Using last iteration's recipe (first title, one (A)) four more times, it failed
every time. The nine-run table splits perfectly on --log_ui_draws: 3 of 3
successes without it, 0 of 5 with it. That would indict my own instrumentation
cvar, which is the one that captures paint order.
It is also perfectly confounded and the entry says so rather than picking the
exciting reading: every success had its title inside 147 s, every failure at
177 s or later. Flag and boot-time move together across all nine runs. A
mechanism is hard to see for the flag (it is read only when F10 arms a capture;
the per-draw hook is one relaxed atomic load), which argues for the boot-time
story — but "I cannot see how" is not evidence.
The separating experiment is named: ON runs with a fast boot. One that accepts
(A) exonerates the flag.
Also recorded: widening the menu signature last commit made screen_id classify
some boot frames as "menu", twice in one run before any menu existed. A dark-blue
movie frame and this menu are not separable from a single grab — navigation
scripts need the classification to hold across two grabs, which is a caller-side
fix, not a classifier one.
screen_id.py called the main menu "other". Its menu rule required a near-white
fraction above 1.5%, measured in 2026-07; the menu reached from the boot title
measures 0.03% (mean 13,26,59 — dark, strongly blue, essentially green-free).
That is worse than a cosmetic miss: a script that waits for "menu" and never
sees it reports the navigation as failed while the menu is on screen, which is
exactly what happened here. Both measured signatures are now documented in the
code and both classify.
menu_draw_capture.sh now taps ONCE on the first title rather than up to 40 times:
repeating was measured to be useless (the attract title accepts nothing) and the
first title accepts a single press.
title_states_capture.sh is new — it captures the draw list in both title states
in one run, which is what refuted the "the attract title omits the button plate"
theory.
The previous entry named the next step: "tap (A) once per second and see whether
the success rate goes to 1". It was run. The answer is no.
40 presses at 1/s on the title change nothing, and neither does one each of
START, B, BACK, X and Y. Every press was delivered (the driver logs down/up for
each) and the new [RE-INPUT] diagnostic never fired, so nothing was swallowed —
the guest received them and ignored them.
Also recorded, because it destroys a run rather than merely wasting it: tapping
(A) through the boot (88 presses over the intro) ends on a permanent black
screen with no crash and no throw.
What is left is a difference in STATE, not in input: the one success came on a
title that appeared ~83 s into a warm boot, the failures on titles that appeared
after a full attract cycle. "The attract-loop title is a non-interactive
presentation" is written down as a hypothesis with the experiment that would
settle it — a log_ui_draws capture in each state, looking for ptbtn00.
The previous entry called the title's Ⓐ a hard blocker. It is not: the main menu
is reached and screenshotted (NEW GAME / LOAD GAME / TUTORIAL / OPTIONS /
EXTRAS). What is true is narrower — Ⓐ advances the title only intermittently,
about one attempt in four, with the press verifiably delivered every time and no
Xenia UI active.
Three candidate causes were eliminated with measurements rather than argument:
* IsUIActive is now observable (Canary logs when it swallows a keystroke) and it
never fires on the failing runs;
* the driver filter is fine — the game polls with flags=3 and the file pad
reports Controller=1, so FilterDrivers keeps it;
* the game makes no content/user/signin call on the press at all — tracing every
Xam call around it shows only input polling.
And two traps in my own measuring rig, which cost more than the bug and are
written down so nobody repeats them: a FIFO trace consumer that exits STALLS the
emulator (the guest stops polling — indistinguishable from a dead pad, and it
produced two runs of false evidence), and phase-A's kernel.return events carry a
placeholder return_value of literally 0, so "every keystroke call returns
SUCCESS" was an artifact of the logger, not a finding.
Chasing the title-screen crash into the emulator's own diagnostics turned it
from a mystery into a controlled experiment, and knocked over two things this
corpus said yesterday.
Withdrawn #1: "the fault address 0x1_0000000C is a pointer with a stale high
word". The crash dump prints r25 = 0x0000000C, clean. Xenia maps the guest's
4 GiB at host 0x1_00000000, so that IS guest address 12. The guest dereferenced
the small integer 12.
Withdrawn #2: "with --mem_watch=false the crash does not happen at all", which
named the crash-oracle handoff's suspect #1 as measured. It was confounded —
every --mem_watch=false run had also had a warm cache. Held cold, the throw
happens with the probe off (2 437 crash dumps). mem_watch is eliminated for this
crash.
What it actually is: the access violation is the guest's own `throw` RETURNING,
because this build logs guest C++ exceptions and continues rather than
unwinding. So the event is the throw, and with --cache_throw_diag=true the guest
names it: std::out_of_range, from the cache-manager flush, with a deque of 38
entries (37 distinct, one duplicated) against a 37-key map, every one of them
absent from the flush's snapshot but present in the live map — the TOCTOU race
the logger's own message describes.
And the new, useful part: the trigger is the on-disc cache. Complete cache, no
throw (2 runs). Directory moved aside or half-rebuilt, throw ~100 s into the boot
(3 runs, including one that threw with NO access violation behind it — which is
why crash dialogs are the wrong thing to count). A suspect in that bisection plan
now costs a `mv` and two minutes instead of a mission.
Two of the paint-order refutations were computed with build 7's keyframe times,
and build 7 is not the build the game runs. Re-checked on build 4: both still
fail, on the same element (ptlogo1 paints after ptlogo_back2 though it starts at
t=26 against t=66, and rests at t=42 against t=80).
A new candidate is recorded and refuted rather than left implicit, because it is
the kind that gets adopted on partial agreement: painter's order by resting Y
reproduces the captured order to within a single transposition — and is wrong
twice, on ptlogo_tm (drawn before a sprite 63 px higher) and on the background
(drawn first, though its resting Y would sort it fourth).
Also scoped, not walked: the guest-code avenue. The splash item vtable
0x820b30b4 is real (25 slots, three construction sites); RTTI carries no class
names anywhere in the binary; the format tags are fourcc immediates behind a
virtual call, not strings; and the obvious shortcut — searching for the 60-byte
declaration stride — leads to a time-string parser, not the table.
Each of these presents as "the pad is dead", and each has a different cause.
1. **F10 opens the emulator's menu bar.** It is the RE capture hotkey AND the
toolkit's menu key. With that menu open — or any Xenia UI window, the
Profiles Menu included — `IsUIActive()` is true, so XamInputGetKeystrokeEx
returns SUCCESS with a ZEROED keystroke before the pad driver is asked. The
guest polls, gets nothing, and does nothing, with no error anywhere. Escape
does not close it; a click on the game surface does. This cost most of an
iteration.
2. **The file-pad was never the problem.** It delivers (vk=5800 down/up per tap)
and A skips movies in the same runs.
3. **(A) at the title needs a signed-in profile.** Without one the guest calls
XamShowSigninUI and Xenia's dialog appears — which is also the proof the
guest reads the button. That dialog cannot be completed here: the gamertag is
an ImGui text field and synthetic X key events never reach it. Hence Canary's
new --create_profile_if_none.
And the blocker, stated as measured rather than as a theory: with the profile
signed in, (A) still does not advance the title across four boots. The
reproducible failure is next door — with the default --mem_watch=true the guest
crashes in the boot content path (read of 0x10000000C at 0x82307128, thread 9,
after a guest C++ throw), and with --mem_watch=false it does not. That is a
second, far cheaper reproduction of the Ready-Room crash the canary handoff
blames on the mem_watch probe: 40 s from a cold boot instead of after a mission.
One observation kept because it did not reproduce: a single (A) once crashed the
guest at a DIFFERENT PC (0x824578A0, thread 6). Two attempts to reproduce it
gave neither a crash nor a transition.
Next step named: those two PCs in xenia-rs/sylpheed.db, and the content exports
around the press.
BACKLOG's UI-ordering item ended at "the next step is no longer static — it is a
per-draw capture of the title screen showing the order the game submits". That
capture now exists.
The order, in submission order: a full-screen background layer, a rotated effect
pair, a second full-screen layer, ptlogo_back2eff + ptlogo_back2, ptlogo1 +
ptlogo_tm, ptlogo2, ptcopyright, and the PRESS (A) BUTTON plate — declaration
indices 13, 22|24, 23, 0, 11, 1, 28, and then two elements that are not in that
build at all.
Two more candidate orderings die on it (keyframe start time, resting-keyframe
time), and one structural fact reframes the item: the visible screen composites
TWO bundles — build 7 plus the one-element build 2 that is the button — so no
single build's element table can be the paint order whatever its order. The
button's quad lands at (384,551) against a declared rest of (383,550), which is
a one-pixel agreement between the static placement decode and the running game
on a bundle nothing had checked.
INDEX's "the screen's draw list is the bundle's declaration table (back-to-front)"
is demoted in place rather than edited away.
Still open and said so: the rule that produces the order; three of twelve quads
unidentified; one screen only.
The paint-order item's remaining step is a per-draw capture, and the obvious
move is to reuse the RE instrumentation already compiled into the Canary build
on `sylpheed-re`. Read the source rather than assume: neither hook works here.
`--log_draws` de-dups by vertex-declaration fingerprint, so a screen's sprites
collapse to one record; it carries no texture identity and only first-seen
order. The F10 ship capture does keep submission order and would separate the
elements, but it returns early on any draw without an `f32x3` position stream —
the comment says "UI/effects — skip" in as many words.
So the item needs a new cvar-gated hook and a `build-canary` run. Stated in the
entry so the cost is known before someone starts, and so this reading is not
repeated.
Records both halves rather than just the fix: the "extra static placements
cannot fail it" line had been stale since `64d372c`, and the residual — a set of
names being blind to a duplicated instance — is what this round actually closed.
Also states what it does not close: the multiset is `e106`'s alone, so the
per-ship table this entry originally asked for still needs a runtime capture of
a second capital ship, i.e. the `capture-ship-placement` build of Canary.
`BACKLOG.md`'s "the declaration table is not a paint order on every screen" had
been reasoned about but never measured. It is now measured, and the entry says
so both ways.
Confirmed: a framebuffer capture of Canary on the title screen shows the
`PROJECT SYLPHEED` wordmarks over the full-screen background, so declaration
order is not the paint order there.
Refuted, and recorded rather than quietly dropped:
* the `0x10`-adjacency first step this entry proposed. The background *is* next
to a `kind = 0x10` `.prm` on both screens that have one — but on opposite
sides (`GP_TITLE` puts the background after it, `GP_MISSION_SELECT` before),
so no rule keyed on the `.prm` orders both. `GP_READY_ROOM` and `GP_OPTIONS`
cannot serve as the third and fourth witnesses: neither carries a `.prm` or a
full-screen background at all.
* the placement-region group order and the RATC child order, as above.
* reverse declaration order, by the same capture — it would draw the background
over the copyright line, which is visible.
And separated out: part of what the bad render showed was not ordering at all
but the pivot/scale defect fixed two commits back. `ui-rat-layout.md` gains that
rule, the correlation evidence, the unmeasured pivot-vs-centre gap, and the
disc-wide demotion of "pivot is exactly half the texture" (49 % agree, 37 % are
off by more than 16 px — `ptcopyright` is 694 px wide in the framebuffer and
`pivot*2` says 618).
Captures: `title-screen-oracle.png` (the game surface cropped out of the
emulator window at (1,45), 1:1, bottom 45 rows clipped by the display) and
`title-screen-composite-2026-08-18.png` beside it.
Building the Explorer's UI Screens browser turned up a limit on a claim the
format doc states flatly. `ui-rat-layout.md` says the element declaration table
is the screen's back-to-front draw list, verified 11/11 on the tutorial pause
bundle -- and that reproduces exactly here. `GP_TITLE.pak` build 7 does not:
painting in declaration order puts the full-screen background art (element 13)
over the wordmarks (elements 0-5).
Ruled out rather than guessed at: there is no depth key in the 60-byte entry.
Across the title build's 30 entries every unknown word is constant, and `+36`
-- the one that varies -- is an instance index on the `kind = 0x4` repeated
entries, not a layer.
Recorded as an open item with the cheapest next step (composite two more
screens that have both a background and overlapping foreground, and check
whether their background sits adjacent to a `kind = 0x10` PRMD entry as the
title's does). The viewer paints in declaration order and does not paper over
it, so a screen whose background lands on top is showing this bug rather than
a decode failure.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
INDEX's XBG7 row still said 5480 decode (87.1%) and 125 of 681 shared resources
failing; it is now 6209 of 6294 (98.7%) with zero real meshes disagreeing across
containers and 46/46 capture-verified drawn buffers. The capital-ship row blamed
an XBG7 mis-decode for the user-reported assembly bug -- that mis-decode (a shared
turret ~100x too large) is fixed, so the BACKLOG entry is marked resolved at the
format layer, keeping the viewer-side pointers and the note that no metric caught
it: it took rendering the ship and looking.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The neighbourhood anchor (f18d591) and its refinement (27a0701) took
cross-container inconsistency from 125 to 51 with coverage unchanged, and made
e106 render as a destroyer rather than a slab. Both are reverted.
ship::tests::static_assembly_matches_runtime_capture is gated on SYLPHEED_ISO, so
it SKIPS in an ordinary cargo test -- which is why the regression was invisible
in every suite run so far. With the ISO it fails:
e106_bdy_01: static M row0 [-1.0, 0.0, 0.0] != captured [1.0, 0.0, 0.0]
e106_bdy_01 and _02 are a mirrored pair whose two buffers hold the same geometry
reflected in X, and BOTH resources currently decode to the SAME buffer (identical
counts, span and mean_x). apply_twin_mirrors picks which instance to reflect from
the sign of that mean_x, so which buffer wins flips the decision:
before both twins mean_x = -66.83 -> mirror bdy_02 (matches the capture)
after both twins mean_x = +66.83 -> mirror bdy_01 (contradicts it)
Neither is right -- two resources sharing one decode is itself the bug and the
mirror heuristic has been compensating. The capture is ground truth, so a change
that contradicts it does not ship. The real fix must give each twin its own
buffer first.
Kept from the attempt: this test now also asserts the SET of static placements
against the capture (allow-list {e303_wep_01} for vbase dedup), so extra
placements can finally fail it -- the direction it could never fail in before.
Docs, backlog, INDEX and the ignored test's message all corrected to say
diagnosed-not-fixed rather than fixed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
anchor_pool_mesh took the FIRST candidate in file order from a container-global
vertex-run scan, so a resource could be handed another resource's block whenever
both shared (stride, vertex count, index count). Both blocks are real geometry and
both pass every quality gate, so only position separates them.
anchor_pool_mesh_near now tries candidates in order of distance from a reference,
and anchor_models_filtered runs two passes: pass 1 anchors first-match to learn
where resources land, pass 2 re-anchors each resource preferring the median anchor
of its +/-2 descriptor neighbours. Too few anchored neighbours -> keep pass 1, so
nothing regresses to guesswork.
before decoded 5480/6294 shared 681 inconsistent 125
after decoded 5480/6294 shared 681 inconsistent 63
Coverage unchanged, inconsistency halved. e303_wep_01 decodes to 49x23x42 in ALL
containers now, and e106 renders as a destroyer instead of a slab -- its two
shared turrets symmetric at X[-203,-154] and X[154,203]. That resolves the
user-reported "capital ships assemble wrong" for this cause.
The filtered path needed care: models_named (what the viewer uses) dropped
non-wanted resources, which would have left filtered decodes with no
neighbourhood and silently kept the old behaviour. Resources are now collected
regardless of the filter, but only the asked-for ones and their +/-2 neighbours
are decoded in pass 1, so a filtered decode stays proportional to what was asked.
63 cases remain; mesh_consistency_disc stays ignored and now records 63, not 125.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
One iteration ago I concluded assemble_ship was drawing a non-renderable
collision volume. That named the wrong cause, and decoding the same resource from
every container that holds it settles it:
Stage_S01 172 verts 110 tris 49 x 23 x 42 <- a turret, correct
Stage_S02 172 verts 110 tris 1600 x 2100 x 4800 <- wrong
Stage_S08 same 1600 x 2100 x 4800 <- wrong
Stage_S26 same 1600 x 2100 x 4800 <- wrong
eleven others 49 x 23 x 42 <- correct
Same resource, same vertex and triangle counts, correct in eleven containers and
wrong in three. So the placement is legitimate (e303_wep_01 is a small shared
turret cross-mounted on e101/e106), the original author's vbase-dedup explanation
of the capture's silence stands, and my "dedup would show one, not zero"
objection does not survive -- at its true size the turret is ordinary geometry.
The defect is in the mesh decoder.
The wider point: the decoder can produce wrong geometry WITHOUT declining. The
XBG7 audit counted 814 honest refusals; this is the other kind, silently 100x too
large. A screen for the signature (bounds exact multiples of 50, span > 1000)
flags 22-32 models each in S02/S03/S08/S26/S27, but it also catches legitimate
e_rou_* composite proxies, so that is a candidate list and not a bug count.
Next: diff the anchor scan's chosen vb0 for this resource between Stage_S01 and
Stage_S02 -- same resource, two outcomes -- and turn whatever distinguishes them
into a post-decode sanity check so a silent mis-decode becomes a decline.
Kept from the previous entry: the test can only fail one way, and cross-id
placement is genuinely narrow (2 pairs across 335 ships).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Did the visual the backlog kept naming as the next step, and it overturns the
conclusion. Rendering e106 from the static assembly and from the baked runtime
capture: the capture has 8 placements, the static assembly 11 -- the same 8 plus
e303_wep_01 x2 and a second e106_eng_01. In the render the destroyer sits inside
a white slab that dwarfs it.
That slab is e303_wep_01: 172 verts, 110 tris, bounds X[-1000,600] Y[-1050,1050]
Z[-2400,2400] = 1600 x 2100 x 4800. A real e106 turret (wep_02_01) is 269x179x417
with 772 tris; the bridge is 105x76x305. So it is not a turret -- it reads as a
collision/trigger volume, and whatever it is the game does not draw it: the e106
capture contains no e303_wep_01 at all.
Why it was missed: assemble_ship treats every rou_* composite node as drawable,
the cross-id mount is documented as intended, and the test asserts
count("e303_wep_01") == 2. The capture's silence was explained as vbase dedup,
but dedup would show one instance, not zero -- and the test walks capture parts
looking them up in the static output, so EXTRA static placements cannot fail it.
Same shape as the earlier include_external gap: a test that can only fail one way.
Scope stated carefully: sweeping 335 ships for "ship-scale span with <400 tris"
flags 20 ships and 58 placements over 28 resources, but ONLY the e106 case is
proven (render + capture absence + geometry). Others may be legitimately large
low-poly parts and need the same three checks.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Mirror handling: the concern was that apply() bakes R*(S*v)+T while the viewer
decides winding from det(m) alone and rotates normals by m alone, so a mirror
encoded as negative SCALE would reflect geometry without flipping winding. It
never happens -- across 1485 assembled parts in all 22 containers there are 22
mirrored parts, every one with det(m) < 0, and ZERO negative or non-uniform
scales. apply_twin_mirrors writes the reflection into m, so the flip always
fires and ignoring s for normals is harmless.
Exhaust cones: the only geometry the viewer invents, drawn at GN_Jet/GN_SJet
frames because the real engine geometry is recessed. Across 335 assembled ships,
192 with exhaust frames, not one cone sits outside its hull bounding box (10% of
span tolerance). Caveat kept in the doc: inside-the-box does not prove a cone is
correctly oriented or sized -- it rules out the reported symptom, no more.
Every mechanism the diagnosis proposed is now eliminated (include_external,
duplicate names, mirrors, cone placement). The visual comparison is no longer
the next step but the only remaining one, and if it agrees the original report
needs re-grounding against a specific ship and expectation.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
build_ship_model resolves each placement with find(|m| m.name == p.resource) --
first match wins -- so a repeated resource name inside a container would draw
the wrong geometry. It cannot happen: decoding every XBG7 resource in all 22
stage containers gives 4603 resources and ZERO repeated names. Per ship it is
tighter still: e106 wants 9 distinct names and decodes exactly 9 models for 11
placements, e105 9 for 9, f105 5 for 6.
Two of the three remaining candidates for the user-reported viewer bug are now
gone (this and include_external), leaving mirror handling and the exhaust cones,
plus the untried visual comparison.
The container sweep also reconciles a number I flagged earlier: the disc has 22
stage CONTAINERS -- Stage_S01..S16 plus Stage_S24..S29 -- i.e. 16 story stages
plus 6 extra. The save screen's 16 STAGE sprites and the notes' "22 stages" are
both right and count different things; neither gets edited to match the other.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Stage-02 outcome write-up ended by naming "turrets near the asset must
become targets" as the fix worth ~50% of the escort damage. That was an
inference from a co-presence attribution, not a measurement. Run it and it
does not hold.
Also corrects the run labelling: pilot.py gained the SYLPH_KILL_TURRETS gate
two minutes before mission02 started, so mission02 was already a treatment run,
not a second baseline. Only mission01 (0 of 3968 pilot frames targeting a
turret) is the baseline.
At a common t=428s: baseline 46.9% escort hull, treatment 44.5% and 53.0%. The
two runs of the same arm differ by more than either differs from the baseline,
and the escort still reaches zero at t~590-670s in all three. So the
transferable finding is the power limit: one 430s flight cannot resolve an
effect below ~9 percentage points, and every single-run pilot conclusion,
including this one's, is inside it.
What does reproduce: the assault is scripted (onset 166/167/166s), and the
e007/e010 damage split is 50/50 in all three arms including the one that never
fires at a turret -- so that attribution measures the wave script, not us.
Also records that the viewer's include_external hypothesis in BACKLOG is dead
(it defaults true and is threaded through unchanged).
Tier 3 matched a part to its hardpoint by trailing index, so `e105_brg`
compared "01" == "" against GN_Bridge_01 and fell through silently. The runtime
capture is what exposed it: the game draws the bridge and places it at
[0, 70, -1850] rel e105_bdy_01, and assemble_ship emitted nothing there.
With no index to match on, take the lowest-numbered frame of the category.
Diffing assemble_ship part counts across every container: 34 (stage, ship)
entries gain parts — e102 +2 (bridge and engine), e104 +1, e105 +1, Stages
02-29. ship_audit is unchanged, so nothing regressed, and the capture now agrees
to dT 0.03 / dR 0.000.
Also fixes the diff itself: correlate_frames compared static against a rotation
sampled from the first block, which can belong to another INSTANCE of the class.
Scoped to the position-agreeing cluster, e105_eng_01 goes 1.711 -> 0.000 and
both e106 nacelles to 0.000. The one remaining rotation delta (e106_wep_02_01,
0.134) is a turret whose rotation varies by 0.182 between blocks that agree on
its position — the runtime disagrees with itself more than with the assembler.
The new rotVar column makes that distinction visible.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A capture at controlled range (ship_capture_close.sh: lock a capital ship, close
on it, F10 per range band) finally draws capital-ship hulls at full detail. Two
correctness fixes were needed before the numbers meant anything:
* one F10 log is ~14 frames with no delimiter, and WV_ref^-1 . WV_p only cancels
the camera within one frame — segment_frames splits on vertex-buffer
recurrence, and correlate_frames cross-checks the blocks against each other
instead of trusting a single shot;
* aggregate by consensus, not median: a stage holds several ships of one class
sharing vertex buffers, so a block can mix two instances.
Result: f105, e105 and e106 reproduce assemble_ship to <=0.43 units in
translation and 0.000 in rotation for every part that does not move. The e106
rules generalise, and the viewer bug report now points at the viewer. Narrow
leftovers: e105_brg is missing from assemble_ship, e105_eng_01 rotation differs
by 1.711.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The runtime-capture write-up declares static ship assembly exact, but its test
covers exactly one ship (e106). Records what is suspect (engine-cluster rig,
the X-reflect twin heuristic, cross-id turrets), and that the F10 capture is
already the oracle to settle it on a second class.
