Sylpheed/docs/re/structures/isl-condition-builtins.md

# 🟡 The three built-ins that appear *inside* clear conditions — read, not named

[`isl-phase-guards`](isl-phase-guards.md) produced the per-phase clear
conditions, but three of the built-ins in them were unread, so the conditions
were only half-readable. All three are read now. **None of them is named** — the
corpus has withdrawn two names taken from usage shape, and what I read is not
enough to name these.

Implementations via the vtable at `0x820A84BC` (control: built-in 69
`unit_state` → slot 184 → `0x8226ADF0`, as recorded).

| built-in | slot | implementation |
|---|---|---|
| **104** | 288 | `0x8226BFE0` |
| **141** | 428 | `0x8226C9B8` |
| **7** | 40 | `0x82266C68` |

## ✅ `builtin104` is a pure getter — and it is the whole tutorial clear condition

```
8226BFE0  lwz  r11, 10160(r3)     ; r3 = the ScriptPhase
8226BFE4  stw  r11, 164(r3)       ; -> special[0]
8226BFE8  blr
```

Three instructions. It returns **`[phase+10160]`** and nothing else. So all six
tutorial stages (S18–S23) end on the value of a **single engine-written word** —
which is exactly why their exits have one dominating condition each and why
`isl-builtins.md` saw built-in 104 only ever inside a poll loop.

**That word has exactly one writer** in the whole image:

```
821AAD74  lwz   r11, 104(r30)
821AAD84  lwz   r11, 12(r11)
821AAD88  cmpi  cr6, 0, r11, 16     ; skip if == 16
821AAD90  cmpi  cr6, 0, r11, 32     ; skip if  > 32
821AAD98  lwz   r11, 4(r10)
821AAD9C  stw   r29, 10160(r11)     ; <- the only write
```

inside `sub_821AA1B0`, gated on a type/kind field being in `(16, 32]`. `r29`
there comes from `or r29, r3, r3` — the return of a preceding call.

## ✅ (2026-08-27) NAMED — built-in 104 is `request_next`, and the game says so itself

`r29` is the return of a **getter on an object fetched from a registry by id**:

```
821AAD1C  lwz  r4, 2424(r30)        ; the registry
821AAD20  addis r5, r0, 0x20FF
821AAD28  ori   r5, r5, 0xFF02      ; <- the id
821AAD38  bcctrl                    ; registry->slot1(&out, 0x20FFFF02)
821AAD40  lwz  r11, 0(r3) ; lwz r11, 8(r11)
821AAD4C  bcctrl                    ; r29 = obj->slot2()      -- GET
821AAD9C  stw  r29, 10160(r11)      ; -> [phase+10160]
821AADA0..DD4                       ; then the SAME id again, obj->slot1(0)  -- CLEAR
```

So the frame loop **reads the value, publishes it to the script, and clears it**.

**The id namespace has exactly three members** — `0x20FFFF00`, `0x20FFFF01`,
`0x20FFFF02` — each built at exactly 4 sites image-wide. Two of those sites are
in `sub_821D5178`, which fetches `0x20FFFF01` and `0x20FFFF02`, calls the same
`slot2()` getter on each, and passes both to a log call whose format string is
at `0x820A4968`:

> `silph::GamePart_ReadyRoom::Impl::OnCommand - Wait() command is requested.`
> `Check flow control valiables. WAIT_MODE : %d, REQUEST_NEXT : %d`

Argument order settles which is which: `r4` (first `%d`, **WAIT_MODE**) is
`0x20FFFF01`'s value, `r5` (second, **REQUEST_NEXT**) is `0x20FFFF02`'s.

> ⇒ **`[phase+10160]` is `REQUEST_NEXT`, and built-in 104 returns it.**

That is the game's own name for the variable, not a shape-based guess.
Corroboration: `GamePart_ReadyRoom::Impl::PrepareScript` sets `WAIT_MODE = 1`
and `REQUEST_NEXT = 0` before an ISL script runs, and `sub_821AA1B0` clears both
each frame after reading. `[phase+10160]` has **exactly one writer and one
reader** in the whole image (the two above); the three other hits on offset
10160 are `lfs` on unrelated objects.

So the six tutorial stages' exit condition `request_next() != 1` is the script
waiting on the surrounding game part's *proceed* flag — which is why S18–S23
each have exactly one dominating condition.

🟡 Still unread: what the `(16, 32]` gate on `[r30+104]`'s `+12` selects, and
`0x20FFFF00`, the third id (used by the `GRAPH_PATH` / `EX_FONT` / `SYSTEM`
code in `sub_821D6350` and `sub_821D6A40`).

## ✅ (2026-08-27) `[phase+10152]` and `[phase+10156]` are an ANSWER and a STATE

The backlog carried these as "9 and 7 writers, unread". They are read now, and
they are not `request_next`'s siblings-by-adjacency: they belong to a **trio of
built-ins** sitting in four consecutive vtable stubs.

| built-in | vtable slot | handler | body |
|---|---|---|---|
| **102** | 70 | `sub_8226BF48` | the state machine below |
| **103** | 71 | `sub_8226BFA8` | returns `state != 0 && answer == 1` |
| 104 | 72 | `sub_8226BFE0` | `request_next` (above) |
| **130** | 97 | `sub_8226BFD0` | `[phase+10152] = 0` |

Built-in 102, read straight off `0x8226BF48`:

```
state = [phase+10156]
  state == 1  ->  return 2                                   ; still pending
  state == 2  ->  special[0] = (answer == 1) ; state = 0 ; return 0
  otherwise   ->  state = 1 ; answer = 0 ; return 2           ; issue the request
```

**`2` and `0` are the dispatcher's thread codes**, so this is a *blocking*
built-in: it yields the coroutine until an answer arrives, then hands it back.
Built-in 103 is the **non-blocking** form of the same test, and 130 clears the
answer slot.

⇒ **`[phase+10156]` is a request STATE — 0 idle, 1 pending, 2 answered — and
`[phase+10152]` is the ANSWER**, compared against `1`.

The answer is published by `sub_821A9DC8`, which writes
`[phase+10152] = [r6+4]` and `[phase+10156] = 2` **under the same `(16, 32]`
gate on `[r30+104]+12`** that `request_next`'s writer uses — so one
engine→script publish path serves both mechanisms.

⚠️ **A name I did not verify.** `isl.py` calls built-in 102 `prompt_yes_no`,
and the mechanism above is *consistent* with a modal yes/no prompt — but
grepping the corpus, that name appears only in a list of built-ins unused by
Stage 02, with **no derivation recorded anywhere**. What is actually read here is
"a blocking request whose answer is tested against 1". 103 and 130 are therefore
left unnamed rather than named off an unverified premise.

🟡 Also unread: what question is being asked. `sub_821A9DC8` has **no strings**
and exactly one reference in the image — a tail `b` from `0x821AC064` — so the
string recipe finds nothing there.

## 🔴✅ (2026-08-27) The gate is `!= 16 && <= 32` — and it is a VALIDITY CHECK, not a selector

🔴 **First, a correction to this page's own notation.** Written as `(16, 32]` it
implies a lower bound. There is none:

```
8225EC90  cmpi cr6, 0, r4, 16
8225EC94  stw  r4, 12(r30)          ; <- the field IS the argument
8225EC98  bc   12, eq, <bail>       ; == 16 -> bail
8225EC9C  cmpi cr6, 0, r4, 32
8225ECA0  bc   12, gt, <bail>       ; >  32 -> bail
```

Everything **below** 16 passes. The condition is `kind != 16 && kind <= 32`.

✅ **`X+12` is a constructor parameter, not engine state.** `sub_8225EC78(X,
kind, …)` — the function carrying the string `"script load cancel\n"` — stores
its second argument into `+12` and applies the identical pair of tests to it
immediately, bailing out of the load. Every later occurrence re-tests the field
it just stored. So the gate is that object's **invariant**, re-checked wherever
it is touched, not a switch selecting between behaviours.

✅ **Where it is checked**: 42 sites image-wide with this exact `cmpi 16` →
`cmpi 32` shape, **34 of them in one code region** — `sub_821A9DC8`,
`sub_821AA1B0`, `sub_821AB570`, `sub_821AB650` — plus 3 in `sub_8225EC78`. One
object, guarded everywhere.

✅ **And the object chain is now read.** `X` is installed at `0x821A78EC`
(`stw r25, 104(r30)`), the previous one torn down first through `sub_8225EB60`;
so `X = [GamePart+104]` is **the game part's current script instance**. From
there `X+4` → `Y`, and `Y+4` → the **ScriptPhase**, `Y+72` = the answer, with
`[ScriptPhase+10152]` receiving the same value. `X+12`'s value arrives as
`[r21+8]`, `r21` being `sub_821A6CF0`'s own second argument.

🔴 **A tempting reading, killed by its control.** GamePart ids (from the
`RegisterToFactory<N,…>` strings) run 0…27, and **16 is not among them** — which
makes "`X+12` is a GamePart id and 16 is the unregistered one" very inviting.
It fails: **1, 2 and 18 are also absent** from that list, so 16 is one of *four*
gaps, not a unique one. Nothing distinguishes it. The reading is not adopted.

🟡 So what the kind **means** is still open — the gate is now read, but the
value's domain is not.

## ✅ (2026-08-27) The code region is NAMED: `GamePart_MainGame`

`sub_821A6CF0` has no `bl` callers, only a tail `b` from `0x821AC04C`. That
address sits in a run of adjustor thunks, and the table pointing at them is a
**vtable at `0x820A319C`** (RTTI locator at `0x820A3198` = `0x8210BB58`):

| slot | entry | resolves to |
|---|---|---|
| 0 | `sub_821ABFA8` | destructor |
| 1 | `0x821AC040` | **`addi r3, r0, 17; blr`** — returns the factory id |
| 4 | `0x821AC048` | `lwz r3, 8(r3); b sub_821A6CF0` |
| 5 | `0x821AC050` | `→ sub_821A82A0` |
| 6 | `0x821AC058` | `→ sub_821A8428` |
| 7 | `0x821AC060` | `→ sub_821A9DC8` |
| 8 | `0x821AC078` | `→ sub_821A9CF0` (guarded `r4 == 9`) |
| **9** | `0x821AC068` | **`→ sub_821AA1B0`** — the per-frame Update |
| 10 | `0x821AC070` | `→ sub_821AB570` |

Slot 1 returning **17** is decisive: `RegisterToFactory<17, class
silph::GamePart_MainGame>`. Every thunk does `lwz r3, 8(r3)` first, so these are
**`GamePart_MainGame::Impl` methods** reached through the outer class. The code
region this page has been reading is now named rather than inferred.

## 🟡 A lead on the kind's domain — better-controlled than the last one, still not adopted

The disc carries **28 mission scripts, numbered `Stage01`–`Stage16` and
`Stage18`–`Stage29`** (as extracted; the corpus's own "all 28 `StageNN.ssb`").
**Exactly one gap: Stage17.** The gate excludes **exactly one value: 16**. Under
0-based indexing `16 ↔ Stage17`, and `<= 32` comfortably bounds indices 0…28.

Compare the controls honestly:

| reading | values missing from the domain | gate excludes | match |
|---|---|---|---|
| GamePart id (refuted) | **4** — 1, 2, 16, 18 | 1 value | 1-of-4 — worthless |
| stage index (this) | **1** — Stage17 | 1 value | 1-of-1 |

That is a much tighter fit, and it is *still* coincidence-shaped. **Not adopted.**

### 🔴 (2026-08-27) The test I proposed for it DOES NOT EXIST

The plan was to find the `Stage%02d` construction site and see whether it is fed
`index` or `index + 1`. Two searches say there is no such site:

* **No stage-shaped format string.** The `'%s%02d'` at `0x820A9C8C` has exactly
  **one** xref, into `sub_822929E0`, a generic string helper. Listing *every*
  short `%d`-bearing string in the image — 43 of them — turns up no
  `Stage`/`_S%02d` pattern at all.
* **No precomputed name hash.** Using the corpus's own `name_hash`, the values for
  `Stage_S00…30`, `Stage00…30`, `StageNN.ssb` and `UnitGroup_S00…30` appear
  **nowhere in the image as a 4-byte word**.

So the executable never builds a stage script's name; the mapping lives on the
data side. **The stage-index lead cannot be settled this way**, and the next pass
should not retry it. The one remaining static handle is another hop up: `X+12`
arrives as `[r21+8]` in `GamePart_MainGame` vtable slot 4, so it comes from
whatever the GamePartTask manager passes when it switches parts.

## ✅✅ (2026-08-27) SETTLED — `X+12` IS A 0-BASED STAGE INDEX, and the game's own strings prove it

Following the *value* instead of the filename does it in two hops.

**Hop 1 — where it is stored.** `sub_82260568` passes the kind as `r7` into the
phase initialisers. `SilphScriptPhase`'s constructor `sub_8225FEF8` **never
touches `r7`** (zero mentions in the whole function); the *base* constructor
`sub_822700C0` does:

```
822700F0  or   r26, r7, r7
82270188  stw  r26, 152(r30)      ; -> [phase+152]
```

**Hop 2 — who reads it.** `sub_82261F70` builds a table of string pointers on its
stack at `r31+144…` and indexes it with that field:

```
8226230C  lwz    r11, 152(r21)        ; the stored kind
82262310  addi   r10, r31, 144        ; the table
82262318  rlwinm r11, r11, 2, 0, 29   ; kind * 4
8226231C  lwzx   r4, r11, r10         ; table[kind] -> a STRING
```

The table is contiguous at `0x820A8880`, 20 bytes per entry:

| index | string |
|---|---|
| 0 | `STAGE01_UNIT_MAX` |
| 1 | `STAGE02_UNIT_MAX` |
| … | … |
| **16** | **`STAGE17_UNIT_MAX`** |
| … | … |
| **32** | **`STAGE33_UNIT_MAX`** |
| 33 | `PLANE` — the table ends |

> ⇒ **`X+12` = `[phase+152]` = a 0-BASED STAGE INDEX**, `index N → STAGE(N+1)`.

The gate falls out of it exactly:

* **`kind <= 32`** is the **array bound** — the table has precisely 33 entries,
  0…32, with `PLANE` at 33 where it stops.
* **`kind != 16`** is **`STAGE17`** — the one stage number with no `.ssb` on the
  disc (files run `Stage01`–`Stage16`, `Stage18`–`Stage29`).

✅ **This retracts the "not adopted" above.** The stage-index reading was recorded
as 1-of-1 but coincidence-shaped and deliberately not believed; it is now read off
the game's own string table, and **0-based is proved by `index 5 → STAGE06`**,
not assumed.

⚠️ **A false positive caught on the way.** `0x82272D88 lwz r11, 152(r11)` inside
the built-in switch looks like a read of this field. It is not: `r11` had just
been loaded from `0(r31)`, so it is the **vptr** — that instruction is a virtual
call to **slot 38**. Offset 152 recurs; the base register decides.

## ✅ (2026-08-27) What `STAGENN_UNIT_MAX` is: two per-stage unit caps, `PLANE` and `VESSEL`

The lookup in `sub_82261F70` continues straight past the table read. `r21` is the
phase there; `r31` is the **stack frame** (`addi r4, r31, 452`, `lwz r11, 376(r31)`
— easy to misread as phase fields).

```
82262308  bl   0x824480D0                 ; open the config node -> [r31+84]
8226230C  lwz  r11, 152(r21)              ; the 0-based stage index
8226231C  lwzx r4, r11, r10               ; "STAGE(N+1)_UNIT_MAX"
82262320  bl   0x82448AA0                 ; find that section
82262328  bc   4, eq, 0x82262340          ; found -> read two keys
   ; NOT found — defaults:
82262334  stw  r11(=20),  360(r21)
82262338  stw  r10(=200), 356(r21)
   ; found:
82262348  addi r4, r11, -29932            ; 0x820A8B14 = "PLANE"
82262360  stw  r11, 356(r21)
8226235C  addi r4, r10, -29924            ; 0x820A8B1C = "VESSEL"
82262368  stw  r3,  360(r21)
```

| phase field | config key | default |
|---|---|---|
| `[phase+356]` | **`PLANE`** | **200** |
| `[phase+360]` | **`VESSEL`** | **20** |

So the section is a per-stage pair of **unit caps** — planes (fighters) and
vessels (capital ships) — with the section name built from the stage index and
the whole thing falling back to 200/20 when a stage has no entry. All three names
are the game's own.

The surrounding reads name the config itself: `GP_SCRIPT` (`0x820A8244`),
`SCRIPTS`, `Resource2D`, `TABLE`. The pak census counts **40 `<?xm` entries**, so
this is an XML config reader (`sub_824480D0` open, `sub_82448AA0` find-section,
`sub_824482D0` read-int, `sub_82448C50` / `sub_824490E0` others).

🟡 **The values themselves are not confirmed from the data side.** A crude
`Z1`+zlib scan over `dat/*.pak` found no `UNIT_MAX` — but that scan did not use
the corpus's own pak reader (`unitgroup.py`), so it is **not** evidence of
absence. 🟡 No reader of `[phase+356]` / `[phase+360]` was identified: the
offsets recur, and the one ISL-region hit that looked promising
(`0x82273400 lwz r11, 360(r11)`) is a **vtable slot 90 call** — `r11` is the
vptr — which is built-in 123, consistent with the known mapping.

### ✅🔴 (2026-08-27) The values: the config is an **INI**, it ships nearly EMPTY, and every stage takes the defaults

🔴 **First, my own label from the previous pass is wrong.** I called
`sub_824480D0` / `sub_82448AA0` / `sub_824482D0` an *XML* config reader because
the pak census counts 40 `<?xml` entries. Those 40 are **XPR2 resource
manifests** — `<RDF Version="XPR2">`, `<XBGMesh Name=… DataFile=…/>`,
`<Texture Name=… Source=… Format="D3DFMT_DXT1_SRGB"/>` — mesh and texture
declarations, nothing to do with game config.

The real config is a **loose INI at the disc root**, `config.ini`, whose header
comment reads *「アプリケーション/ゲームパート初期設定テーブル」* — "application /
game-part initial-settings table" — and notes that the `SYSTEM` section holds
settings the game and every game part share. That matches every key this code
touches: `SYSTEM`, `GP_SCRIPT`, `SCRIPTS`, `Resource2D`, `TABLE`, `GRAPH_PATH`,
`EX_FONT`, `SCRIPT_ID`, `SCRIPT_PATH`, `BASE_INFO`, `PLANE`, `VESSEL`,
`STAGENN_UNIT_MAX` — INI section and key names.

**And the shipped file is 400 bytes.** In full it has two sections:

```
[SYSTEM]

[LANGUAGE]
        = eng   ; default
#0x01   = eng   ; XC_LANGUAGE_ENGLISH
#0x02   = jpn   ; XC_LANGUAGE_JAPANESE
… deu / fra / esp / ita
```

`[SYSTEM]` is **empty**, and there is no `[STAGENN_UNIT_MAX]` anywhere.

| check | result |
|---|---|
| `.ini` files on the extracted disc | **1** — `config.ini`, 400 bytes, printed in full |
| `[STAGENN_UNIT_MAX]` sections in it | **0** |
| `UNIT_MAX` in any pak entry | **0** |
| `VESSEL` in any pak entry | **0** |

The pak scan is **controlled**: the same loop over all 41 paks decompressed
**26 443 entries** and found `MSG_DEMO` 192×, `mapmesh_box_500km` 162× and
`<?xml` exactly 40× — the census number. So the reader works and the absence is
real. (An earlier hand-rolled scan used `s[8:]` instead of `s[10:]` for the `Z1`
header and searched the index file rather than the `.pNN` data — that one's null
result was worthless.)

> ⇒ **The lookup always misses, so every stage runs on the defaults:
> `PLANE = 200`, `VESSEL = 20`.**

🟡 A hash search was run too, in case the section were name-hashed like a pak
entry: `name_hash("STAGE01…33_UNIT_MAX")`, `name_hash("PLANE")`,
`name_hash("VESSEL")` — **0 hits as pak entry keys**. Eight of the 35 hashes did
appear as 4-byte words *inside* entries, seven of them in the 1.1 GB `sound.pak`;
that is the chance rate for a 32-bit needle in a blob that size, not a finding.

### ✅ (2026-08-27) The absence re-tested properly — and the reader is NOT INI-only

🔴 **Correction to the label above.** `sub_82448AA0` is not "the INI config
reader". Resolving the `r4` string at **every** call site of it and of
`sub_824482D0` (artefact `data/config-keys.txt`) gives **65 section names and 54
int keys** — `MISSIONS`, `FONTS`, `StageResource`, `EPILOGUE_MOVIES`, `WEAPONS`,
`UNITS`, `SOUNDS`, `ACHIEVEMENTS_REQUIREMENTS`, `LINE_PITCH`, `MSG_FONT_SIZE`,
`SUBTITLE_Y` … far more than the 400-byte `config.ini` holds. And the decisive
check: **`BASE_INFO`, `SYSTEM`, `MISSIONS` and `FONTS` are all present as IDXD
record keys on the disc.** So this pair is a **generic named-section accessor
over the IDXD `.tbl` containers**, with `config.ini` as one small extra input.

⚠️ **My first re-test was invalid, and only its control showed it.** IDXD record
keys are **not** `name_hash` — they are **`tag_hash`** (case-sensitive, modulus
`0x00FFFFDF`; `unitgroup.py` documents it). Searching 190 782 records with
`name_hash` returned 0 for `STAGENN_UNIT_MAX` *and* 0 for every control name,
which is what exposed it.

Redone with `tag_hash`:

| | |
|---|---|
| IDXD entries parsed | **7 750 / 7 750** (matches the corpus census) |
| records scanned | **190 782**, 3 496 distinct keys |
| control names found as record keys | **`BASE_INFO`, `SYSTEM`, `MISSIONS`, `FONTS`** ✅ |
| `STAGE01…33_UNIT_MAX` records | **0** |

With the literal search (0 in 26 443 decompressed entries) and `config.ini`
printed in full, the section exists **nowhere on the disc** — not as an INI
section, not as an IDXD record. The earlier worry that the `GP_HANGAR_ARSENAL`
word-hit might be real is settled: it is not a record key.

> ⇒ **`PLANE = 200` / `VESSEL = 20` for every stage** — now on a controlled test
> rather than a scan whose reader had never been checked.

### ✅ (2026-08-27) `Stage30`–`Stage33` do not exist — the 33-slot table is headroom

Probing the pak entry names (`name_hash`, the right hash for *entry* names) for
each stage-table family across all 41 paks:

| family | stages present |
|---|---|
| `stage\UnitGroup_S%02d.tbl` | **01–16, 18–29 — 28 stages** |
| `stage\AIParams_S%02d.tbl` | 01–16, **24–29 — 22 stages** |
| `stage\Stage_S%02d.tbl` | 0 — expected; `stagetbl.py` documents that the per-stage record is *not* name-addressed |

The control is built in: 28 hits out of 33 probes for `UnitGroup`, so the loop
resolves entry names correctly. **S17 and S30–S33 are absent**, so the 33-entry
`STAGENN_UNIT_MAX` table is oversized headroom over a 28-stage game.

🔑 **An independent corroboration falls out of the second row.** `AIParams` is
missing for exactly **S18–S23** — precisely the six stages this page identifies
as the tutorials by a completely different route (their only dominating
condition is `request_next() != 1`). Tutorials ship no AI parameters.

### 🟡 Who consumes `[phase+356]` / `[phase+360]`: not findable by offset

Image-wide there are **243 non-stack `lwz` reads from +356/+360**, spread across
the UI, sound, render and script subsystems — the offsets are far too common to
attribute. Inside the ISL region the only hits are the two writers in
`sub_82261F70` and two **vtable calls** (`0x82273400` slot 90, `0x82290ED8`).
Dropping this sub-item: it needs type information the offset index does not
carry.

### ✅ (2026-08-27) The two-function model is confirmed by a 64/65 vs 0/54 dissociation

Taking every literal key from `data/config-keys.txt` and testing it against the
3 496 distinct IDXD record keys on the disc (`tag_hash`):

| class | argument to | present as an IDXD **record key** |
|---|---|---|
| section names | `sub_82448AA0` | **64 / 65** |
| field keys | `sub_824482D0` | **0 / 54** |

The two classes are each other's control, and the split is total. It confirms the
model exactly:

* **`sub_82448AA0(node, NAME)` finds a RECORD**, keyed by `tag_hash(NAME)`.
* **`sub_824482D0(node, KEY)` reads a FIELD**, whose name is a literal string in
  the record's pool — so of course none of them is a record key.

🔑 **The single absent section is `LANGUAGE`** — which is precisely the one real
section in the disc-root `config.ini`. The same API serves both stores, and the
data shows the seam.

So the earlier worry that "64 sections have unknown hardcoded fallbacks" was
overblown: **all but one are real data on the disc**. `STAGENN_UNIT_MAX` is the
genuine exception, and it is absent because its name is *built at runtime* from
the stage index rather than being one of these literals.

📎 The artefact now records which pak each record lives in — `tables.pak` for the
global tables (`MISSIONS`, `STAGES`, `SOUNDS`, `SUB_OBJECTIVE`,
`SQUADRON_ORDER_OBJECT`, `EPILOGUE_MOVIES`, `ACHIEVEMENTS_REQUIREMENTS`,
`FONTS`, `BASE_INFO` …), **`GP_HANGAR_ARSENAL.pak` for `WEAPONS`, `UNITS`,
`Camera`, `ControlTweak`, `Rendering`, `AUTO_SETTINGS`, `IGNORE`**, and the
per-language `GP_MAIN_GAME_*` paks for the enum tables.

### ✅ (2026-08-27) The records DECODED — and my "canonical definitions" guess was half wrong

`parse()` + `named()` on each record (artefact `data/config-records.txt`):

**Real, directly portable constants** — these are the find:

```
Rendering      Brightness 1.4   Contrast 1.0   Saturation 1.00   Hue 0.00
               ExposureKeyValue 0.18   BrightPassThreshold 0.6
               BrightPassOffset 1.5    LuminanceMin 0.15  LuminanceMax 3.6
               BloomScale 0.5   AfterimageScale 0.05   StarScale 1.20
               GlareType 8      ColorLayer R/G/B/A 0.00
ControlTweak   CameraSpeed 0.50   TargetSpeed 0.50
               TargetMovingRange -10.0 … 10.0
               mov_trigger_play 20   eye_stick_play 6000   mov_stick_play 6000
Camera         NoseCameraFOV 0.92   ChangeTime 0.20
```

🔴 **But `WEAPONS` and `UNITS` are not what I called them.** Last iteration I
said `GP_HANGAR_ARSENAL.pak` holds "the canonical weapon/unit definitions". Half
right at best:

* **`WEAPONS`** is a **59-entry name roster** — `No_Equipment`,
  `Machiene_Cannon_MG1`, `Broad_Sword_SG1`, `Frail_GP37`, `Stiletto_BG1`,
  `Pilum_BP`… — an enumeration of ids, **not a stat sheet**.
* **`UNITS`** is not units at all: seven fields pairing a craft slot with a pilot
  — `Bird1-Sandra`, `Bird2-Billy`, `Bird3-Antonius`, `Bird4-Carl`,
  `Rhino1-Raymond`, `Rhino2-Katana`, `Rhino3-Ellen`. **The wingman roster.**
* **`IGNORE`** is a **13-name blocklist** over that roster —
  `Long_Spear_HBP`, `Twin_Saber_LG2H`, `Eagle_120AM`, `Maelstrom_Bomb`,
  `Cluster_Mine_B10`, and the placeholders `Wep_82`, `Wep_85`. Weapons that
  exist in the enumeration but are withheld.

🔑 **`AUTO_SETTINGS` lists exactly 28 files** — `stage01_settings.tbl` …
`stage29_settings.tbl` — a **third independent confirmation of the 28-stage
count**, after the `.ssb` census and the `UnitGroup_SNN.tbl` probe.

🟡 **And the UI records are UI, not gameplay.** `SUB_OBJECTIVE` is a HUD layout —
`MAX_OBJECTIVE_OFFSET 353,242`, `TEXT_OFFSET 85,271`, `NUM_QUALIFY_OFFSET
309,242`, `DOT_OFFSET 0,20` — with a **second variant** (`275,222` / `85,251` /
`231,222` / `0,0`) in one pak, so there are two layouts. `SQUADRON_ORDER_OBJECT`
is 29 `.prt` sprite names for the squadron-order menu. `MISSIONS` holds six
positional members — `TimeAttack`, `ScoreAttack`, `Extra01`…`Extra04`, the
**challenge** categories, not the campaign. `STAGES` has **zero** fields.

🟡 Its neighbours belong to the same cluster: `builtin103` reads
`[phase+10156]` and `[phase+10152]` (9 and 7 writers), and a sibling vtable stub
clears `[phase+10152]`. The shape is an engine→script status trio, but that is a
description, not a name.

## ✅ `builtin7`'s mechanism — and an independent confirmation

```
82266C84  lwz   r11, 324(r25)      ; the unit array
82266C88  lwz   r10, 4(r26)        ; local[4]
82266C94  lwzx  r11, r10, r11      ; -> the record
82266C98  lwz   r10, 4(r11)        ; the handle …  (alive?)
82266CA4  lwz   r11, 16(r11)       ; rec+16 = the unit STATE
82266CA8..CBC                       ; bail if state == 1, 3 or 4
82266CC0  lwz   r10, 12(r26)       ; local[12]
82266CC8  lwz   r11, 244(r25)      ; [phase+244] = SYMBOL TABLE 1
82266CD8  lwzx  r10, r10, r11      ; -> resolve local[12] as a symtab-1 index
82266CD0  lfd   f31, 25600(r9)     ; a double constant
82266CD4  lwz   r9,  16(r26)       ; local[16]
```

🔑 **`isl.py`'s `SYM1_SLOTS` already lists slot 12 for built-in 7**, derived
purely from operand ranges. Reading the implementation shows the *mechanism* —
`local[12]` is indexed into `[phase+244]`, which
[`isl-bytecode.md`](isl-bytecode.md) documents as symbol table 1 (routes,
messages, objectives). **Two independent methods, same conclusion.**

That makes the Stage 02 phase-2 condition

```
builtin7(TCT206, 1, Route_TCT206_p2S, 4294967295, 500) == 1
```

structurally coherent — a unit, a **route symbol**, and two numbers — but *what*
it asks about the route is not read, so it stays `builtin7`.

## 🟡 `builtin141` — only the entry read

```
8226C9D4  lwz  r11, 324(r29)   ; unit array
8226C9D8  lwz  r10, 4(r31)     ; local[4]
8226C9E8  lwz  r8, 4(r8)       ; handle
8226C9F0  bc   … 0x8226CA0C    ; if alive, continue
8226C9F4  addi r11, r0, 0
8226C9F8  stw  r11, 164(r29)   ; dead -> special[0] = 0
```

The same unit-array opening as every unit predicate, returning 0 when the unit is
gone. Everything past `0x8226CA0C` is unread. Stage 16 calls it twice with
arguments differing in one position (`0` vs `-4000`), which *looks* like a
coordinate — and looking like one is precisely the evidence this corpus does not
accept.

## 🟡 Not settled

* **No name for any of the three.** For 104 that needs the meaning of `r29` at
  `0x821AAD9C` and of the `(16, 32]` gate; for 7 and 141, their bodies past the
  entry checks.
* `builtin103`'s fields `[phase+10152]` / `[phase+10156]` have 9 and 7 writers,
  none read.
* This does not change any artefact — the conditions already printed
  `builtin104`, `builtin7`, `builtin141`, and still do.