# Live unit definitions from a running Stage 02 (2026-08-12)

The Route-B question — what value does a field that the disc leaves **defaulted**
actually take at runtime — needs the game running. This is the first snapshot of
the parsed definition objects taken straight out of guest RAM, plus the recipe
that works in the box today, because getting there was most of the work.

## The recipe that works

Everything must happen inside **one blocking foreground call**. A process that
outlives the call that started it is reaped — `nohup`, a harness background job
and a `wait`-holding wrapper were all tried and all died within a minute of the
turn ending (see the session-lifetime notes).

```bash
tools/re-capture/launch_mission.sh fly     # title → LOAD → slot 01 → READY ROOM → TAKE OFF
python3 tools/re-capture/gmem.py find hex:820af844 400   # live definition objects
python3 tools/re-capture/gmem.py words <va> 96           # dump each one
```

Two environment facts that cost an hour each:

- **Audio must be `--audio --apu=sdl` with `SDL_AUDIODRIVER=dummy`.** There is no
  PulseAudio server in the box, so muting (`--apu=nop`, `run-canary`'s default)
  looks like the safe choice. It is not: the log then fills with
  `AudioSystem::RegisterClient: CreateDriver failed for index=0`, the guest never
  gets past the movie, and the window stays black for 8+ minutes. `nav_probe.sh`
  and `launch_mission.sh` already pass the right pair — do not "fix" them.
- **Boot is fast once the caches are warm.** The first boot took minutes; with
  the shader/code cache populated, `launch_mission.sh` reached the flight HUD in
  **25 s**, which is what makes a launch-and-dump fit in a single call.

`entities2.py`, `flight_probe.py` and the other analysis scripts **do not run
here** — no `numpy` (and no `PIL` for image work). `gmem.py` is pure stdlib and
works, so the pattern is: dump raw words to disc inside the call, analyse
afterwards.

## What came out

14 live definition objects (vtable `0x820af844`), 96 words each:
[`captures/stage02-live-unit-definitions.txt`](captures/stage02-live-unit-definitions.txt),
with the flight screenshot they were taken from.

| object | `+0x30` | `+0x34` | `+0x38` | `+0x54` |
|---|---|---|---|---|
| `0xbd3b6a00` | 2500 | 3100 | 2600 | 10000 |
| `0xbd3d2d80` | 80 | 80 | 350 | 4000 |
| `0xbd3da100` | 10 | 7 | 29 | 1500 |
| `0xbd3db980` | 350 | 70 | 200 | 10000 |
| `0xbd3dbd00` | 10 | 5 | 20 | 300 |
| `0xbd3dea80` | 20 | 9 | 22 | 100 |
| `0xbd3dfc00` | 10 | 7 | 29 | 1000 |
| `0xbd3e0680` | 300 | 300 | 2800 | 3000 |
| `0xbd3e1f00` | 400 | 400 | 1400 | 25000 |
| `0xbd3e6f80` | 300 | 300 | 2100 | 10000 |
| `0xbd3ee300` | 200 | 200 | 2000 | 10000 |
| `0xbd3faa80` | 100 | 40 | 50 | 500 |
| `0xbd3fd800` | 600 | 600 | 3800 | 30000 |
| `0xbd40e200` | 700 | 700 | 3800 | 30000 |

This **corroborates** the layout solved earlier from a single in-mission
snapshot (`+0x054` = HP, `+0x30…0x38` = `Size_X/Y/Z`) rather than re-proving it:
the `+0x54` column reads as hull values across three orders of magnitude
(100 for something small, 30 000 twice for capital ships), and the sizes scale
with them. It also gives independent support to the **`Size_Y` inherits `Size_X`**
default rule — five of the fourteen have `+0x30 == +0x34` exactly
(300/300, 400/400, 200/200, 600/600, 700/700) while the rest differ.

## Identity without a name: match on the values

Nothing in the first 96 words is a name pointer, so identity comes from the
numbers themselves. `examples/unit_signatures.rs` prints `(HP, Size_X, Size_Y,
Size_Z)` for every craft unit and vessel on the disc; a live object is then the
record whose *known* fields it reproduces exactly.

Seven of the fourteen match a disc record on all four values — the player's
Delta Saber (`10, 7, 29`, HP 1500 in its player form and 1000 otherwise), the
ArrowHead, an ADAN turret, an Attacker. **The other seven match nothing, and
that is the point:** their disc records leave a field defaulted, so there is
nothing to match against. **18 of the 23 vessel records on disc are missing at
least one of these four fields**, nearly always `Size_Y`.

## ✅ Route B: 13 defaulted fields read out of the running game

Matching each disc record that defaults exactly one field against the live object
that reproduces its remaining fields resolves the missing value:

| unit | field | runtime value | live object (sx, sy, sz, hp) |
|---|---|---|---|
| `UN_e201_ADAN_ISCMissile` | `Size_Y` | **300** | `0xbd3e0680` (300, 300, 2800, 3000) |
| `UN_f106_TCAF_Destroyer` | `Size_Y` | **200** | `0xbd3ee300` (200, 200, 2000, 10000) |
| `UN_f106_TCAF_Destroyer_Inv` | `Size_Y` | **200** | `0xbd3ee300` |
| `UN_e105_ADAN_Cruiser` | `Size_Y` | **600** | `0xbd3fd800` (600, 600, 3800, 30000) |
| `UN_e105_ADAN_CruiserEX` | `Size_Y` | **600** | `0xbd3fd800` |
| `UN_f105_TCAF_Cruiser` | `Size_Y` | **700** | `0xbd40e200` (700, 700, 3800, 30000) |
| `UN_f105_TCAF_Cruiser_EX5` | `Size_Y` | **700** | `0xbd40e200` |
| `UN_f105_TCAF_Cruiser_Inv` | `Size_Y` | **700** | `0xbd40e200` |
| `UN_e108_ADAN_ASFrigate` | `Size_Y` | **80** | `0xbd3d2d80` (80, 80, 350, 4000) |
| `UN_e108_ADAN_ASFrigateEX` | `Size_Y` | **80** | `0xbd3d2d80` |
| `UN_e106_ADAN_Destroyer` | `Size_Y` | **300** | `0xbd3e6f80` (300, 300, 2100, 10000) |
| `UN_e106_ADAN_DestroyerEX` | `Size_Y` | **300** | `0xbd3e6f80` |
| `UN_f101_TCAF_Acropolis` | `Size_Y` | **400** | `0xbd3e1f00` (400, 400, 1400, 25000) |

**Every resolved value equals that unit's `Size_X`** — so the running game
confirms the `Size_Y` inherits `Size_X` rule that was derived statically from
sibling records, this time from the engine's own parsed definitions.

Two honest limits. Variants share a live object (`_Inv`, `_EX5`, `EX` map to the
same definition as their base), so the snapshot cannot distinguish them — the
value is right for the class, and per-variant differences in *other* fields are
not addressed. And this run only reaches the units Stage 02 instantiates: the
`…Ratio` / `…Count` family (`HQRatio`, `PowerRatio`, `ShieldRatio`, `HatchCount`,
`NodeCount`, `SequencingCount`, …) is defaulted on 5–23 records each and is
**not** resolved here, because those fields' offsets in the live object are not
yet known. Finding them is the next run: dump deeper than 96 words and look for
the constants the disc *does* set on the records that set them.

## Locating fields inside the live object

The next step — reading the `…Ratio` / `…Count` family — needs to know *where*
each field sits in the live object. The method is to anchor on a unit whose disc
record **sets** a field and look for that value in its dumped words. Anchoring on
`UN_f106_TCAF_Destroyer` (disc sets `Size_Radius 0.1`, `ThrusterRatio 0.8`,
`ResistanceParalyze 0.97`) and then checking the same offsets across five capital
ships and the player's fighter:

| offset | reading | `f106` | `f105` | `e105` | `e106` | `f101` | Delta Saber |
|---|---|---|---|---|---|---|---|
| `+0x30` | `Size_X` ✅ | 200 | 700 | 600 | 300 | 400 | 10 |
| `+0x34` | `Size_Y` ✅ | 200 | 700 | 600 | 300 | 400 | 7 |
| `+0x38` | `Size_Z` ✅ | 2000 | 3800 | 3800 | 2100 | 1400 | 29 |
| `+0x54` | `HP` ✅ | 10000 | 30000 | 30000 | 10000 | 25000 | 1500 |
| `+0x74` | 🟡 `ThrusterRatio` | 0.8 | 0.8 | 0.8 | 0.8 | 0.8 | 1 |
| `+0x84` | 🟡 `ResistanceParalyze` | 0.97 | 0.97 | 0.97 | 0.97 | 0.97 | 0.97 |
| `+0x40` | ❔ | 0.1 | 0.1 | 1 | 1 | 1 | 0.1 |

`+0x74` and `+0x84` are marked 🟡 because they rest on one anchoring unit: the
value is right for the Destroyer and constant across the others, which is
consistent with a field the rest default, but a second anchoring unit that sets
them to something *different* is what would prove it. `+0x40` is 0.1 on two units
and 1 on three, so it is a real per-unit field — just not identified.

⚠️ **The automated version finds almost nothing, and the reason is not what it
first looked like.** `examples/live_offsets.rs` correlates every `get_f32(key)`
against every dumped word and keeps offsets all units agree on. Over five units
it produced exactly **one** hit, `YawDragFactor → +0x0c`, and that hit is
**wrong**: `+0x0c` holds the integer 2 (printed as the denormal `2.8e-45`), which
collided with `YawDragFactor = 2.0`. Any float-compare against small integers has
that failure mode.

The first explanation written here — that `get_f32` is unreliable on
default-heavy records — was **wrong**, and measuring killed it: `get_f32`
resolves **46–57 numeric fields per unit** (Destroyer 46, `e105` Cruiser 57,
`f105` Cruiser 49, `e106` Destroyer 51, Acropolis 45). It is the documented,
reliable API and it works.

The real problem is that **the values are not in this object**. Dumping 512 words
(2 KB) instead of 96 changes nothing at all: still exactly **4 of ~50** disc
values found, for every one of the five units. So the structure at vtable
`0x820af844` is not the parsed definition record — it carries the handful of
fields above (sizes, HP, a couple of ratios) and the rest of the record lives
elsewhere, presumably behind a pointer.

Next probe, therefore, is not a bigger window: search guest RAM for a value that
is distinctive to one unit (e.g. the Acropolis' `HP 25000` = `0x46c35000`) and see
what *other* structure holds it, then map that one.

## Where the rest of the record is — three probes, and a hypothesis

Following "search RAM for a unit-distinctive value" produced no parsed record:

1. **`HP 25000`** (`0x46c35000`, the Acropolis) — **200+ hits**, the search limit.
   Too common to anchor on: it is also a generic constant.
2. **`AttackVesselPoint 0.08`** (`0x3da3d70a`, an unusual float) — 64+ hits, and
   **none of the eight windows dumped around them contained the unit's other
   values** (`Size_X 400`, `Size_Z 1400`, `HP 25000`, `RadarRange 60000`). One hit
   sits in the executable range (`0x820b62f4`), i.e. a code/rdata constant.
3. **The ID string itself.** `UN_f101_TCAF_Acropolis` appears **twice** in RAM,
   each copy referenced by a pointer exactly **16 bytes before the string**. But a
   256-word window around it contains **none** of the unit's numbers — because
   that region is the IDXD **string pool**, where `25000.0` is stored as *text*.

🟡 **Hypothesis: there is no flat parsed record.** IDXD is a reflective format,
and the evidence fits the engine reading values out of the pool by key when it
needs them, caching only the few it touches per frame. That is exactly what the
`0x820af844` object looks like: sizes (collision), HP (damage), a couple of
ratios — and nothing else, no matter how deep it is dumped.

If that is right, it reframes Route B. For a **defaulted** field there is no value
in the pool at all, so the default is applied by *code* at read time and can never
be found by scanning data structures. It would have to come from either the read
path (trace where a key hash is looked up and what the miss path substitutes) or
from the title's own code — which is the DuckDB static route this project already
has. The 13 `Size_Y` values recovered above worked precisely because `Size_Y` is
one of the cached, every-frame fields.

This is a hypothesis, not a finding: it explains three negative probes but has not
been confirmed by watching a read happen.

## The static route says where the defaults must be

The DuckDB image (`xenia-rs/sylpheed.db`, 12 156 functions / 1.87 M instructions)
has the reflective machinery under real names:

| address | name |
|---|---|
| `0x824486c0` | `IdxdLoad_Dispatch` |
| `0x82448d00` | `IdxdLoad_Variant2` |
| `0x82449640` | `Idxd_Parse` |
| `0x8244a2f0` | `Reflect_FindFieldIndex` |
| `0x8244a4b0` | `Reflect_SetField` |

`Reflect_FindFieldIndex` measures a C string, indexes a buffer and calls a
compare routine — it looks a field up **by name at runtime**, which is the
reflective read the hypothesis above predicted. But `Idxd_Parse` is a **text
parser** (it checks for a UTF-16 BOM, walks characters through a jump table at
`0x824497f8`) and it calls `Reflect_SetField` on a target object passed in `r3`.

That last point **weakens the "no parsed record" hypothesis**: parsing writes
values into a real object. So a parsed record does exist — my three RAM probes
simply did not find the right one — and, more usefully, it tells us where a
*defaulted* field's value comes from. `Idxd_Parse` only ever writes the fields the
text actually contains, so a defaulted field keeps **whatever the object had
before parsing** — i.e. the value its **constructor** stored.

So the defaults are neither in the disc data (by definition) nor in a table to be
scanned for: they are immediates in the constructor of each definition class.
Reading them is a disassembly job — find the allocation site above
`IdxdLoad_Dispatch`, then the constructor it calls, and read the stores. That is
the next step, and it would settle the whole `…Ratio` / `…Count` family at once
rather than one field per flight.

(The `vtables`, `classes` and `methods` tables are empty in this DB build, as
`CLAUDE.md` warns, so the class has to be reached through the call graph.)