02d3c9c82b
Disassembling the three Stage-02 entry_a targets shows opcodes 0x19 and 0x1A, and the
ISL dispatcher's table has 25 entries (cmplwi 0x18). They are not instructions. Each
phase region ENDS with a trailing data table of 8-byte typed records -- tag 0x19 = int,
tag 0x1A = IEEE float (0.0, 0.5, 1.0, 4.0) -- and entry_a is where it starts.
Confirmed across the disc: in 44 of 44 phases the first offset whose opcode exceeds
0x18 is exactly that phase's entry_a, with zero exceptions, and only two tags ever
appear (1394 x 0x19, 675 x 0x1A). So the record is
0x1883, base, size, 0, code_end, force_end_handler
one boundary and one entry, not two entries as the previous commit said.
That also retires this thread's own "82 of 88 land on a valid instruction = 93.2% vs a
38.6% control" as TOO WEAK a test: a data record has length 8 and passes "nonzero,
even". The entry_b result stands on different evidence -- those targets were matched
against isl.call_sites(), an independent enumeration.
isl.linear_offsets was decoding all 2069 data records as instructions, 1.23% of the
stream. Now each phase's walk stops at its boundary:
decoded instructions 168251 -> 166182 (= 168251 - 2069, as predicted)
opcode > 0x18 2069 -> 0
call sites covered 25705/25705 -> 25705/25705
exits unreachable 0 -> 0
conditions unknown 400 -> 400
Recorded because the first attempt at the fix was worse than the bug: it destroyed 36%
of the stream (168251 -> 107596, exits 0 -> 74) because linear_offsets is ONE global
walk from the first phase base, so stopping at phase 1's table lost every later phase.
It has to skip the region and resume at the next base. A count moving hard in the
wrong direction is the same signal as one that will not move.
Still open: the table's contents are undecoded -- its int values land on the
instruction stream 46/51 against a 29.5% chance rate, but 0 of them are unreached
run-starts, so this is not what starts the unreachable code either.
Runtime-capture harness (sylph-re container)
Screenshot-driven scripts for reading the running retail game's menus under Xenia
Canary + lavapipe, headless. They assume the container helpers screenshot,
vgamepad, pad are on $PATH and HOME=/sylph-home/re.
| Script | What it does |
|---|---|
skip_intro.sh |
Boot → main menu, unattended. Taps A only while the intro movie is actually playing (frame-to-frame RMSE), then once at the PRESS Ⓐ BUTTON title. Static logo screens are left alone, so a stray tap can never land on NEW GAME. |
wait_title.sh |
Older variant: wait for the title (green Ⓐ glyph at px 625,618) and tap A. Superseded by skip_intro.sh. |
step.sh |
One Arsenal navigation step (down/up/next/prev/none) + a compact capture: weapon list stacked over the DATA SHEET. |
sweep.sh |
Walk a whole weapon-type list, capturing only rows that show a DATA SHEET — locked rows (a "Conditions to Develop" panel) are detected by the brightness of the Range Class label box and skipped. |
type.sh |
Change weapon-type tab N times (RB) and report the header strip. |
hp.sh / cyc.sh |
Hangar hard-point carousel: cyc.sh steps it (d-pad down, not left/right) and captures the Name + DATA SHEET. |
Input timing under lavapipe: the game polls input at its own low frame rate, so a 60 ms d-pad tap is dropped roughly half the time. 200 ms is reliable; 300 ms starts to auto-repeat (two rows per press).
Findings produced with these: docs/re/weapon-datasheet-runtime.md.
Guest-memory tools (no screenshots)
| Script | What it does |
|---|---|
gmem.py |
Read the live guest address space out of /dev/shm/xenia_memory_* (Xenia's backing file), addressed by guest VA. find / read / words. |
weapon_runtime.py |
Solve the Weapon/Shell struct layouts against the disc records and read the fields the disc defaults. |
unit_discover.py |
Find which runtime class carries a set of ID strings, assuming no vtable: tallies the word at pointer_site - k across distinct IDs. |
unit_runtime.py |
Same solver for the unit\UN_*.tbl definition objects (vtable 0x820af844). Unions several snapshots — unit definitions are per-stage. |
schema_order.py |
Merge a sub-record's field-declaration order across all tables (topological sort); the layout check that pins fields no table ever values. |
order_check.py |
Test offset = base + 4*index for one table's sub-record against solver output. |
grab_tutorial.sh |
Cold-boot Canary, walk to the Nth TUTORIAL entry, wait for the stage load, snapshot guest RAM. One emulator per capture — backing out of a loaded mission wedges it. |
Snapshot first — cp --sparse=always /dev/shm/xenia_memory_* snap.bin (~2 s) — and
point $GMEM_FILE at the copy; the running emulator pegs every core under lavapipe.
🔴 pgrep -f / pkill -f match YOUR OWN shell
An agent driving this toolkit runs its commands through a wrapper shell whose command line contains the pattern being searched for. So
pgrep -f 'pilot\.py' # matches the wrapper running this very command
pkill -f 'fly_session|pilot\.py' # kills that wrapper — the script dies mid-way
This has cost four separate mistakes in one session: two scripts killed
mid-execution, and twice a "is it already running?" guard that answered yes
because it had found itself. The [p]ilot bracket trick does not help when
the literal invocation (python3 pilot.py …) also appears on the wrapper's
command line.
What works:
pgrep -x xenia_canary # exact NAME match, no -f
ps -eo pid,args | grep 'python3 pilot.py' | grep -v snapshot-bash
kill -9 <explicit pid> # look it up first, then kill by pid
The snapshot-bash filter is the reliable tell: the wrapper's command line
always contains the shell-snapshot path.
🔴 The emulator does not survive the end of an agent turn
/work/.claude/settings.json defines a Stop hook that kill -9s every
xenia_canary when a turn ends, printing "Stop hook killed N stale xenia
process(es)".
So:
- never launch a run intending to read it in a later turn — it will be dead;
- an experiment has to produce its evidence within the turn that starts it;
- prefer measurements that land in the first minutes of flight.
REMAINING OBsteps4 → 8 → 12inside the first few minutes, which is why a two-pass differential works in one turn while a 25-minute freeze watch does not.
This cost three iterations of investigating a "mysterious external SIGKILL", complete with cgroup and host memory forensics, before the hook was found. When a process dies at a session boundary, check the harness first.