured shape. Fitting both readings to the
captured fade-out gives RMS alpha error **4.05** shifted against **12.13** as
decoded, on two elements independently (`gamearts` and `sqex`).

It also removes a special case. The decoder currently notes that a group's data
"stops 4 bytes short of its final block's time slot", so the last block's time is
unreadable. Under the shifted reading the last block simply *has no successor*,
so it has no time word — the same bytes, no special case.

**Not adopted, for three reasons:**

1. It does not explain the missing fade-in, which is the larger anomaly.
2. It does not fix the 17-frame lateness of the fade-out.
3. The `_eff` elements **do not discriminate** between the two readings — I
   checked: with four blocks the shift merely relabels which phase is which, and
   both reproduce the observed fade-in / hold / fade-out. So the entire case for
   the shift rests on one element's fade-out shape.

**The decoder is unchanged.** 🟡 What a port should take from this: the
interpolation *law* is settled (linear, 2 units per frame), and the *group
timeline* for multi-keyframe elements is not.

### 🔴 One thing I got wrong in the course of this

I first reported the `_eff` glows as holding "a constant α ≈ 33" and read that as
contradicting the declared 255 plateau. They do no such thing — they ramp 34 →
255 in exact steps of 34. I had printed only the series' **minimum** and read it
as its range. Withdrawn; the trap is in [METHOD.md](METHOD.md).

### ✅ Settled the next iteration: the hold duration decides it

The section above left the candidate unadopted because its whole case rested on
one fade-out's *shape*. There is a much blunter measurement in the same capture,
and it needs **no calibration at all** — just the ratio of two observed spans:

| | full-alpha hold : fade-out | ratio |
|---|---|---|
| **observed** | 83 frames : 13 frames | **6.38** |
| as decoded | 4 units : 16 units | 0.25 — **off by 26×** |
| `+36` = the NEXT pose's time | 160 units : 20 units | **8.00** |

With the glow's 2 units/frame fixed and nothing else free, the current reading
predicts `palogo_gamearts` holds full alpha for **2.0 frames**. The capture holds
it for **83**. The shifted reading predicts **80.0**.

That is no longer a shape argument, and it is not a fit — it is a factor of 26.

### A fifth argument, from the corpus's own behaviour

`Element::rest()` tries `rest_plateau()` first and falls back to picking the
keyframe with the longest **dwell**. Run that fallback on `palogo_gamearts`:

| reading | dwell winner |
|---|---|
| as decoded | **`a=0`** (dwell 160) — a fully transparent pose, for a logo |
| shifted | **`a=255`** (dwell 160) — the visible hold |

Under the current reading the plain dwell rule picks an invisible pose for a
publisher logo, and only the `rest_plateau` special case rescues the render. That
special case has been repeatedly troublesome — it is the one the port agent
reported a bug in on 2026-08-28. Under the shifted reading the dwell rule is
simply correct on its own.

### What the shift is, exactly

Blocks `0…n−1`; `W[k]` is the word at block `k`'s `+36`. `W[n−1]` lies outside the
group, because a group owns `frames·40 − 4` bytes.

* **as decoded:** `kf[k].time = W[k]`; the last keyframe has no time, as a special
  case for the missing word.
* **shifted:** `kf[k].time = W[k−1]`; the **first** keyframe has no time, and
  `kf[n−1]` takes `W[n−2]` — which exists. Nothing is missing and nothing is
  special-cased: `W[k]` is simply *the time at which pose `k+1` is reached*.

Gated by `SYLPHEED_KF_TIME_SHIFT=1`, default unchanged. See below for what
adopting it would cost.

### 🔴 …and what it costs — which is why the default is UNCHANGED

Rendering every build of six UI paks under both readings, and byte-comparing:

| | result |
|---|---|
| builds rendered identically | 10 of 11 compared |
| builds that changed | **`GP_TITLE` build 7** — 13.1 % of pixels |

Build 7 is the **Japanese twin of build 4**, and build 4 — the one verified
against a live capture — is byte-identical under both readings. So the one build
the shift moves is the one with no capture to adjudicate it. The available proxy
says the shift makes it *worse*:

| build 7 rendered | corr vs verified build 4 | mean luminance |
|---|---|---|
| as decoded | 0.6206 | **70.94** (build 4: 71.41) |
| shifted | 0.6201 | 76.32 |

Two language twins are the same artwork with different text, so their brightness
should match closely. As decoded it matches to **0.5**; shifted it is **4.9**
brighter. Correlation does not separate them (0.6206 vs 0.6201) — the luminance
does.

### The honest position

These two results are about **different things**, and both are real:

* For **animation timing** — what the port actually needs to play a screen — the
  shifted reading is favoured by a factor of 26 on a calibration-free
  measurement, and the current reading predicts a 2-frame hold where the game
  holds 83.
* For **resting-pose selection** — what `screen render` does — the current
  reading matches the EN/JP twin brightness and the shifted one does not.

They can both be true: `rest()`'s dwell fallback is a heuristic layered *on top
of* the times, and moving the times moves its tie-breaks. Adopting the shift
would mean revisiting that heuristic in the same change, and there is no capture
of build 7 to verify the result against.

**So the default stays as decoded.** 🟡 The port should treat a multi-keyframe
group's *timing* as unverified — and specifically should not expect a
2-frame hold where the game holds 83 — while `rest()` and `screen render` remain
as they are.

---

## ✅ Replicated: three elements, two screens, and the shifted reading wins every time

**2026-08-29.** The case for reading `+36` as *"the time the NEXT pose is
reached"* rested on one element's fade-out shape, then on one element's hold
duration. Both splash halves supply more, and they all say the same thing.

Phase durations in frames (2 units/frame, from the `_eff` glow ramp):

| element | screen | observed hold at `a=255` | as decoded | shifted |
|---|---|---|---|---|
| `palogo_gamearts` | developer splash | **83 f** | 8 f | **80 f** |
| `palogo_seta` | developer splash | **83 f** | 6 f | **80 f** |
| `palogo_sqex` | publisher splash | **≥ 77 f** * | 6 f | **102 f** |

\* the capture opens mid-hold at frame 1, so 77 is a floor, not the length.

The two readings predict **opposite structures** for these elements. On
`palogo_gamearts`, as decoded: `hold 8f, in 80f, hold 2f, out 6f, out 2f` —
an eighty-frame *fade-in* and a two-frame hold. Shifted: `in 8f, hold 80f,
out 2f, out 6f, out 2f` — an eight-frame fade-in and an eighty-frame *hold*.
The capture shows a **83-frame hold** and no fade-in at all.

### And the elements that cannot discriminate are not contradicted

`palogo_gamearts_eff` observed `in 7f, hold 7f, out 8f`. As decoded it reads
`in 8f, hold 8f`; shifted it reads `hold 8f, out 8f`. Both fit — with only four
blocks the shift relabels which phase is which without changing any duration. So
the glows, which are what Q1's linear law was measured on, **do not argue against
the shift**; they simply say nothing about it.

### 🟡 Why the decoder's default is still unchanged

The one thing that opposes the shift is `Element::rest()` on `ptlogo_eff3.t32`,
where the shifted reading makes the longest-dwell fallback return the bloom's
200 % peak instead of an invisible frame. But that fallback is **unsound whenever
it runs** ([resting pose](structures/ui-resting-pose.md)) — it returns an
endpoint of a movement, and neither endpoint is held. Checked: the shift does not
fix it either, so this is not a case of two readings disagreeing about the times.
It is a heuristic guessing, in both.

**So the two questions are separable, and only one of them has evidence.** The
times govern animation; `rest()` picks a static pose through a rule that consults
them only after its own precondition has failed.

### ⚠️ What a port should do

For **animation timing**, read `+36` as the *next* pose's time: three elements
across two screens, each off by an order of magnitude under the other reading.
For **static composites**, nothing changes — `screen render` is unaffected, and
the five screens' correlations stand ([acceptance](five-screens-acceptance.md)).

🟡 Classified **measured, not decoded**: this is three elements in one game
screen family, not a disc-wide field check, and our own decoder still defaults to
the other reading behind `SYLPHEED_KF_TIME_SHIFT=1`.