The URL is public, so this is the first thing a stranger meets.
ENTER, PINNED. The exploit shipped because the rule lived in a keydown closure
inside boot(), behind a canvas and a WebGL context — nowhere a test could
reach. That's WHY it went out: not carelessness, no seam. `enterCommits(phase,
cardOpen)` is a value now, asserted exhaustively over PHASES so a sixth phase
is covered the day it appears, and the safe default (may not advance) is the
one it gets. Mutation-checked: reverting the guard reddens forecast, storm and
aftermath.
SPLASH. A cold visitor used to land on a job sheet — an invoice-shaped card
from a business they'd never heard of, quoting money for a job nobody had
explained, controls revealed later on prep's help line. It's a good SECOND
card. One more .card, not a menu system: there is nothing to configure. Built
on E's letterhead because the masthead IS the game's name. boot({splash:false})
for benches and D's harness.
P PAUSE — and the test is the story. The pause lives in frame(), frame() is
only called by rAF, and rAF DOES NOT FIRE IN A HIDDEN TAB. My first probe drove
the real loop and reported success: simT advanced 0 while paused... and 0 while
running. It was measuring a frozen tab, and it would have passed no matter what
I wrote. So the rule is a value: accumulate(acc, raw, paused). Paused DRAINS
the accumulator rather than carrying it — a kept `acc` spends time that elapsed
while you were paused on the first frame after resume, which is a free sixtieth
of a second of storm and exactly how two runs of a deterministic sim come to
disagree. Storm-only, and cleared at every phase change: a pause flag left set
where P is inert would be a soft-lock built out of a comfort feature.
M MUTE — wired, honest, and not advertised yet. C's bus has no tap: createAudio
holds `master` inside skyfx's closure and exposes only unlockAudio(). Asked in
THREADS. `sky.setMute?.(on)` behind a key the splash promises would be D's
night-3 soft-lock with a keycap on it — `rigging.setWorld?.(world)` sat in this
same file for a sprint doing nothing, silently, because ?.() on a missing method
looks exactly like a call. So the UI asks before it advertises, and M lights up
by itself the day C lands the tap.
TOUCH NOTICE. The naive check is the bug: `(pointer: coarse)` asks what the
PRIMARY pointer is, so a touchscreen laptop answers coarse and gets locked out
of a game it can play — worse than the dead canvas it replaces. `(any-pointer:
fine)` asks whether a mouse exists at all. Detection errs toward letting people
in; unknown browsers get the game, not a lecture.
DEV LINE. Out from under B's anchor panel (both were top-left, typesetting on
each other), and off unless localhost or ?dev=1 — a stranger on partly.party was
reading "60 fps · storm 12.3s · debris 4" on the front door.
selftest 338/0/0 (was 335).
Gate 3.3: collateralFor('gutter') returns {90, 'the gutter'} — canonical in
backyard_01.json beside the reasoning, GLB extras as fallback, per E's
collateralKey chain. House wreck wired on the carport's swap contract
(no-op until lane/e's GLB merges; window_glow state copied). site_02 stays
null: no house, no gutter. hud.js bars + main.js verdict now read
rating × ratingHint, matching sail.js's failure line (B's flag).
selftest 319/0/0.
Gate 3.2: bike_kid_01_v1 on backyard_01's south fence, rotYDeg 180 so the
baked -Z lean lands ON the fence — bars measured to the rail (0.286 m
reach, z 9.69 -> grip at 9.98), tyres 0.31 m clear, between posts. Unpriced
by ruling: the sim can't knock a bike over, and billing an unseen event is
the lie the invoice kills; the windward-fence pose is the honest one.
a.test pins lean direction, standoff band, and the ruling. 318/0/0.
audit.mjs's hand-typed backyard dump goes full-precision: the 2-dp rounding
moved t2 ~5 mm against t1's shelter cone and swung its ice-night peak
4.8 -> 9.3 kN — the node and browser front-ends disagreed by one whole
winnable line until the positions matched to the float. Reconciled by
measurement (node on full floats reproduces the browser exactly).
Re-audit verdict (THREADS, full tables): every night still passes, no retune
proposed — E's numbers stand. backyard: hints move exactly one number (night
2 honest $30 -> $40; fascia appears in zero bed-covering quads so its 0.35
is audit-invisible). site_02: 64/66 -> 23/66 affordable, every cheap carport
line dead, cheapest now $40 honest steel, cheapest >=70% line $45 through a
$30-rated carport post — the trap is priced, not fenced off.
Selftest 319/0/0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Gate 3.1: the CUT block from tools/jobsheet verbatim — letterhead, billto,
brief, sched, .void, AMOUNT DUE, terms. Docket numbers shared across the
pair (JOB/INV 1040+night); week.js nights carry an addr. Cards scroll when
taller than the window (safe center + overflow). Both rulings adopted and
pinned by DOM asserts; selftest 317/0/0.
Every windchange event gets an in-world tell: a dark bank standing on the
horizon in the quarter the new wind comes FROM, rising across the 12 s
before the change and clearing after the swing, plus a low distant rumble
layer in the audio. Soft-edged via vertex alpha (the hard band read as a
floating rectangle on the bench). Deterministic off (dt, t): progress is a
pure function of t, azimuth comes from dirAt at a fixed post-change
instant. Night 2 teaches the tell at ~18 s; night 3 shows the same wall at
t=6, which is 'looks like night 2 and isn't' made visible.
dev_skyfx.html: the skyfx bench — one storm, bare ground, scrubbable
clock, so the tell can be judged by eye without playing to night 3.
Two new asserts in c.test.js (window, monotonic rise, direction, mesh
orientation, changeless storm stays clear, bit-exact determinism).
Selftest 322/0/0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Five new asserts (Lane E 67 -> 72; selftest 315 -> 320):
- price exists, sits between gnome and carport, keyed to the anchors'
collateral string, wreck carries the same bill and names its twin
- the torn gutter is DOWN in the wreck and UP in the intact house — negative
control run: rebuilt with the ground run at eave height, selftest went red
('gutter_down tops out at y=2.68', 1 failed / 319 passed) while the Blender
verify PASSed the same bad file (plan position unmoved, no box span changed);
restored build round-tripped to identical hashes
- wreck height equals house height (a house does not fall when its gutter does)
- no fascia_anchor_* survives into the wreck
- house pair added to the shared ground-plane test
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The fascia anchors have said collateral:"gutter" since Sprint 6 and nobody
priced one — collateralFor('gutter') returns null, backyard_01's house is a
free failure. Proposal rides in the asset like the carport's did: 90, between
the gnome (25) and the carport (180), with the band reasoning next to the
constant. A rules the number.
house_yardside_wrecked_v1: same origin/footprint/facade (shared _house_facade
helper — intact house BIN chunk byte-identical to the committed GLB, only the
three new price extras differ), fascia torn through the anchored span, left
run hanging by its last bracket, right run kinked ON the grass, downpipe out
of plumb, offcuts flat on the lawn. No fascia_anchor_* empties survive — you
cannot re-tie to a ripped eave.
collateral_key is new and explicit: unlike the carport, the structure and the
thing you take have different names, so the key names which collateral string
the value prices.
All 33 GLBs byte-identical across two full factory runs; the house pair also
byte-identical between --only and full runs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
sail.js _checkFailure now fails a corner on load > hw.rating * anchor.ratingHint
(read live from the anchor — dress() mutates in place). Consequence asserted,
not the formula: on site_02's carport line (uniform rated shackles, funnel on,
12 s calm settle) a cb corner blows before any q corner — at DEFAULT tension
(t~29 s into the storm, honest posts hold) and at max. In-suite control pins
the attribution: the same rig with every hint forced to 1 holds 4/4, so the
break is the hint, not the load. Mutation-checked: unwiring the line fails both.
rigging.js weak-link (D's #7): summary.weakest reduces on rating * ratingHint
via _effRating, not bare hw.rating — uniform hardware now flags the genuinely
worst steel instead of the first click. Mutation-checked.
site_audit prices the anchor, not just the steel: tierFor(peak, hint) needs
h.rating * hint >= peak, both front-ends carry ratingHint through (browser off
world.anchors, node dump extended with the live-dumped hints), rows print the
hint. sweep.selftest asserts hints reprice without touching peaks and land
over-ceiling corners in unholdable. Mutation-checked.
Selftest 319/0/0 (315 + 4).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Second harness on B's post-ratingHint audit: measures what the STORM
delivers at every dressed anchor — peak speedAt, tPeak, dynamic pressure
(downdraft folded back in), dose, and Pa/hint (the wind-side failure-order
prediction under load > rating x ratingHint). No cloth, no rig, no tension:
independent by construction.
Browser front-end dresses the yard the way the game does (createWorld +
dress) and wires venturi + tree shelters exactly as main.js:446-447 does.
envelope.selftest.js wired into c.test.js, every assert written to fail if
a wiring line is deleted (sweep.selftest's pattern). Selftest 320/0/0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
C AND D BOTH LANDED WHILE I WAS IN GATE 2, AND BOTH CORRECT ME.
VENTURI — reverted to axis 2.1 / gain 1.5. C reached the same physics I did,
independently (|dot| means an axis is a line, mod PI), and then went one step
further and caught my reasoning: -1.08 was never a proposed AXIS. It's
dirAt(40), the southerly's HEADING once the change settles. 2.1 is the GAP'S
ORIENTATION — site geometry, which doesn't move when the wind does. They're
different quantities that sit a hair over PI apart only because the gap was
BUILT along the southerly.
So "measured beats authored" was the wrong rule for this question: I'd have
swapped a geometry fact for a weather fact because they nearly coincide, and
the site's axis would silently mean the wrong thing the day someone authors a
storm from another quarter. "A number gathered from the wrong harness" caught
both of us on the same question from opposite sides in one sprint.
Gain reverted too, and that one I simply jumped: it's a balance lever, B's audit
calls it, and C's standing offer is to drop to 1.35 only if no $80 line survives.
B hasn't run. D then measured the site winnable off honest anchors at 98% — so
the funnel isn't killing the line and the gain stays.
D's LANDMINES, both mine, both in gate 2's path:
· CALM_STORM loaded only because night one HAPPENED to be storm_01_gentle.
Gate 2 rewrote every night entry. Drop gentle and windTime() throws every
non-storm frame on "Cannot read properties of undefined (reading 'duration')"
— naming neither storms, nor calm, nor the week. Now explicit in the preload,
and the assert calls stormsToPreload() with a gentle-free ladder, because
asserting it against the shipped NIGHTS passes either way. That's how it hid.
· boot({site}) was a documented lie that failed INVERSELY — markers on the
requested yard, world on the week's. Honoured through the week now (seek to
that site's night), because a site without its night is half a game: no
client, no brief, no storm, no bank. Plus opts.bank, so D can boot the corner
block at the $237 they measured instead of night one's $80 and OFF THE JOB.
RATINGHINT — D is right that nothing reads it, and the naive fix is worse than
the bug: honest posts are built from site JSON, never touch adoptAnchor, so
their ratingHint was undefined — and `load > hw.rating * undefined` is
`load > NaN`, always false. The strongest anchors would have become UNBREAKABLE
while the carport became the only thing that can fail: the trap would look like
it works, for the wrong reason, and read as a successful playtest. Every anchor
carries a real number now (DEFAULT_RATING_HINT), pinned, so the wire is safe
before it's made. The wire itself is sail.js — B's — and the ruling is in THREADS.
selftest 306/0/0 (was 304).
site_audit's headless false-negative is fixed by the browser front-end
(A's loadSite/createWorld landed in main — that was all it waited on), and
fixing it surfaced a worse bug underneath: the sweep built its wind from the
STORM def alone and never called setVenturi, which main.js:424 does at every
site load. The venturi is SITE data, so every site_02 audit ever run — including
the numbers quoted at C in SPRINT10 — was of a corner block with no gap. The
SPRINT6 p1=7.4kN failure inverted: there the tool called a good site unriggable,
here it called a mean site cheap.
- sweep.js: setVenturi on both winds (the gap doesn't switch off for the settle);
both front-ends pass the site's funnel and print it, so a funnelled run is
legible as one.
- sweep.selftest.js (new, wired into b.test.js): the auditor gets an auditor.
Funnel must strictly raise every corner; a funnel-less site must be
byte-identical. Verified it fails when the setVenturi calls are deleted.
- rigging.js: setWorld(world) + session.setBudget(n)/setAnchors() — A's two asks.
Markers are rebuilt (the anchor set changes across sites) and disposed; stale
picks die with the old yard. Deletes A's "one private touch" into this module.
- dev_rigging.html: dead since SPRINT10 (called createWorld without a site, which
now throws) and nothing noticed, because a dev page has no selftest to go red.
Fixed, dressed, and given the site switch it should have had (N = next yard).
Verdict for C/A, in THREADS: site_02 PASSES with the funnel on — 64 of 66 lines
affordable, cheapest $20. The funnel kills nothing, so the gain does NOT need
dropping and E's tree stays put. It's a reach problem, not a gain problem: the
throat's radius is 5 m and the bed centre is 6.08 m away, so the gap screams over
the carport at +50% and reaches the garden at +0.0%.
selftest 300/0/0 (296 + 4 new).
The letterhead is the point: one business, two documents, same masthead, so
the job sheet and the invoice read as a pair of papers off the same outfit's
printer rather than two game screens. DESIGN.md has opened with "you run a
small landscaping outfit" since day one and it has never been on the glass.
tools/jobsheet/ — paste-ready for Lane A between CUT markers, purely additive
(.card/.row/.ledger untouched), rendered to four PNGs before handover, each
state addressable as ?v=. The ledger already WAS an invoice; it just never
said so. The line worth keeping is .void: a bonus that was on the job sheet
and didn't survive the night, struck through in the same column as the money
that did — a bonus you never watch fail is one you never learn to protect.
Two design calls left as A's with the reasoning beside them, same as the
carport's $180: the business name (proposal: HARD YARDS) and the ABN
(sequential digits on purpose — a plausible one could be a real business's).
bike_kid_01.glb — 1.15 x 0.33 x 0.74 m, 700 tris, lean baked in because a
bike doesn't stand up on its own. Two bugs caught by LOOKING at the render,
neither catchable by the existing suite: a comment claiming a step-through
above a diamond frame, and a fork with negative rake that drew head tube and
fork as one pole through the front wheel. Both passed the Blender verify —
a bounding box can't answer "is this the right shape".
The lean is an export-time axis fact (+Y in Blender is -Z in three.js), so
e.test.js now pins it in browser coords. Proved the assert fails: rebuilt at
lean=0 and Blender still said PASS while the selftest went red.
selftest 299/0/0. 32/32 GLBs byte-identical across three full factory runs.
DESIGN.md has opened with this since line 3 — "every callout is a different
site, a different client" — and the game had the site and not the client. The
forecast card is a job sheet now and the aftermath is an invoice.
THE WEEK, once it was asked to name a client, mostly wrote itself: four nights
are the same yard, so they're one client on retainer while they're away (the
Hendersons, in Cairns from Friday — SPRINT11's own line). Night 3 is a
different SITE, so it is necessarily a different CLIENT: a short-notice callout
to a corner block you've never rigged, on the night the change comes early.
That isn't flavour bolted on, it's what `site: site_02_corner_block` already
meant, finally said out loud.
BUDGET Y IS THE FEATURE, not the letterhead. DESIGN.md's brief is "client wants
X, budget Y, forecast Z" and this card had Z alone. The fee has existed since
Sprint 8 and the player has never seen it until the money was already spent —
you chose what to rig without knowing what the job was worth, which is a reveal,
not a decision. The sheet quotes base + garden + clean up front, and a test pins
that a perfect night pays exactly what the sheet advertised: a job sheet that
over-promises is the game lying on paper.
THE CLEAN BONUS is the new money, and it's the one number in week.js that isn't
measured. It gives the invoice three axes that mean three different things —
you turned up, their garden lived, you didn't wreck the place — and it's NOT
gated on winning, because a lost night where you broke nothing of theirs is a
different fact and deserves its own row. It also makes the carport bite twice:
180 + 20 = 200, inside E's own "over ~250 is a silent game over" ceiling.
Measured: it adds ~$100 to a competent week (385 -> 485), which makes
BROKE_BELOW's known runaway worse. I did NOT pay for it by quietly trimming
feeFor or gardenBonusMax — those carry measured evidence, and re-tuning someone
else's evidence to fund my own feature is the move this repo keeps catching. If
gate 3 says the bank runs away, this is the lever to cut first and it's data.
TWO BUGS, BOTH CAUGHT BY LOOKING AT THE CARD, neither by a test:
· "THE HENDERSONS" over "The Backyard — Your own place". The blurb and the
letterhead called each other liars on screen. The blurb gave way: "your own
backyard" is a leftover from when the game had one yard, and DESIGN.md never
wanted it. The line's job survives — night 1 is the tutorial because you know
this yard, now from having worked it for years rather than owning it.
· C's `lead` is a 0..1 haze dial (confidence = 1 − lead), not "nights out". I
read it as nights, passed 1, and the sheet advertised "forecast confidence
0%". Now 0.6, because C's own hail rule calls L > 0.55 too distant to promise
ice, and tomorrow is exactly that. C: the slot is mine, the number is yours.
Lane C's forecastLines(def, lead) finally has a caller above 0 — two sprints
after they built it. Lane E: .letterhead / .brief / .jobsheet / .tomorrow are
structure only and yours to style.
selftest 304/0/0 (was 300).
Played lane/a 0f288be. Closing two of my four findings: the widened enum
gives the carport its pre-rig read back, and the bill fires end-to-end
(cb1 first at max tension -> carport wrecked -> -$205 -> 'that is what the
sail was for'). Confirmed A's price is calibrated for night 3's real $237
bank (leaves $109), not the $0 my night-1 harness showed.
Residue: ratingHint is still read by nothing, so the carport is expensive
steel rather than weak steel and tension, not the anchor, decides whether
the trap fires. Needs a ruling since E's tests pin those ratings.
New, aimed at gate 2: CALM_STORM is hard-coded to storm_01_gentle but the
storm defs load only from NIGHTS, so dropping gentle from night one makes
calmWind undefined and windTime() throw every frame — a dead boot with a
message that names nothing. A is about to rewrite every night entry.
Night 3 remains soft-locked; lane/b has not moved.
Gate 1.3, four sprints owed. Played the real path (nights 1-2-3, mouse).
The corner block is a good yard; two gate-1 blockers stand in front of it,
both measured rather than reasoned:
- Night 3 is unplayable. rigging.setWorld is undefined, so main.js:450's
optional call is a no-op and createRiggingUI's markers (rigging.js:254)
stay bound to the boot-time backyard. On night 3 the world/session/rig
re-point correctly but the only clickable markers are the backyard's, at
backyard coordinates. Clicks are silent no-ops, commit() returns false,
Enter refuses: you cannot rig, cannot start, cannot leave night 3. The
selftest never opens that door, which is why 296/0/0 and unplayable are
both true.
- The carport trap has no teeth. anchor.ratingHint and anchor.collateral are
written by world.js:556-557 and read by nothing; sail.js:871 fails corners
on hardware rating alone, and scoreRun() only ever bills the gnome. 46 m2
hung off the carport through the funnel pays +$82 at garden 83%.
Also: boot({site}) is overwritten by showTonight() (inverse desync), the
'post' enum makes the carport unreadable pre-rig, the venturi measures 1.5-1.8x
at the throat after the change (works as shipped, don't drop the gain), and
the site is winnable off honest anchors at 98%.
THREADS-only; no code touched on lane/d this sprint.
The corner block's seams, all of them A's to close.
ENUM (D's flag). Anchor.type was a closed JSDoc enum 'house'|'tree'|'post',
so site_02 typed its carport 'post' to fit — quietly enrolling it in the
sail-post family that C's venturi and B's audit both read. Widened to E's own
asset strings (carport, carport_post) and, more to the point, CHECKED: it was
documentation for ten sprints, and documentation cannot fail. validateSite now
rejects an unknown type by name. The ladder is untouched either way — D keyed
it on `work`, the mechanism, which is exactly why that was right.
CARPORT (E's proposal, adopted). $180, with E's reasoning intact: gnome 25, a
night's budget 80, so 180 is 2.25 nights — a good week turned broke and felt
for the rest of the run, without ending a strong one outright. Under ~120 it's
a shrug, over ~250 a silent game over that teaches nothing.
The number lives in site JSON, not the GLB: sites are data, and what a carport
COSTS is this site's economy, not a property of the mesh. E's baked value stays
as the fallback and the proposal.
And it BILLS now. E shipped the trap and said it plainly — the anchors said
collateral "carport" and nothing said what one cost, so nothing scored it. You
could lose the worst steel in the game and pay for a $15 shackle. One broken
corner is enough (the gnome needs two: it needs the sail to land on it; the
carport doesn't need the sail at all), priced per structure so two beams is one
carport gone, not $360. The wreck swaps on the same event — E built it to the
same origin for exactly this.
VENTURI (C's question). There was no disagreement to settle. weather.core
aligns on |dot(wind, axis)| because a gap funnels either way through it, so an
axis is a LINE: -1.08 + PI = 2.0616, and my shipped 2.1 was the same gap read
from the other end. Took C's number because C MEASURED it off storm_03b and I
eyeballed mine at authoring time; the 2.2 deg of my eyeballing was worth
0.44 m/s. gain left at C's 1.35 — that's a balance lever, theirs with B's audit.
Also wired validateSiteWind at site load, C's ask from Sprint 10.
selftest 300/0/0 (was 296).
Answer to A's blocking gate 1 question, by measurement. venturiFactor
aligns on |dot|, so an axis is a line defined mod pi. 2.1 - (-1.08) =
3.1800 vs pi = 3.1416: the two are 2.2 degrees apart, the same gap read
from opposite ends. Measured at the throat through all of storm_03b,
axis 2.1 peaks 33.5061 m/s and axis -1.08 peaks 33.5329 — 0.08% apart,
same gust, same second. Ship 2.1 as it stands.
The error was mine and it is the repo's own lesson again: -1.08 was
dirAt(40), a wind DIRECTION, never a proposed axis. I compared it to a
gap ORIENTATION and called it a conflict. A number gathered from the
wrong harness, fourth time.
Also posted for B: gain 1.35 vs 1.5 measured at the throat (30.16 vs
33.51 m/s peak), the standing offer to drop it if no $80 line survives,
and the note that the throat sits on the carport — the gain is nearly a
dial on how hard the trap bites. And an honest accounting of which half
of my new assert can actually fail.
forecastLines(def, lead) always took the param; what was missing was the
map from "this night is N nights away" to a lead. leadFor(nightsOut,
weekNights) is that map, linear onto forecastFor's documented 0..1 domain
so tonight is exact and the week's far end is lead 1. Tomorrow at a
five-night week reads 0.25 — 75% confidence, sustained band ~±8%.
hud.js is Lane A's; this only exports the argument.
The new assert pins the property that makes a band safe to print: it
RESOLVES. lo rises and hi falls monotonically onto the truth as the night
approaches, because the centre-wander is the same seeded draw at every
lead — so tomorrow's band always contains tonight's and the card never
jumps. Verified against a mutation that reseeds per lead (caught).
Its containment half is honestly decoration today and is labelled as
such in the test: 0.6-of-half-width wander means c±w never crosses v, so
deleting the clamps does not make it fail. Kept as a contract guard for a
future rewrite of band(), not counted as coverage of today's code.
Selftest 298/0/0.
Gate 2's actual deliverable. The node tool cannot dress, so it can only audit
the built-in snapshot or a resolved export and REFUSES A's dress-source site
JSON. This is the honest audit: audit.html builds the site the way the game does
— createWorld(await loadSite(name)) then dress() — and reads world.anchors, the
single source of DRESSED positions (GLB fascia + branch anchors, 8°-raked posts,
all of it), then runs the SAME sweep. It is the only way to audit a site with a
carport (site_02) before it ships.
tools/site_audit/audit.html?site=backyard_01
tools/site_audit/audit.html?site=site_02_corner_block&storm=storm_03_southerly
The winnability math now lives in ONE place, sweep.js, imported by both
front-ends — a tool built to catch reimplemented-formula drift doesn't get to
keep two copies of its own sweep. audit.mjs is refactored onto it (output
byte-identical); the node refusal message now hands you the exact audit.html URL.
Validated in-browser against backyard_01: dressed 12/12 anchors (GLBs loaded),
PASS, best line p1,p2,p3,p4 $65+$15 spare — identical to the node snapshot
verdict and to balance.test, off A's real data with no snapshot. That
cross-agreement also re-proves A's extraction is byte-identical.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
