Sylpheed/docs/re/s00a-drive-blocked-by-focus.md

# 🔴 `S00A` is unreachable here — the new-game path crashes on an on-disc cache it cannot finish building

**Four drives.** The first three failed on tooling and are recorded below because
each defect is live in shared scripts. The fourth drove perfectly and hit a
**guest crash**, which is the real answer.

**Classification: measured.** Two drives, 2026-08-29. Reported as a **route
finding** rather than retried silently, because the cause is a defect in shared
harness tooling that will bite the next drive too.

## What was wanted

`S00A` is the new-game intro and the second asset for the centre-channel voice
result ([`voice-three-streams-are-concurrent.md`](structures/voice-three-streams-are-concurrent.md)):
its second full-length stream is **digital silence** where `ADV`'s is a 0.60 ×
copy, so agreement there would stop the finding resting on one movie. It starts
~4.5 s after Ⓐ on the save slot, so it needs a **driven, rendered** run.

## 🔴 Drive 1 — `screen_id.py` cannot see the title it was waiting for

396 s of `other`, two spurious `menu` hits, abort. The guest was healthy
throughout — its audio was decoding the whole time.

`screen_id.py` thresholds on **green** and only returns `title` once the
`PRESS Ⓐ` plate has faded in. This corpus's own finding is that the boot title
shows **build 4 first, plate-less**. Both defects reproduce on committed frames
(see [`METHOD.md`](METHOD.md)):

| frame | `screen_id.py` | should be |
|---|---|---|
| `live-title-build4-no-plate.png` | **`other`** | title |
| `difficulty-screen.png` | **`menu`** | not the main menu |

⚠️ `newgame_path.sh`, `nav_probe.sh` and `boot_menu.sh` all gate on it.

## 🔴 Drive 2 — reached the title, then pressed Ⓐ on the wrong button

With a replacement classifier controlled **6/6** (three oracle frames at 1.000;
`difficulty-screen` and both movie frames correctly `other`, at a 0.85 threshold —
at 0.60 it repeated `screen_id.py`'s own difficulty/menu confusion), the drive
reached `title_plate` at **t = 398 s**, took Ⓐ to the menu, and drove on.

**It ended up in a tutorial mission.** The screen 60 s after the last Ⓐ
correlates **+0.960** with the committed
`tutorial-mission-reached-then-crash.png`. No `S00A` voice stream ever decoded —
the probe logged `ADV`'s three, `BGM_102`'s two, `BGM_103`'s two and two SE cues,
and **none of S00A's** (1 810 432 / 1 263 616 / 98 304).

**Why:** `newgame_path.sh`'s header says *"NEW GAME is the menu's FIRST item, so
this presses (A) with no d-pad movement"*. 🔴 **That assumption contradicts this
corpus's own Q5 result** — initial focus varies boot to boot, four boots giving
`TUTORIAL, TUTORIAL, NEW GAME, NEW GAME`
([`menu-navigation-semantics.md`](menu-navigation-semantics.md)). This boot
started on TUTORIAL, so the first Ⓐ opened the lesson list.

✅ **Incidental corroboration:** `BGM_103`'s two waves (3 876 864 / 3 930 112)
decoded on reaching the menu — an independent confirmation of HANDOFF's
"the menu's music is `BGM_103`", from the runtime rather than from the cue table.

## ❔ What is needed, and why it is not just "press up four times"

Focus must be **detected**, not assumed — and wrap-around makes it
unreachable by counting: ⬆ from the first item goes to the *last*, so no fixed
number of presses lands on a known item from an unknown start.

🔴 **And I do not currently have a focus detector.** A per-row brightness
statistic **failed its control**: on `live-main-menu-options-focused.png` — whose
answer is in its own filename — it picked NEW GAME. Brightness across a button
row is dominated by something other than the focus highlight. Until a controlled
detector exists, a driven run cannot reliably choose a menu item on this boot
path.

## 🟢 A refutation attempt on the port's focus identification — it FAILED

The port identified `live-main-menu.png` as **NEW GAME focused**, by minimum
render difference, with the `-options-focused` capture as a control it picked
correctly by 4.7×.

Differencing the two captures directly and binning by button row, I got NEW GAME
and **EXTRAS** changing while OPTIONS stayed still — an apparent refutation.
**It was my error.** I placed the row bands as `rest_y ± 24`, treating the
resting position as a band *centre*; the changed bands then mis-assigned.

The offset-independent check settles it without needing that geometry at all:

| | |
|---|---|
| changed band centres (design y) | 195.2 and 450.1 |
| separation | **254.9** |
| 3 button pitches (3 × 80) | 240 — off by **14.9** |
| 4 button pitches | 320 — off by 65.1 |

**The two focused buttons are 3 apart, not 4** — NEW GAME → OPTIONS, not
NEW GAME → EXTRAS. ✅ **The port's identification stands.** ⚠️ And as they
themselves noted, it identifies *one frame*, not a rule: Q5's instability is
untouched.

---

## ✅ Drive 4 — the drive worked; the GAME crashes at `SELECT DATA`

With focus detected and the detector validated **live against a known
transition** (`NEW GAME` → ⬇ → `LOAD GAME`, expected `LOAD GAME`, **CONTROL
PASSED**), the drive navigated to `NEW GAME` and pressed through. Every step is
confirmed against a committed capture:

| step | correlates with | r |
|---|---|---|
| after navigation | `main-menu-items.png` | **+0.999** |
| after Ⓐ | `newgame-difficulty.png` | **+0.999** |
| after Ⓐ (NORMAL) | `newgame-selectdata-crash.png` | **+0.997** |

Then the guest throws: **`PC: 0x82307128` ×349**, one
`Guest attempted to throw a C++ exception!`. No `S00A` voice stream ever decodes.

✅ **It is on the new-game path, not the boot.** Ordering in the log is
`ADV` (attract) → **`BGM_103`** (menu music, i.e. the menu was reached) → throw →
crash dumps. So this is *not* the ~100 s boot throw that
[`title-crash-stl-tree.md`](title-crash-stl-tree.md) documents.

## 🟡 Refining that page — the mechanism holds, the container does not

That page attributes the throw to an **incomplete on-disc cache**, and Q4's note
says the `SELECT DATA` crash "is already in the corpus", citing it. **Checked,
and the attribution survives — but not via the container the page names.**

| cache container | state here | |
|---|---|---|
| **`aab216c3`** — the one the page names | **7 files, complete** | not the trigger |
| **`1b556564`** — resolved immediately before the throw | **1 file + `1b556564900c8dcd.tmp`** | ⚠️ exactly the page's **run C**, "partially rebuilt (1 file + a `.tmp`) → 1 GUEST-THROW" |

The log line before the exception is
`HostPathDevice::ResolvePath(\1b556564\9\00c8dcd)`. So the **new-game path
builds a different cache container from the title path**, and it is that one
which is incomplete.

🔴 **And the page's remedy does not transfer.** Its fix is to restore a
*previously complete* cache (run D). **No complete `1b556564` has ever existed in
this container** — the game crashes while building it, leaving the `.tmp`, so the
cache cannot complete and the crash cannot be escaped by re-running. The page's
own run B shows a fully cold cache throws too, so deleting it does not help
either.

## The answer

❔ **`S00A` is not obtainable in this container**, and the reach is:

* the drive is **not** the obstacle — four screens confirmed against committed
  captures at r ≥ 0.997, and the focus detector passed a live transition control;
* the obstacle is a **guest crash** with a known site and a known class of
  trigger;
* the documented remedy needs an artefact — a complete `1b556564` — that has
  never existed here and that the game cannot produce because it crashes
  mid-build.

⚠️ **Consequence for the voice finding:** the centre-channel result stays resting
on `ADV` alone. `S00A` was wanted because its second stream is digital silence
where `ADV`'s is a 0.60 × copy; that corroboration is **not available from this
container** without first solving a crash that is outside the menu-port scope.