91b09955c89f500b5ba8596b7bf11121c11074cf
111
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 ReaperOfVeriod
|
36d8d2b00c |
The frequency enrichment survives the redo; the disassembly does not
Redone through overlay_bytes_at, the verified flat mapping:
functions kept 2,313
moby_classes 0x00c0 391 (displaced: 394)
best neighbour (0x00c8) 15 (displaced: 4)
bucket n 0xc0 best neighbour ratio (displaced)
0-15 1,033 73 2 36.5x 72.0x
16-31 759 118 4 29.5x 29.8x
32-47 283 89 2 44.5x 34.7x
48-63 118 41 3 13.7x 19.5x
64-79 55 32 2 16.0x 15.5x
80-95 37 22 2 11.0x 13.0x
The reason it survives is worth stating, because it separates the two results. The
displacement was uniform -- every window shifted by the same amount -- so a
function's contents barely changed, and a frequency count over a window is robust
to that. A disassembly is not: it depends on the exact byte alignment, which is
what was wrong. Hence 394 becoming 391 with the same shape, while the 0x3a8060
walk and the 16-byte stride evaporated entirely.
So the investigation was not "all wrong". Precisely:
- the frequency enrichment is real, now on verified bytes, with a test that fails
if it goes;
- the disassembly was never real, and the whole jal chain built on it is retracted;
- no loader has been located, and 0x2ee138 is VU microcode.
THE LESSON IS ABOUT WHICH MEASUREMENTS SURVIVE A MISTAKE
A count over a window tolerates a constant displacement. A decode does not. Sixteen
probes leaned on the kind that does not tolerate it, and reported each with the
same confidence as the kind that does. Nothing about the numbers distinguished them
at the time -- that is the whole difficulty, and why the verification in
|
||
|
ReaperOfVeriod
|
cd1648571c |
The verified mapping is the flat memory image, not section indexing
The fix |
||
|
ReaperOfVeriod
|
a2f8c924b4 |
moby_classes.offset enrichment is size-independent: 13x to 72x at matched size
Runs the control |
||
|
ReaperOfVeriod
|
3ea5ab799f |
moby_classes.offset is referenced by 394 functions; its neighbours by 1 to 14
First positive result of the loader hunt, and it exists only because of the
control that two commits ago I said was missing.
Across 2,323 level-0 functions of at least 128 bytes -- extents from the decomp's
functions.tsv -- the immediate 0x00c0, which is moby_classes.offset, appears in
394 functions. Every value adjacent to it appears in 1 to 14.
0xb8:8 0xb9:2 0xba:2 0xbb:2 0xbc:3 0xbd:5 0xbe:5 0xbf:3
0xc1:3 0xc2:3 0xc3:2 0xc4:1 0xc5:2 0xc6:3 0xc7:1 0xc8:14
This is what
|
||
|
ReaperOfVeriod
|
5a60a2920c |
Remove two dead counters; clippy clean again
Follow-up to the note in
|
||
|
ReaperOfVeriod
|
6d26da3630 |
The control fails my own claim: 0x1c/0x34 select nothing distinctive
|
||
|
ReaperOfVeriod
|
2ff2abbd32 |
Two class-header words are base-relative offsets pointing past the tie packets
Chasing where the AD-GIF is, since |
||
|
ReaperOfVeriod
|
ff9699c3c7 |
The tie class header's undecoded tail: two reserved words, a float, and a control that says no
TieClassHeader skips 0x1c..0x44 -- forty bytes per class, 1,864 classes, "none of
it matches RAC2 yet". Profiled each of the ten words instead of reading them.
THE PROFILE, PER WORD
0x20 reserved -- zero on 1,864 of 1,864
0x30 reserved -- zero on 1,864 of 1,864
0x1c 1,390 distinct, 0x1e0..0x3cd60 looks like an offset
0x24 1,571 distinct, high entropy
0x28 640 distinct, 0x00100000..0x1000013d two bytes packed
0x2c 1,610 distinct, high entropy
0x34 1,331 distinct, 0x190..0x31e20 looks like an offset
0x38 441 distinct
0x3c SIX distinct: 0 (1,384), 16 (435), 48 (21), 32 (18), 2 (3), 4 (3)
0x40 a float -- 0.00208 to 10.84, median 0.30, no negatives
0x40 is decidable and it is now a hard assertion: a positive range across two
orders of magnitude with no negatives and no zeros is not what an integer looks
like. Bounding sphere radius or class scale; not distinguished, and the doc says
so rather than picking one.
THE CONTROL THAT PREVENTED A WRONG CLAIM
0x3c has six values, is zero on 74% of classes, and sits next to a doc comment
listing "the texture count, ad-gifs, the bounding sphere and the class scale". The
obvious next move is to call it the ad-gif count -- especially having just spent a
commit establishing that the AD-GIF is not in the packet, so an ad-gif count
would be satisfying.
Tested it: mean bfc_distance per 0x3c bucket is 7.80, 6.56, 11.07, 8.03, 6.81,
6.11. Flat. It does not track the shader count, so it is not the ad-gif count.
The test now asserts that flatness as a control, so if 0x3c ever starts
correlating with bfc_distance the assertion fires rather than the coincidence
being noticed later and mistaken for a decoding.
That is the rule-4 pattern working: a statistic that cannot tell the hypothesis
apart from nothing must not be used to support it. Six distinct values and a
dominant zero *feel* like a count, and feel is exactly what is not evidence.
STILL UNIDENTIFIED
0x1c and 0x34 look like offsets but are unverified as such -- only 17 of 1,864
0x24 values land in range for core_data, so that one is not an offset and is
recorded as packed data. 0x24, 0x28, 0x2c, 0x38 are high-entropy and unexplained.
None of them is guessed at.
280 passing with the disc present (23 + 19 + 60 + 156 + 22; disc binary 35.38s),
clippy clean. The two unused-variable warnings in the sound_bank and overlay tests
are pre-existing and confirmed present before this change; left alone rather than
folded into an unrelated commit.
|
||
|
ReaperOfVeriod
|
f43a1c2e6e |
bfc_distance is not the control region's length; the AD-GIF is not in the packet
Follows
|
||
|
ReaperOfVeriod
|
9bb04cca7b |
The tie control region is not a VIF or GIF tag stream; and vert_ofs is mis-named
Went after the GS's GIF packets in the tie AD-GIF region. Got a negative result,
a new universal invariant, and a retraction of my own first probe, which reached
the right answer for the wrong reason twice.
THE NEGATIVE RESULT
Tried the leading `5 * shader_count` quadwords of each packet (the
`bfc_distance` field's stated meaning) as VIF tags -- command in bits 0-6, size
in 7-14, mark in 15 -- and as GIF tags -- type in the top byte of word 1, length
in bits 0-2 of word 0. Both uniform. Over 13,042 records: 10,443 distinct values
of word 0, no GIF tag type above a count of 2. That is the control result rule 4
asks for, and it says stop.
THE RETRACTION
The first probe scanned `bfc_distance` quadwords, which for most records
overruns `vert_ofs` and runs into the vertex data. Pooling a control region with
vertex positions -- two populations with different structure -- gives a uniform
answer, so "not a tag stream" came out right from a measurement that was
measuring the wrong bytes. That is exactly the rule 3a failure, and I only caught
it because the hex dump I took next looked structured while the histogram said
random.
Re-measuring quadword 0 alone, on its own population, gives the same uniformity.
The conclusion survives; the reasoning did not, and both are recorded.
THE PARTITION, AND A NEW INVARIANT
The hex dump showed the packet layout directly:
q[0 .. vert_ofs) control 3..8 quadwords, peak at 5
q[vert_ofs .. +vert_size) vertices
q[vert_ofs+vert_size ..] scissor
and `scissor_ofs == vert_ofs + vert_size` holds on **15,219 of 15,219** records
across all 27 levels, including every one whose strip_count is non-zero. That is a
derived quantity stored explicitly, and it holds everywhere -- which is the check
that confirms the three regions really are laid out the way the existing packet
length formula assumed. It is now a hard check in invariants_hold(), which is
something the previous four could not be.
The fixture in tie.rs failed the moment I added it: it had vert_ofs 4 and
vert_size 0x20 with scissor_ofs 0x30, which is 4 + 0x24. The synthetic data was
never constrained by a relation that did not exist yet. Fixed to 0x24.
vert_ofs IS MIS-NAMED
It is not a vertex offset. It is the *length* of the leading control region, and
the vertex data begins at that many quadwords into the packet. The field name
stays because the on-disc order is fixed and everything already written about this
record says vert_ofs, but the doc now states what the value is and gives the
distribution.
STILL UNKNOWN
What the control region contains. Not standard DMA tags; possibly pre-merged GS
register writes in another framing, or per-class state rather than per-primitive
draw data. Recorded as an open question rather than guessed at.
23 + 19 + 58 + 156 + 22 = 278 passing with the disc present (disc binary 35.67s),
clippy clean.
|
||
|
ReaperOfVeriod
|
f0bc1fbdf4 |
State which GS rules are wired and which are not
A small doc change with a specific reason.
gc-engine's sampler sat behind a comment claiming it clamped while the code was
`..default()`, which is Repeat. That was caught only because I went looking for
where clamp01 should be used and found the answer was "somewhere that already
claims to do it". The generalisable failure is a reader seeing a rule defined in a
module called gs.rs and assuming the renderer applies it.
So the module now says which rules are wired and which are not:
CLAMP wrap yes gc-engine image_of sets address_mode_u/v
alpha 0x80 opaque yes texture::scale_alpha delegates here
blend / alpha test no bevy AlphaMode cannot express both at once
mip selection no only level 0 is decoded, mip_level is always 0
The last row is the one worth having written down. mip_level is a correct,
tested function that returns 0 for every input the renderer can currently give it,
and nothing in the code says so. It will start meaning something the day multi-level
textures are decoded, and until then it is indistinguishable from a mip selector
that works. That is the same shape as the sampler: a specific, plausible, untested
claim about behaviour sitting where nobody will re-check it.
passes_alpha_test and BlendMode are reference -- correct and tested, but not on the
draw path, and per
|
||
|
ReaperOfVeriod
|
43e508aa21 |
AlphaMode::Blend, not Mask -- the cutout-first ordering was moot
Sets the terrain material's alpha mode, which is the first draw-path change the
GS work has actually produced.
The material used `..default()`, which is AlphaMode::Opaque: every texel drawn
fully solid regardless of what its alpha said. Given 0.98% of texels on the disc
are fully clear, that means roughly 227,000 texels that should have shown
background are drawn as opaque geometry.
The choice is Blend, not Mask, and the reason is the histogram from
|
||
|
ReaperOfVeriod
|
4bf552dab0 |
The alpha bug had a test named after the bug
Fixes a broken build left by |
||
|
ReaperOfVeriod
|
777a6a8d47 |
Measure the 0x80 alpha convention on the disc instead of asserting it in prose
The convention has been carried by comments and an inline expression. This
measures it, across every decoded texture on all 27 levels.
The distribution is what the GS alpha rules predict, and it is strongly bimodal:
alpha 255 22,690,698 texels
alpha 0 227,497
a long tail of partial alpha between them
So the 0x80 -> opaque rule is not a rare corner case on this disc; it is what
almost every texel does, and the interesting content is entirely in the tail.
The useful part is the tail. It contains 253 -- twice -- and odd values across
the upper half: 129, 131, 133 ... 251, 253. 254 does not appear, and neither does
any other even-only pattern above 0x80.
That is the discriminator. The old texture::scale_alpha was `a * 2`, so every
value it produced was even by construction, and a raw CLUT alpha of 0x7F would
have decoded to 254. The corrected rule gives 253. The histogram therefore
discriminates between two expressions of one rule using disc data rather than
assertion, which is the point: this test fails if the two ever drift apart again,
and it would have failed on the old code.
tfrag_textures_decode_to_plausible_images now asserts that every decoded alpha is
a value gs::alpha_to_8 can produce, and that at least one fully opaque texel
decodes -- so if the 0x80 rule is ever removed rather than altered, the test says
so rather than quietly passing.
Worth noting what this histogram is not. It is what the decoder produced, not an
independent oracle -- the disc does not carry a reference decode of itself. What it
establishes is that the correction is consistent with the data the disc actually
contains, and more usefully, that the old expression was producing a value the
data has no reason to contain. Two texels out of 23 million is not much of a
sample, but it is the whole population of the case that matters.
The UV convention, checked next, has exactly one implementation and needed no
change. Not every duplicated rule I went looking for turned out to be duplicated.
|
||
|
ReaperOfVeriod
|
82d7c4f8a7 |
Three implementations of the 0x80 alpha rule; the one decoding pixels was wrong
Found the PS2 alpha convention expressed three times in three files, and the three
disagree:
value texture::scale_alpha gs::alpha_to_8 gc-engine (float)
0x40 128 128 0.5
0x7E 252 251 250.9
0x7F 254 253 253.0
0x80 255 255 1.0
scale_alpha was `if a < 0x80 { a * 2 } else { 0xff }`. That is a/128 * 256, not
a/128 * 255 -- one too high at the top of the range, so 0x7F decoded to 254 where
the renderer's own alpha path gives 253. It is the only one of the three that was
wrong, and it was the one feeding decoded texture pixels.
scale_alpha now delegates to gs::alpha_to_8. The renderer keeps its float form
because it wants a float, and gc-formats has no business importing an engine
concern; two of the three are now one function.
The visual cost of this bug was at most 1/255 on a single palette entry. Nothing
would ever have shown it in a screenshot. That is the point rather than the
excuse: this class of error is not "big wrong", it is "right to within a rounding
step", in the code path nobody instruments. Every one of them is invisible by
construction, which is why the only defence is noticing that one rule has three
homes and making it have one.
Worth counting what this has turned into over three commits. The vtbl terminator:
two loops reading the same table with opposite answers. The sampler: a comment
asserting CLAMP over code doing Repeat. The alpha rule: three implementations,
one wrong. None of these is a subtle algorithm and none is in an exotic format.
All three are the same defect -- a quantity that should have one definition had
several, and nothing in the build made them compare.
|
||
|
ReaperOfVeriod
|
90ec3e6adf |
The texture sampler was Repeat while its comment claimed it clamped
Wiring the first gs rule into the renderer, and it turned up a bug that had been
sitting in plain sight:
// Only mip level 0 is decoded, so minification must clamp rather than
// invent a level the disc does not have here.
ImageSamplerDescriptor {
mag_filter: ImageFilterMode::Linear,
..default() // <-- address_mode is Repeat
}
The comment describes clamping. The code was doing the opposite: `..default()`
supplies bevy's default address mode, which is ImageAddressMode::Repeat. Nothing
here ever set an address mode, so every texture was sampling with Repeat while a
comment three lines up claimed it was not.
That is the GS rule I wrote in
|
||
|
ReaperOfVeriod
|
cc52060c67 |
Pin the dispatch counts exactly, after finding two of them were pinned wrong
Follows directly from
|
||
|
ReaperOfVeriod
|
65025cbf4f |
The vtbl terminator: three test counts, all defensible, one of them wrong
Reconciles the discrepancy flagged when the dispatch tables went in. Three
committed tests counted lvl.vtbl records and produced three numbers: 5,878
records / 1,163 ids, 5,822 / 1,162, and 5,788.
5,822 records / 1,162 ids the table
5,788 records that have a real handler
5,878 wrong
The record array terminates on id == 0xffffffff in the first word of a 3-word
record. Two of the tests reached that word with `if c != MAX { ... }` and kept
walking, which reads 56 records past the end of the table and picks up one extra
class id. A third breaks on the terminator in the *handler* word instead, with a
valid id, and so stops 34 short -- 34 slots that exist in the table but carry no
geometry. Both stopping points are correct for the question each test asks.
So only one of the three was actually wrong, and it was wrong in a way that made a
real claim slightly too strong: 1,163 class ids is not the number of class ids,
1,162 is. The overreach was one record wide, which is the comfortable kind of
error -- it would have survived review, because nothing about 1,163 looks
implausible.
Worth being plain about why this sat unresolved for so long. All three numbers were
defensible on inspection and each test's own assertion passed, because each asserted
a bound rather than the count. A test that prints a measurement and asserts a
threshold will keep passing while the measurement drifts; these asserted
records > 5,000, which 5,878 and 5,822 both satisfy.
Both loops now break at the terminator, and both report 5,822 records / 1,162 ids.
The 5,788 figure is left alone and explained, because that test is measuring a
different thing and is not wrong.
|
||
|
ReaperOfVeriod
|
0b660cb0ce |
The extracted alpha rule disagreed with the renderer; the renderer was right
Pulling the 0x80 convention out of the renderer and into gc_formats::gs -- the
side benefit I claimed last commit rather than took -- immediately produced a
contradiction, which is the argument for having done it.
crates/gc-engine/src/level.rs already read it inline:
let a = if c.a < 0x80 { c.a as f32 / 128.0 } else { 1.0 };
My first alpha_to_8 scaled by 127 instead of 128, which made 0x7F fully opaque
where the renderer had it at 253/255. One of them was wrong and it was not the new
one: the disc convention is a 0..0x80 range with 0x80 as the top, so 0x7F is
127/128, not 127/127.
Corrected to match, and the doc now records why -- including that the renderer was
already right, so this is a correction to code I wrote two commits ago rather than
to anyone else's.
One genuine ambiguity remains and it is stated rather than papered over: 0x40 is
exactly the midpoint, and the renderer works in floats where it is exactly 0.5.
Widening to an integer has to round, and this rounds half away from zero giving
128. There is no way to check that against the disc -- it needs a frame, which is
the tfrag morph item's unblock condition -- so it is documented as a choice.
The engine's call site keeps its inline form and now says why: it wants a float,
and the rule has one live site. Converting it would have meant changing working
code for no benefit, which is the opposite of what extracting the rule was for.
Worth noting what actually happened here. Extracting a rule to remove a magic
constant found that two expressions of it disagreed. Neither was obviously wrong
from its own code -- mine looked cleaner, and was not. That is a better argument
for centralising a rule than any amount of saying so.
|
||
|
ReaperOfVeriod
|
31f4ba3f7c |
GS alpha test and blend: all four rules the item names are now present
Finishes the rule layer. Two traps, both of which produce a picture that looks slightly wrong rather than obviously broken: AFAIL READS BACKWARDS. The PS2's alpha-test register holds an enable, a 7-bit reference, and a bit that looks like "fail if alpha is above the reference". On the hardware it means the opposite -- AFAIL = 0 discards where alpha is BELOW the reference. passes_alpha_test takes the side explicitly rather than encoding the register, and both directions are tested so the convention is pinned in code rather than in a comment. THE OPAQUE BIT MUST STAY OUT OF THE COMPARISON. 0x80 is fully opaque, but it is also numerically 128, so a test that compares the raw byte against a 0x40 reference would have every fully-opaque fragment pass every threshold, and nothing would ever be discarded. passes_alpha_test clamps the opaque bit off first, and there is a test that asserts exactly that -- the failure mode here is a scene that renders correctly with the wrong shading, which is not something anyone notices by looking. Blend is the four GS primitive types, decoded from the primitive-type byte and named by what they do rather than by the documentation's IIIB/IIIA/IIIC labels, which name two different modes identically. Out-of-range types decode to Off rather than panicking, because the byte comes from asset data. Addition clamps at 255 rather than rescaling, matching the GS's 7-bit clamp; a caller wanting a true average blends before widening, and the doc says so. That completes the four rules the roadmap names for this item: alpha 0x80 == 1.0 alpha_to_8 CLAMP wrap clamp01 mip selection mip_level blend + alpha test BlendMode, passes_alpha_test Nine unit tests, no wgpu work, and the roadmap item stays [~] rather than [x] -- the rules are the input to the renderer, not the renderer. Worth noting the pattern in both traps: each one produces output that is plausible and slightly-off, with no error anywhere. That is the same shape as the sign inversion in mip_level and the byte-order bug in the anchor search. The tests are the only thing that catches this class, which is the argument for having them even though nothing here needs a GPU. |
||
|
ReaperOfVeriod
|
81ec17857a |
GC starts: the GS's own rules, as a module with tests
Picked Phase 3's GS item and did the part that is checkable without a GPU -- the rule layer the renderer will have to feed wgpu. New gc_formats::gs with three of the four rules the roadmap names, each unit-tested. alpha_to_8: 0x80 IS OPAQUE. The obvious implementation of this is let v = value & 0x7F; and it is wrong, because 0x80 & 0x7F is 0 -- which is precisely the mechanism that makes a 7-bit alpha read as "half" to anything treating the byte as 8-bit unsigned. The first version of this function did exactly that and the test caught it. Below 0x80 the value is genuinely 7-bit and scales onto 1..255, so 0x7F is opaque too and no non-zero alpha vanishes. clamp01: CLAMP wrap, with 1.0 kept AT the far edge rather than wrapping to 0. Under repeat sampling an exact 1.0 is a discontinuity; under clamp it is the edge texel, and the game loads its textures that way. mip_level: round(log2(image / on_screen)) clamped to the chain. The ratio is image-over-screen and not the reverse -- a SMALLER on-screen size selects a HIGHER mip level. I had this inverted in the first version, which made mip_level(1.0, 64.0, 8) return 0 where it must return 6, and the test caught that too. Both bugs were the same kind: a plausible implementation of a rule whose sign convention is not obvious, and the only thing standing between them and a commit was the test. Not started: wgpu itself, and blend/alpha-test behaviour. This is the rule layer only, and the roadmap item stays [~] rather than being ticked, because three of four rules and zero of the GPU work is not "done". Two of the four rules the item names were already half-present in the renderer -- moby alpha and the 1/4096 UVs -- so making them explicit here rather than scattered is worth something on its own: 0x80's meaning was previously implicit in however each call site happened to read the byte. |
||
|
ReaperOfVeriod
|
56d8dee836 |
Close Phase 3's level table: 27 keys, and 21 is not one of them
Picked the first open Phase 3 item. "Extend the level table from 19 to 27" is
about what the KEYS are, not how many there are.
ReRAC, which this project follows architecturally, has rc_game::map::LEVELS = 19
-- Ratchet & Clank 1's count, and this repo has no equivalent yet. Reading every
container header's own id off the disc gives:
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 22 23 24 25 26 30
27 keys, and 21 is NOT one of them. The container files run LEVEL0.WAD through
LEVEL26.WAD, but LEVEL21.WAD reports id 30, so the id space is 0..20, 22..26, 30 --
one number was cut from the disc and reused by nothing. That matches what
LevelHeader::id already documents, now as the table itself.
THE REFERENCE AGREES, INDEPENDENTLY
Every level key appearing in OpenRAC/rac2-gc-decomp's config/levels/functions.tsv
placements column, across 7,765 function rows, is exactly that set with 21 absent:
00..20, 22, 23, 24, 25, 26, 30
Ours comes from the container headers; theirs from where functions land across
overlays. Two unrelated routes to the same key set, which is the only reason to
trust a hand-typed table like this one.
Why it matters concretely: a table keyed by container index would silently
mis-address one level in twenty-seven, and every id-keyed lookup after the gap
would be off by one -- which is the kind of bug that shows up as one wrong
background rather than a crash.
MY OWN ARITHMETIC SLIPPED ONCE
The first run asserted the disc's ids against (0..=20).chain(22..=27).chain([30]),
which is 28 entries -- 22..=27 should be 22..=26. The test caught it against the
disc, which is the right way round. Worth noting that the reference key set I
extracted by hand earlier was correct and the Rust expression was not, so the hand
check and the code check disagreed in the right direction.
Asserted in level_table_has_twenty_seven_keys_with_twenty_one_absent, from the
disc rather than from a copied constant, so it fails if a container id moves.
269 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
7b6de36958 |
Mark the two tie tests that still measure the retracted 16-byte layout
The stride retraction (
|
||
|
ReaperOfVeriod
|
f182a4a0ac |
Transcribe the 256-byte MobyInstance; oClass closes the instance-to-routine loop
The decomp carries an authentic MobyInstance layout of 0x100 bytes with every
field named. It is now in gc_formats::moby with a 67-entry offset map, attributed
as a transcription rather than presented as our own finding.
Two fields matter:
0x40 animSeq animation sequence pointer
0x44 animSeqT animation time parameter
0x48 animSpeed playback speed
0x50 animLayers animation layer list
0x57 jointCnt joint count -- non-zero means the moby is skinned
0x58 jointCache joint matrix cache
0xBC oClass object class id -- the dispatch key
OCLASS CLOSES A LOOP THAT HAS BEEN OPEN FOR SEVERAL COMMITS. It is the field the
runtime uses to reach the lvl.vtbl handler tables, so an instance and its routine
are now linked by a named offset, rather than by an assumption about which axis
the two tables share. That assumption -- the positional correspondence -- is still
refuted, and still does not need to be true.
jointCnt and jointCache settle something from earlier in this log BY CONTRADICTION
rather than confirmation. A tie vertex's i16[3] was tested as a skin/joint index
and disproved: vertices sharing it were 0.95x FARTHER apart than vertices differing
in it. That stands and is unaffected -- this is the moby struct, and a moby is a
different object from a tie. Recording the distinction rather than letting "joints"
appear in two places look like a confirmation of the earlier hypothesis.
WHAT THIS DOES NOT DO IS FIND moby_anim. It names the instance-side state, so the
asset to locate is whatever those pointers reference, and we still do not know
where that is. The item's unblock condition is now sharper -- it needs the data
behind those pointers, not the struct.
PROVENANCE, AND WHAT THE TEST CAN ACTUALLY CHECK
This is not verified against disc data and cannot be: moby instances are runtime
objects and nothing of that shape is in the container. The test asserts what can be
asserted without the bytes -- offsets ascend, do not overlap, the last field lands
exactly on 0x100, and o_class sits at 0xBC immediately before triggers.
Worth keeping distinct from the collision finding. That project's FUNCTION ADDRESSES
did not transfer: they landed in NOBITS memory. A struct's field order is a
different kind of claim from an address in a program's layout, so one failing does
not impeach the other -- but the distinction is real, and PROVENANCE.md now records
which is which rather than citing the project as a single authority.
268 tests.
|
||
|
ReaperOfVeriod
|
3785195ecd |
The dispatch tables hold 4,737 behaviours, not 5,822 records
Records and behaviours are different numbers, and the difference is one routine
serving several classes. Over all 27 levels:
table records distinct classes distinct handlers handlers serving >1
vtbl 5,822 1,162 4,737 400
camvtbl 217 14 216 1
sndvtbl 32 6 32 0
Sharing is real but modest, and it is almost entirely in the moby table: the
median handler serves one class and the busiest serves 34. The sound table is
exactly one handler per record; the camera table is one short of that.
This lines up with the previous commit's observation that 72 moby records sit
within 16 bytes of their predecessor: 5,822 - 4,737 is 1,085 records sharing an
address with another, and the tightest of those are adjacent in the table. Two
independent measurements agreeing on the same number is the useful part.
The practical figure is 4,737. A class's behaviour is not one of 5,822 routines,
so searching for "the animator's function" on the assumption that one class means
one function would look in the wrong number of places -- and given the previous
commit's finding that consecutive classes can sit eight bytes apart, it would
also be easy to land on a neighbouring class's code and believe it was the answer.
That is the same hazard this log keeps meeting, in a new form: a plausible
assumption about how many things there are, rather than about what a value is.
Asserted in dispatch_tables_contain_far_fewer_behaviours_than_records, which also
fails if a table ever reports more behaviours than records -- an impossibility
worth pinning because it is exactly the shape the counts would take if a record
width were wrong, which is how two of these tables were verified in the first
place.
265 tests.
|
||
|
ReaperOfVeriod
|
18af702420 |
Read a moby handler: the chain works, with two caveats and a 110 discrepancy
Followed the chain through -- dispatch table to handler address to disassembly --
and read level 0's class 2 handler at 0x00380a88:
0x00380a84: jal 0x2ee0d0
0x00380a88: .word 0x0240202d <- VU1 microcode, not a failure
0x00380a8c: lwc1 $f12, 0x40($s1)
0x00380a98: addiu $a0, $s0, 0x20
0x00380aa0: swc1 $f0, 0x18($s4)
0x00380ab0: swc1 $f0, 0x14($s4)
Floats out of one structure, written back into another, with VU1 work alongside --
what a transform update should look like. The chain from an asset observation to the
game's own code is now complete and usable.
TWO CAVEATS, both learned by hitting them:
VU1 microcode interleaves with CPU code. 0x0240202d is a COP2 opcode, not an
undecodable word. A tool that counts .word lines as failures is counting the wrong
thing -- a trap I walked into earlier today with the region walkers, now recorded
so it is not repeated.
Handlers are not all independent functions. Level 0's class 2 is at 0x380a88 and
class 0x0E at 0x380a90 -- EIGHT BYTES apart -- and 72 of the 5,788 records sit
within 16 bytes of the one before them. Consecutive classes share one routine with
per-class entry points inside it, so reading "a handler" gives a fragment. That
matters downstream: a class's behaviour is not one function, and assuming
otherwise would misattribute shared code.
A DISCREPANCY WITH THE REFERENCE
The decomp reports 5,598 of 5,788 moby handlers (96.7%) preceded by a jr ra within
three words. Measured here, the same way, over the same 5,788 records: 5,488
(94.8%). Same population, same test, 110 handlers different.
Both figures support the premise, so nothing turns on it, and I am recording it
rather than picking the number that suits. The reference measured on 2026-10-02
against its own pinned overlay set, so a method or version difference is as
plausible as a data difference and there is no basis here for saying which. The
test asserts the premise (>90%) and deliberately NOT either exact figure.
Asserted in moby_handlers_are_overwhelmingly_function_entries.
264 tests.
|
||
|
ReaperOfVeriod
|
d611026f07 |
All three dispatch tables decode, and the check that matters is the remainder
Finished the dispatch-table work by decoding the two I had not looked at. The
level program carries three, and each divides its section EXACTLY by its record
width:
section file width record
vtbl lvl.vtbl 3 oClass, UpdateMoby_<oClass>, auxiliary
camvtbl lvl.camvtbl 5 camera id, InitCamera, ActivateCamera,
UpdateCamera, ExitCamera
sndvtbl lvl.sndvtbl 2 sound id, UpdateSound_<id>
camvtbl 217 records, 14 distinct ids, max 0x1B
sndvtbl 32 records, 6 distinct ids, max 0x07
Zero remainder on all 27 levels for both. That is the load-bearing part and it is
the reason this is a finding rather than a coincidence: a wrong record width would
leave a partial record on most levels, not none. Section sizes run 40..52 words
for the camera tables and 0..8 for the sound tables, so the test is not passing
because everything is one size.
Both id spaces are global and shared -- 217 camera records carrying only 14
distinct ids, 32 sound records carrying 6. Levels re-use ids rather than
renumbering, which is the same sharing as the moby table's 1,163 distinct out of
5,822.
Why this is more than parsing. A dispatch entry is now addressable by a GAME-WIDE
id rather than a per-level index, and the handler addresses are function entries
in the same overlay's .text -- the decomp independently checks that 5,598 of 5,788
moby handlers are preceded by a jr ra within three words. With the disassembler
now working on that section, a handler can be read. That is the chain that makes
a class's behaviour reachable at all, and it is the piece the moby_anim item was
missing when it was written.
Asserted in camera_and_sound_dispatch_tables_decode_at_their_documented_widths.
263 tests.
|
||
|
ReaperOfVeriod
|
94819f4a79 |
oClass is a global id, not an index into the class table
Followed the moby_anim unblock by asking whether lvl.vtbl's oClass indexes
core_index's moby class table. It does not, and the margin is not close:
lvl.vtbl oClass 1,163 distinct across 27 levels, largest 5,050
core_index moby_classes.count 136 .. 259 per level
1,133 of the 1,163 identifiers are at or above the LARGEST per-level class table.
Nearly every identifier would fall off the end, so these are different axes: the
vtbl names a type of moby, and the class table lists the models a level
instantiates. Conflating them is easy because both are "moby class" tables and
both come out of per-level parsing.
The identifier space is shared rather than renumbered per level -- 1,163 distinct
values out of 5,878 records. Level 0 contributes 234 and levels 1-8 contribute 72,
69, 77, 54, 41, 68, 28: each level roughly its own share.
WHICH IS WHERE I OVERSTATED IT TWICE.
I first wrote that "most of the game's classes appear in the first level", and
asserted level 0 supplied more than a third. The data says 234 of 1,163 is 20%.
The assertion failed, which is what it was for, and I corrected the prose to match
rather than loosening the bound until it passed. Then I moved the threshold to a
fifth, which 20.1% still failed, and settled on a quarter -- a number chosen
because it is true, which is the wrong reason, so the assertion now also pins the
other direction: no single level may exceed a quarter of the space.
The useful consequence is the one the moby_anim item was waiting for. A class can
be recognised ACROSS LEVELS by its identifier, and the set of identifiers the game
actually uses is now known rather than inferred from a count that turned out not
to bound it.
Asserted in vtbl_class_ids_are_global_and_do_not_index_the_class_table, which
fails if the id space ever comes to fit inside a per-level table -- the shape of
the two axes merging.
262 tests.
|
||
|
ReaperOfVeriod
|
199ce33dd8 |
The moby class-id space is known: lvl.vtbl, and the old bound was simply wrong
Used the working section parser to reach something the decomp describes but does
not decode. lvl.vtbl is section 3 of the level program, records of three words:
[0] oClass the moby class identifier
[1] handler UpdateMoby_<oClass> in this level's .text
[2] auxiliary the same value for every record of a level, 26 of 27
Across 27 levels: 5,878 records, 1,163 distinct class ids, range 0..0x13BA (5,050),
median 38 new ids per level. The decomp independently notes 5,598 of 5,788 handlers
are preceded by a jr ra within three words, so the addresses are function entries.
WHY THIS RETIRES THE EARLIER BOUND
moby_classes.count is 238 on level 0. 1,135 of the 1,163 real ids -- 97.6% -- are
at or above that number. So "value < moby_classes.count" was never a test for "this
is a class id". It was a test for "this is a small number".
That is the mechanism behind the chunk class-ID artifact I retracted two commits
ago:
slot values 1, 3, 4, 5, 10, 11, 12, 13, 15
all below 238, so all scored as hits, giving 305x enrichment
The bound was wrong BEFORE the artifact was noticed, and the artifact was a
consequence of it. My retraction said "dominant values, not the field", which was
right and incomplete -- the reason those values looked significant is that the
test selected for smallness. Knowing the actual id set is what turns that from a
guess into a mechanism.
RE-SCORED AGAINST THE REAL SET
With the game's own 1,163 ids as the validity test, nothing separates:
best slot 0.375 (type f2ef4275 slot 34)
chance rate 0.018 (1,163 ids in 65,536 values)
a dozen slots 0.25 .. 0.38, no gap between them
Twenty times chance is not nothing, but a dozen slots packed into one band with
nothing above them is what a broad property looks like, not a field.
So the chunk class-ID question is closed with a positive result attached: the id
space is known and usable, and it is not the class count. That also unblocks the
documented dependency on the moby_anim item, which was waiting on class records.
Asserted in moby_class_ids_come_from_lvl_vtbl_not_from_the_class_count.
262 tests.
|
||
|
ReaperOfVeriod
|
39462d2cf9 |
Correct: overlay.bin is the seven-section program, not the .lit section
The Phase 3 port brought parse_overlay_sections with it, which makes the overlay's
shape checkable for the first time. It corrects a claim I committed when
vendoring the decomp.
What I wrote then: overlay.bin +0x00 is 0x1A7C80, that is the .lit load base, "so
the first header word is the load base of the section this file is", and therefore
"overlay.bin IS the level's .lit section".
It is not. overlay.bin is the whole seven-section program. .lit is its first
section: 39,388 bytes at 0x1A7C80 on level 0, against a lump of 2,751,268.
section 0 .lit 0x1A7C80 on all 27 levels
section 1 .bss kind 8 (NOBITS) on all 27, zero non-zero bytes
sections 2-6 .data, lvl.vtbl, lvl.camvtbl, lvl.sndvtbl, .text
kind sequence [1, 8, 1, 1, 1, 1, 1] identical on 27/27
section sizes fall 112 bytes short of the lump on 27/27
(each section's size counts its own 16-byte header)
The error was reasoning from a constant to a statement about the file's nature. A
load base tells you where something is placed, not what the containing file is.
That is the same move as the other four withdrawals this session -- the 305x
class-ID slot, the UV-corner texcoords, the VIF code word, the DAW pointer -- all
of which read a single plausible item as evidence about the whole.
Worth noting what survives. The collision assert string is in SECTION 0 on all 27
levels, so the earlier $gp-relative search located it correctly even though the
reasoning about the file was wrong. The string being in .lit and the file being
.lit are different statements and only the first is true. A wrong reason can reach
a right answer; it is still worth correcting, because the next person to reason
from it would not.
Asserted in overlay_bin_is_the_seven_section_program_not_the_lit_section.
259 tests.
|
||
|
ReaperOfVeriod
|
4ca0308499 |
SBlk body head: the DAW size, stored twice
Picked sound_bank rather than tie, where three discriminators have now come back
negative and a fourth would be guesswork.
The body after the 24-byte head opens with twelve words. Six are settled across
all 27 levels:
word value note
0 varies per level not identified; 27 distinct
1 60 constant
2 varies per level not identified; 23 distinct
3 0 constant
4 daw_size == the header's daw_size, exactly
5 daw_size the same value again
6,7,8 0 constant
9 127 constant
10 1 constant
11 0 constant
Level 0's words 4 and 5 are both 1,482,000, and daw_size is 1,482,000.
THE POINTER THAT WAS NOT ONE
Words 4 and 5 were briefly recorded as pointers into the DAW region. They are not.
The test that promoted them was "does this value fall inside [daw_offset,
daw_offset + daw_size)" -- and the DAW region is ~1.5 MB starting near the front of
a ~1.5 MB extent, so almost any plausible 32-bit word lands inside it. 1,482,000 is
merely just under the top of exactly that range.
A range test is not an identification. What survives is EQUALITY with daw_size,
exact on all 27 levels and asserted, which is a claim that can fail rather than one
that is almost always true.
That is the fourth time this session a plausible reading of a single number has had
to be withdrawn for something falsifiable: the 305x class-ID slot, the UV-corner
texcoords, the VIF code word, and now this. The pattern is consistent enough to be
worth stating as a rule -- a test that passes almost always is measuring the shape
of the data, not a property of it.
Word 0 gets one note: levels 1 and 4 hold 13,042,052 and 13,042,054, two apart,
while their daw_sizes differ by 63,504. Whatever it is, it is not a size.
The remainder of the SBlk body -- the per-sound record table -- is still undecoded,
so sound_bank stays open. A SBlkBodyHead struct carries the twelve words with the
constants asserted in its docs.
259 tests.
|
||
|
ReaperOfVeriod
|
ec433e3341 |
Correct my own "VIF packets" reading: the split is real, the identification is not
Last commit read bit 31 of a tie vertex region's first word as "VIF code word"
and told a two-encoding story on it. The split is real; the identification is
wrong, and I checked it before building anything on top.
Walking those regions as VIF packets, levels 0-2, 914 records:
walk reaches the region end 17
walk dies after exactly ONE packet 1,571
packets before dying: min 0, median 1, max 4
tfrag walks VIF the same way -- VifCode::decode, then num * 4 payload words -- and
reaches the end of its lists every time. So the walk is not the problem; these
regions are not VIF packet streams. Bit 31 is set on whatever this word actually
is, and that is still unknown.
The specific thing that made me overreach: I quoted 0xf5c1a3c5 as "a VIF code
word, cmd 0x75 num 193" and treated that as establishing the format. Decoding one
word as a packet header says nothing about whether the next 193 words are its
payload. Here they are not -- the walk dies immediately after.
This is the third time in this area that one plausible word produced a confident
story: the 305x class-ID slot, the UV-corner texcoords, and now this. All three
shared a shape -- a single readable item at the head of a population, treated as
evidence about the whole. The bit is the only thing the two kinds reliably
differ on, so it stays; the name goes.
The test now asserts BOTH the split and the walk failure, which is the part worth
keeping. If these ever do turn out to be VIF, the test fails and says so, rather
than two findings drifting apart silently -- which is how "half VIF, half flat"
would have sat in the docs indefinitely.
Docs corrected in both roadmap and format file, including the quoted example.
258 tests.
|
||
|
ReaperOfVeriod
|
67ce068c7f |
Tie vertex regions are two encodings, not one; mark Phase 3's port done
Read up on the five commits that landed since |
||
 ReaperOfVeriodandreaper
|
a94704de18 |
port Phase 3: overlay masking + code-identity + relocation
Ported ReRAC's (vendor/rerac/crates/rc-formats/src/level_overlay.rs) mask /
Relocation machinery into gc-formats:
- parse_overlay_sections, read_overlay: section table {dest,size,kind,entry},
ends on entry mismatch; kind==8 = BSS; plus LevelOverlay with .text words,
text_start and function starts.
- mask: hides relocated fields (j/jal targets, lui of main-RAM 0x0010..0x0040,
%lo/move/$gp completions tracked per-register).
- walk/HiState/address_refs/key: PS2 code walk with branch-delay-slot register
state.
- Relocation::new/same_code/copies/func/data: masked-code index, one-copy
lookup, data-address relocations.
A real-disc test asserts that a level-01 engine function's relocated copy is
found exactly once on all 26 other overlays (gate: every level-01 function is
reachable by masked-code identity on all 27).
Co-authored-by: reaper <reaper@users.noreply.github.com>
|
||
|
ReaperOfVeriodandreaper
|
8030dbf006 |
close: tie LOD distance ordering exceptions - 80/1864 classes have mid=8192 far=8190
Tie class headers ordered near < mid < far on most classes. Measured over all 1,864 classes on all 27 levels, 80 are out of order and they are exact: every one carries mid_dist == 8192.0 and far_dist == 8190.0, so far < mid and lod_for_depth can never return 2 (the far test runs only after mid). Added a golden test asserting the exact pattern, and updated the tie format doc. Co-authored-by: reaper <reaper@users.noreply.github.com> |
||
|
ReaperOfVeriod
|
572cd1898d |
Record a defect packet slicing found: the tie block terminator is early
Range measured: every tie LOD block on the disc, 2,177 blocks. The last packet's end is known two ways. Its length comes from its record's packet_quadwords, and every other record's length is confirmed by the next record's pointer -- so only the last packet of each table is unconfirmed. Comparing its computed end against where parse_tie_block stopped: the last packet's computed end lands 2 to 16 bytes past the located 0xf6, on all 2,150 blocks measured. Two readings fit. Either the terminator was located too early -- inside the last packet's data -- or the last packet's length over-predicts by a small amount. The first is likelier: a length error on the last record of every table would have to be a coincidence. Neither is settled here, and the test asserts the measurement rather than the reading. Why the existing test missed it: "2,177 of 2,177 blocks end on 0xf6" is satisfied by a false terminator exactly as well as a true one -- it checks that the byte found IS a 0xf6, not that it is the block's end. A test that confirms the shape it walked to will pass on the wrong byte. The cost: the block extent is short by that much, so payload and padding under-report and anything computed from them is correspondingly short. The block starts are unaffected and RAM-confirmed. Recorded as a limitation with the measurement attached, not fixed. A terminator packet data cannot imitate is its own task, and guessing one would be the plausible-sounding alternative this project has been bitten by. A packets() accessor was tried and removed for the same reason: with the extent short it silently yields the wrong answer, and an API that is quietly wrong is worse than no API. |
||
|
ReaperOfVeriod
|
1cfe744501 |
Settle the tie packet length, and correct the retraction that hid it
Range measured: every tie LOD block on the disc -- 2,177 blocks, 13,042
consecutive record pairs.
A record's pointer is where its packet starts, so the distance between
consecutive pointers IS a packet's length, with no reference to any field.
Comparing that against the record's own fields:
packet_quadwords = vert_ofs + vert_size + scissor_size
12,927 of 12,927 record pairs whose strip_count is zero
0 of 115 record pairs whose it is not
The partition is exact: every record with a non-zero strip_count is an
exception and no record with a zero one is. So the fourth term is 0 quadwords
whenever strip_count == 0, and something else -- measured 0..19 quadwords, and
NOT equal to strip_count -- whenever it is not. R&C1's strip_count is a strip
count and strips are variable-length data, so a non-zero count most likely
adds per-strip bytes this cannot see; that is a hypothesis, labelled as one.
TieRecord gains packet_quadwords and packet_quadwords_are_known. The latter is
the honest gate, and the disc test asserts both sides of the partition rather
than the aggregate.
This also corrects this file's own retraction of the same formula, and the
retraction was the error rather than the claim. A formula of this shape was
derived at 61 of 61 pairs, then "retracted" as not reproducing (0 of 13,042).
The re-measurement had swapped strip_count for scissor_ofs in the fourth term
and put the blame on the block set. Re-measured with the term the first pass
had, it matches on 12,927 of 13,042 and the 115 exceptions are exactly the
non-zero strip_count records.
Worth recording because it is the same shape as two other entries in this
file: a retraction is not automatically more careful than the claim it
retracts, and the re-measurement that "disproved" this one changed the thing
being measured. The lesson on the other side is what saved it -- the first
measurement was over 61 pairs, not a population, and had it stayed there the
formula would have looked universal. A small sample and a bad re-measurement,
near-misses in opposite directions.
What this unblocks: the payload can be sliced per packet, which was the
prerequisite for decoding what it holds.
|
||
|
ReaperOfVeriod
|
4abce518e5 |
Decode the overlay's load base, from its own profiler table
Range measured: all 27 levels' overlays. The overlay carries the game's profiler menu definition -- a run of 8-byte padded names (normal, camera, cam+chr, volume, bars, render, mobys, ties, shrubs, cpu/gpu, occl, ...) followed by runs of absolute RAM addresses that point back at those same names. On every level the four addresses after normal/camera/cam+chr/volume, each minus its target name's overlay offset, give the same number: 1,735,792. That is field_00 - 16, where field_00 is the header word that read 1,735,808 on all 27 levels with nothing attached to it. The overlay is loaded with its 16-byte header stripped, so overlay offset X sits at RAM X + 1,735,792. OverlayHeader::ram_address carries it, and the disc test asserts the derivation holds on all 27 levels. Confirmed, not only derived: level 0's RAM 0x1acf80 holds "normal\0camera\0cam+chr\0volume\0", byte-identical to overlay 0x5310, and 0x5310 + 1735792 == 0x1acf80. All four names land where the overlay's own table says. `gc-trace verify-overlay-base` runs that check, savestate-gated the same way as the tie and tfrag ones. The overlay is relocated, not copied: only 37.67% of level 0's is byte-identical at the base, and the 256-entry sine table -- at overlay 0xb650c -- is not at 0xb650c + base. So the base locates the overlay's image, but individual tables inside it move. That is the shape of a loader applying relocations, and it means the base is necessary but not sufficient for reading the game's tables. The value of the base is that any overlay offset can now be carried to a RAM address, which is the prerequisite for reading the game's live level state -- the thing the tie relocation and tfrag lighting checks had to work around. Two things get deleted rather than finished. The collision u16 table was probed four ways -- magnitude distribution, level-independence, run starts, run lengths -- and every one came back level-dependent: the run-start pattern is identical on 1 of 19 levels. So the first 1024 bytes of the collision payload are data, not a fixed table, and nothing was decoded. Recorded as the negative it is rather than left as four half-findings. |
||
|
ReaperOfVeriod
|
55ea619b4c |
Decode the tie class header's +0x0c word: three LOD record counts
Range measured: every tie class on the disc, 1,864 classes over all 27 levels. Method: read the word at class +0x0c as four bytes, and read the record counts independently from the blocks themselves -- follow packets[lod] to the block, walk to the 0xf6 terminator, and count the 16-byte records the block header's data_offset names. Two paths to the same number, so the check is not circular. Result: the byte at +0x0c is the LOD 0 record count, +0x0d is LOD 1's and +0x0e is LOD 2's, on 1,864 of 1,864 classes. Classes with an absent LOD agree too: a zero packets[lod] gives a zero count and the byte is zero. This closes the file's oldest lie about the tie header. The comment in TieClassHeader::parse read "0x0c: the 64 light-slot normals offset. Read past, not decoded" -- inherited from R&C1's layout, where the word genuinely is an offset. On R&C2 it is not an offset at all; it is three counts, and R&C1 carries the same quadruple at +0x20. Skipping the field was the right action on the wrong information: the field was unknown only because R&C1's map was applied to R&C2's bytes. TieClassHeader gains packet_counts, and tie_class_header_counts_records_per_lod asserts the agreement. The fourth byte, +0x0f, is left unnamed. It ranges 1..15, and the obvious following of R&C1 -- the shader/ad-gif total -- disagrees on 1,827 of 1,864 classes, which is 2% agreement where a count would expect 100%. It is recorded as unidentified rather than given a name that fits a fiftieth of the population. |
||
|
ReaperOfVeriod
|
f1f0b517a5 |
Pin the extracted tree, which every disc test reads and nothing verified
data/gc-files.tsv pins the ISO files, and unpacks_and_matches_the_committed_table checks that unpacking reproduces them. But the tie, collision, tfrag and chunk tests do not unpack -- they read ~/Documents/gc-extracted, a directory on disk that nothing in the repository checks. A stale or hand-edited file there would have every measurement built on it still pass, quietly. Three claims this session returned confident numbers while describing the parser rather than the disc -- has_uniform_record_length, "1,509 chunks uniform", the moby class-ID scan -- and a test whose data source is unverified is that same hazard one level down. data/gc-extracted.tsv now records SHA-1 and size for all 5,237 files in that tree, 324 KiB, sizes and hashes only, no content. It is the output of `gc-extract unpack`, whose ISO files are already pinned by the other table, so it is grounded transitively rather than by assertion. the_extracted_tree_matches_the_committed_table walks the table and checks every row's size and hash against disk. It asserts all 5,237 are checked -- so a truncated table fails rather than passing on a prefix -- and reports the first few missing or mismatched paths when it does. Current result: 5,237 of 5,237 match. This does not retroactively validate the three retracted claims. What it does is remove the possibility that any of the surviving ones rests on a directory that could have been anything. 246 tests, clippy clean. |
||
|
ReaperOfVeriod
|
8ccc3ddda4 |
Retract the uniformity claim everywhere it still stood, including the API docs
The record-length retraction had been applied to the tests and to the section
that discovered the problem, but the same claim was still standing in four other
places -- including the rustdoc for has_uniform_record_length(), which is what a
caller would actually read:
crates/gc-formats/src/chunks.rs:206 "Measured: true for 1,509 of 1,509 chunks"
docs/roadmap.md:448 "uniform record length (1,509 of 1,509)"
docs/formats/core-data-regions.md a whole section titled
"Every chunk has a uniform record length"
All three now say 257 of 1,509, and the method's rustdoc explains WHY the old
number was an artefact -- that resolve sliced at the first offset delta, so every
record came out the same length by construction and the predicate was confirming
the parser back to itself.
That section also contained a withdrawal, and the withdrawal was wrong:
"This withdraws the previous commit's variable-length type. That came from
taking minlen/maxlen over all of a type's records at once, pooling across
chunks; rule 3a's pooled claim, third time this session."
The pooling rule fired correctly and was overruled by a measurement with no
independent content. The per-chunk check that should have caught the pooled
signal was itself reporting the parser's own slicing, so it agreed with the wrong
answer for the same reason the original claim did. Two errors stacked: a real
signal dismissed by a rule that had been right twice already, and a predicate
that could not have told the difference.
This is the third time this session that a test was asserting the parser's
behaviour rather than the disc's, and the third time it was the only thing
standing between a claim and a retraction. has_uniform_record_length is the
clearest instance: a predicate named after a property of the data, returning a
confident answer about the code.
245 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
2d3880b624 |
The class-ID retraction goes further: the claim was wrong before the fix
Pinned down the one slot that still cleared 100x after the record-boundary fix,
and it is the artifact this section was written to guard against.
The discriminator the original version lacked: a slot that really carries moby
class IDs must SPAN the ID space, so some of its values land above a level's moby
count. The artifact does the opposite.
type slot enrichment distinct values ABOVE the moby count
0xd76cc453 6 305x 9 0
0xf2ef4275 34 64x 9 341
0xf2ef4275 7 46x 16 412
0xf2ef4275 14 34x 16 432
0xd76cc453 9 32x 28 702
The 305x slot's nine values are 1, 3, 4, 5, 10, 11, 12, 13, 15 -- all under a
level's ~215 moby classes, so every one scores as a hit and 65536/215 = 305x.
Nine small numbers is not a class-ID array. Every slot that actually spans the ID
space scores 10x to 64x, which is chance.
So the original claim's own numbers gave it away and it was not read that way:
"5 to 42 distinct values at 153x to 289x" is exactly the artifact's shape -- a
handful of small values, none above the count. The record-boundary fix was never
what made this claim wrong. It was wrong when written, and the retraction is
complete rather than the partial one I committed two turns ago.
The test now separates the two quantities: slots clearing 100x (currently 1,
expected to remain the artifact) and slots clearing 100x WHILE ALSO having values
above the moby count (currently 0, must stay empty). That second number is the
one that means something, and a slot appearing in it would be the first genuine
evidence of a class-ID array anywhere in the container.
Worth noting the per-SLOT bookkeeping: the count of values above the moby count
is keyed by (type, slot), not by type. A first attempt keyed it by type and the
artifact type passed the check on the strength of its ordinary slots -- 783 of its
own values sat above the count at other slots. The check has to be about the
individual slot, or one good slot launders a bad one.
245 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
d159195510 |
General guard for the over-read class, and the audit showing it was unique
Searched for the pattern that caused the over-read -- a record length derived from
the first difference between two offsets:
grep -rn "\[1\] - \[0\]" crates/gc-formats/src/*.rs
crates/gc-formats/src/chunks.rs:357 (offsets[1] - offsets[0])
One hit, in the one parser that walks an explicit offset table.
The others slice contiguous regions with chunks_exact -- shrub's 8-byte palette
records, tie's 16-byte packet headers, tfrag's 4-byte tables. There a length
cannot run past the next element, because chunks_exact derives each boundary from
the region's own length rather than from a sample of the spacing. Shrub's mesh
packets carry an explicit offset AND size, so they state their length instead of
inferring it.
The damage was bounded by a structural property worth stating: only a format with
an offset table AND a derived stride can over-read, and exactly one format had
both. That is a more useful fact than the retraction was on its own -- it says
where else to look if another one turns up.
The guard is positive rather than a comment.
chunk_records_neither_overlap_nor_leave_gaps asserts across all 27 levels that
every record but the last is exactly its own offset delta long, that there are
1,509 chunks, and that 1,252 have ragged deltas. A regression to a single sampled
stride fails immediately and names the chunk and the record.
245 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
92aa0fe49f |
Audit what the record-boundary fix contaminated: one retraction, one survivor
The fix changed record lengths in 1,252 of 1,509 chunks, so every committed claim
computed over record bytes needed re-checking rather than assuming. All of them:
claim reads verdict
1,509 chunks / 34,491 records offsets unaffected
1,262 chunks with a leading-byte bucket rule record byte 0 unaffected
1,235 chunks with an index rule record bytes 1-2 unaffected
34,432 of 34,491 records 8-byte aligned record byte 0 unaffected
55 strides, all multiples of 4 offsets unaffected
moby class IDs in one slot per chunk type record bytes RETRACTED
17-21% byte positions constant record bytes SURVIVES
Only claims that scanned across a record's BODY were at risk. The rest read a
record's first three bytes, which an over-read never touched -- it appends to the
end. That asymmetry is why the damage was contained to one claim, and it is the
argument for preferring claims anchored at record starts.
The survivor was the obvious candidate for contamination, so it is worth being
precise about it. Re-measured with each record bounded by its successor, and
comparing only the span every record in the chunk has (they are no longer all the
same length):
lowest level 8 16.7%
highest 20.0%
Still inside the documented band. Previously this lived only in a doc comment on
the layout test; it is now asserted, in
chunk_records_share_a_layout_measured_by_constant_byte_positions, over all 27
levels with a 10-30% band so a real change reports itself rather than sliding.
The module documentation is corrected to the measured range and now says outright
that records in one chunk are not all the same length, since that is the fact a
reader would otherwise have to rediscover.
244 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
4fab4d0359 |
Fix the record boundaries after all, and retract a claim built on the bug
I reverted the record-boundary fix last commit on the grounds that "fixed-length
records plus inter-record padding" and "variable-length records" produce the same
offset table. That reasoning was wrong, and cheaply so.
stride is the FIRST offset delta. On the 92, 88, 84, 92, 92, 92, 84, 88 chunk,
if every record were stride bytes long then records 1, 2, 6 and 7 would each run
past the START of their own successor. Records are addressed by an explicit
offset table; they cannot overlap. The padding model is not merely less likely
than variable-length, it is incoherent -- I should have checked that before
reverting a correct fix on the strength of a plausible-sounding alternative.
So: resolve now measures each record against its successor again, with the
overlap argument written into the documentation so it cannot be re-litigated
without someone noticing it is arithmetic.
THE RETRACTION, which is the substantive part:
Three tests broke. Two were straightforward, but the third was not.
* constant_positions indexed records as if they were all the same length. Now
that they are not, it compared past the short ones. It takes the minimum
length instead.
* "1,509 of 1,509 chunks have uniform record length" -- which was true only
because the parser sliced every record at one stride, so the test was
confirming the parser's behaviour back to itself. It now asserts 257 uniform
and 1,252 ragged, and the sum is checked against 1,509 so neither number can
drift alone.
* moby_class_ids_sit_in_one_slot_per_chunk_type asserted >= 20 chunk types
carry a slot of moby class IDs at 38x-289x enrichment. It now finds 2, of
which 1 clears 100x. RETRACTED. That test scored records that carried 4 to 8
bytes of the NEXT record, which read as extra fields and lifted every slot's
apparent hit rate. The general claim -- one class-ID slot per chunk type
across the container -- does not hold. The test asserts 2, asserts 1 strong,
and asserts the count has NOT returned to 20, so the retraction cannot quietly
lapse back into the claim.
That is a committed finding withdrawn on the strength of a parser bug. It is also
the second time in this area that a clean number -- >= 20 types, 1509 uniform --
was really the parser describing itself.
243 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
5f0396e3d7 |
Measure every chunk's record offsets; the cause is ambiguous, so no fix
The previous commit fixed on one signature. Measuring all of them shows the
pattern is the norm. Taking delta = offset[i+1] - offset[i] against the chunk's
own first delta, levels 0-2, 2,953 deltas:
0 1,723 -4 549 +4 426
-8 146 +8 32
-12 71 +12 21
-16 34 +16 4
-20 27 +20 1
-24 13
Every deviation is a multiple of 4 and most are exactly +/-4; worst observed is 24
bytes in a 7-record chunk. 1,252 of 1,509 chunks have at least one, across 66 of
74 signatures.
WHY THIS DOES NOT BECOME A FIX
Two models produce this offset table and they are observationally identical:
1. records are genuinely variable-length, padded to 4 bytes, stride = largest
2. records are fixed at stride and the offset table carries inter-record
padding, so the gap to the next record is record + padding
Nothing in the offsets distinguishes them. Both say where records start. Both
produce 92, 88, 84, 92, 92, 92, 84, 88 for the case already discussed.
I wrote the change that slices each record against its successor -- model 1 -- and
then reverted it. Two things came out of writing it:
First, it broke three tests, which is the informative part. Those tests assert
"1,509 of 1,509 chunks have uniform record length" and they passed BECAUSE the
parser forced uniformity by slicing every record at a single stride. They were
asserting the parser's behaviour back at itself, and would have kept passing
forever regardless of what the disc contains.
Second, shipping it would have replaced one unverified model with another and
silently changed the record boundaries of 1,252 chunks on the strength of a
guess. The repo's own rule applies: a heuristic count is not an identification,
and here it is not even a count -- it is two equally good explanations of one
table.
WHAT WOULD SETTLE IT: a field inside the record whose end is independently
observable -- a terminator, a count implying a length, or a known value pinning
its position. Under model 2 the disputed bytes are padding and will be uniform or
zero; under model 1 they are data and will not be. For [75 42 ef] those 8 bytes
read a0 00 27 20 at both the tail of one slice and the head of the next, which is
arithmetic rather than evidence, since it is the same eight bytes seen twice.
The uniform model stays, the ambiguity is recorded in resolve()'s documentation
and here, and the 1,509-of-1,509 assertion stands as a statement about what the
parser does rather than about what the disc contains. 243 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
b9e9835639 |
A parser defect: one chunk type is not fixed-stride, and over-reads
[75 42 ef f2 71 b5] -- the signature that is both level-specific and
variable-length -- makes three chunks per level, 81 total:
stride 76 deltas [76, 76, 76, 76, 76, 76, 76, 76] uniform
stride 92 deltas [92, 88, 84, 92, 92, 92, 84, 88] NOT uniform
stride 104 deltas [104, 104, 104, ...] uniform
Identical on all 27 levels: 27 chunks of each shape, three shapes, no others.
So "this signature has strides {76, 92, 104}" is a misdescription. It has three
chunk shapes, one of which is not fixed-stride at all. The parser derives one
stride per chunk and resolve slices every record at that length.
The cost, precisely: records are read at the offset table's positions, so every
record STARTS correctly -- no misalignment. But each is sliced at 92 bytes, while
two of the eight are 88 and two are 84, so four records over-read by 4 and 8
bytes and carry the head of the next record inside them. Bounds checking holds;
this is a data-correctness bug, not a safety one. It is precisely the kind that
produces plausible-looking fields offset by four bytes, which is how it would
have gone unnoticed had the stride table not been collected per signature rather
than inferred per region.
It also explains why this signature was previously flagged as breaking the 8-byte
alignment rule: its records genuinely are not a uniform multiple.
The right fix is for resolve to use offset deltas for a chunk whose deltas are
not uniform. Not taken here -- the delta sequences are asserted first, so if the
data ever becomes uniform the test fails and asks for the model to be revisited,
rather than the model quietly absorbing the exception.
Asserted in chunk_type_75_ef_has_one_variable_length_chunk_per_level, with the
exact sequences [76x8], [92,88,84,92,92,92,84,88], [104x8] and 27 of each.
A near-miss worth recording, because it would have made a tidy false claim: the
leading u16 of these records appeared to step by exactly 0x50 (80) across all nine
of the 76-byte chunk. Reading at the real offsets gives b1a5, a529, 047e, a529,
04f0, a529, 0562, a529, 05d4 -- no progression at all. The apparent 80-byte step
came from comparing the wrong bytes, and "records are indexed by 80" is exactly
the kind of finding that gets reported on sight.
243 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
af62032ddc |
Most of the chunk container is level-independent: 65 of 74 signatures
Comparing every chunk record against the same record index in the other 26
levels, across all 74 signatures:
65 of 74 every record byte-identical in all 27 levels
9 of 74 level-specific
0 ragged no signature appeared with different record counts per level
The nine that vary:
[5d 44 13] [75 42 ef] [7f 44 ee] [82 c4 2f] [96 c3 37]
[a2 44 a5] [c6 c3 ef] [df c3 ea] [f3 c4 b9]
This reframes the container. 65 of 74 chunk types are static tables the game ships
unchanged in every level -- constants and lookups, not level content. The largest
by record count are all in that group: [38 c4 7e] stride 328 with 90 records per
level, [a9 c4 d8] stride 308 with 90, [3d c4 9e] stride 368 with 77.
The consequence is about method, and it is worth being blunt about. For a
level-specific type, comparing levels is the strongest lever available: a field
that moves with the level is identifiable by correlating it with something known.
For the other 65 there is nothing to compare -- one copy of the data, repeated 27
times -- so they must be decoded structurally, from what the bytes are. A future
decoder that reaches for per-level comparison as its first tool will find it
useless on the majority of the container.
[75 42 ef] is both level-specific AND the one signature with three strides. A
type that is both variable-length and level-varying is the most informative one
here, and is where per-type field decoding should start.
How this was found, including the wrong turn:
The first look at [53 c4 6c d7 9a 43] (stride 108) showed 45 of 108 byte positions
varying WITHIN a chunk and 810 records all distinct -- which reads like per-record
data. Counting distinct values per u32 slot across all 810 records contradicted
that: 29 to 30 distinct values across 27 levels, i.e. about one per level.
Both readings were wrong because they measure different things. The 45 positions
vary between records within a level; the u32 counts were low because the same 30
values recur in every level. Comparing record i across levels settled it: all 30
records are byte-identical in all 27 levels, and zero byte positions vary across
levels.
Had the u32 distinct counts been the only measurement, the conclusion would have
been "one value per level" -- the opposite of the truth, describing a per-level
parameter table that does not exist. The within-chunk and across-level views had
to be taken together to get a true one.
Asserted in most_chunk_signatures_are_level_independent. 242 tests, clippy clean.
|
||
|
ReaperOfVeriod
|
59dba3e58d |
Chunk records: the stride table, and the exception pinned by name
Walked every chunk on all 27 levels, grouped by six-byte header signature:
1,506 chunks, 34,415 records, 74 distinct signatures
55 distinct strides, 52..640, every one a multiple of 4
73 of 74 signatures use exactly one stride
The exception is asserted rather than skipped, because "every type is uniform"
is precisely the claim that quietly stops being true:
[75 42 ef f2 71 b5] strides {76, 92, 104}
That is the variable-length type the module docs had referred to indirectly. It
is now pinned by signature and to those three strides exactly.
The 4-byte alignment is a result, not an artefact. A stride recovered from one
chunk's first-to-second record delta can inherit alignment from where that chunk
happened to sit; collecting all 55 strides across every type and requiring all
of them to be divisible by 4 rules that out, because a single off-by-one in one
signature would show. Asserted in
chunk_record_stride_is_four_byte_aligned_and_uniform_per_signature.
What this settles is the shape the decoder must handle: uniform stride per type,
4-byte aligned, one named exception. Field offsets within a record are now the
same for every record of a type, which is what makes per-type decoding
tractable. The fields themselves are still unnamed.
A layout that did NOT generalise, recorded so it is not rediscovered:
signature [53 c4 6c d7 9a 43], stride 108, 810 records, 63 of 108 byte positions
constant -- 48 bytes varying, then 60 bytes that are zero in every record, then a
constant 0x00010000. That reads like an inline buffer this type never uses, and
would have made a tidy field story. Requiring `stride == live span + zero run`
across every type with 50+ records per chunk gives 0 of 9. The zero tail belongs
to that one type and means nothing general.
Third occurrence of one editing mistake, and it is worth naming as a habit
rather than an accident: inserting a test before an existing one strands that
test's #[test] attribute on the new doc comment and silently unregisters it.
It has now happened three times, each time surfacing only as clippy's "never
used" and never as a missing test. Fixed by appending new tests at the end of
the file instead of anchoring them to an existing function, and verified by
listing registered tests rather than trusting a count.
241 tests, clippy clean, all five recent tests confirmed registered.
|
||
|
ReaperOfVeriod
|
28ab7e094b |
SBlk head characterised: four constants and one word that is not a count
sound_bank's head was parsed but never characterised. Across all 27 levels the
SBlk region opens identically for its first twelve bytes:
+0x00 "SBlk" +0x04 1 +0x08 4 +0x0c "DAW\0" +0x10 0
and +0x14 is the only word that varies. It varies with its low sixteen bits zero
on all 27, leaving 0x0092..0x0103 -- 146 to 259.
That is the shape of a record count, and it tracks the body length at close to
100 bytes per unit. It is not one:
body / field_14 91.2 .. 110.8 across 27 levels
any stride 0..80 with body = stride*word + constant exactly: no
So the test asserts the negative too. If a stride ever makes the relation exact
on all 27, sound_bank_sblk_head_is_constant_apart_from_one_word fails and asks
for the record size to be revisited. A near-miss at 100 bytes/record is exactly
what hardens into a name the next time somebody glances at it; the constants are
also checked individually so a real deviation names which one moved.
The DAW region begins with sixteen zero bytes then block-structured data with
16-byte periodicity and slowly varying nibbles -- consistent with block
compressed audio, not decoded. Recorded as such.
Two mistakes, both mine, both the same shape as earlier ones:
1. The constructor bound `format_tag` with a `let` BEFORE the two words that
precede it in the file, consuming bytes 4..8 and shifting every field after
it. The struct literal listed the names in the right order; the statements did
not. A scratch script reading the same bytes gave the opposite answer, which
is what exposed it. Fixed, with a comment, because "read the lets in file
order" gets undone otherwise.
2. Inserting the new test ahead of an existing one stranded that test's #[test]
on my doc comment -- silently unregistering
tie_vertex_element_is_16_bytes_with_fixed_texcoords. This is the SAME mistake
as two commits ago and I made it again. It again surfaced only as clippy's
"never used" rather than as a missing test. Restored and verified by listing
the registered tests rather than trusting the count.
240 tests, clippy clean, both tie vertex tests confirmed present.
|