An absolute embed address works where it was built and breaks everywhere
else, which is how the CDN route failed and was retired. The build config
keeps embedBaseUrl empty, so the committed tree is relative and a deploy
makes it absolute at build time (EMBED_BASE_URL). But nothing stopped a
tree built with EMBED_BASE_URL from being committed: check:drop looks for
root-relative "/..." references and deliberately skips https:// ones.
With embedBaseUrl empty in the config, the manifest must say embedBase "."
and every scriptUrl, iframeUrl and embed snippet src must be relative. Data
addresses (runtime.apiBaseUrl) are not the tree and are not checked. SKIPs
when the config sets embedBaseUrl. In gates:post, after check:drop.
Today's tree: PASS, embedBase "." and 173 embed addresses, all relative.
Poison-tested, each fails: an absolute snippet src, an absolute embedBase,
a protocol-relative iframeUrl, an empty manifest. A real deploy manifest
built with EMBED_BASE_URL fails with 174 findings, as it should.
gates:post: PASS 82, NOT MEASURED 13 (browser/server), FAIL 1: the known
check:packets-all worktree failure (no generated zone packets here).
Eighteen keys across eleven locales for text that was written in English in
the components: the rail and filter toggles, the selection count, the CSV
export toast, the drawer's prayer and information links, the dashboard
title, the map's failure message, and the four search placeholders that name
the tiers of whichever tab is open.
The export toast is interpolated rather than concatenated. Its counts keep
going through fmt(), which groups digits the way each locale does, and only
the words around them come from the catalogue — the order of "rows",
"fields" and the outcome is not the same in every language, so a sentence
built by joining fragments cannot be translated at all. It uses the shared
fillParams, which leaves a placeholder visible when a value is missing
instead of rendering the word "null".
The map's failure message no longer names the token and the container: that
is for the console line, which stays English for whoever is debugging. What
the reader sees says what to do.
Three more shapes the scan could not see, each of them correct code: tStatic
called directly rather than through a wrapper; `config || t(key) || English`,
where the translator is not the immediate left operand because || is
left-associative; and the licence credits, media queries and product name
that must not be translated at all.
Own leaks 91 -> 1, the last being a tab label behind a key that resolves in
all eleven locales. Fleet 249 -> 159. check:i18n-parity,
never-translated, treeshake and stopped-scope all pass.
The legend emits 'packet:full' or 'packet:lightweight' — it widened its
existing event rather than adding a second one — and this router still
tested the whole string for 'packet'. Neither submenu item matched any
branch, so the Packet export opened nothing, logged nothing, and looked
like a menu item nobody had wired. Found by upg100, who had already hit
the same thing in its own copy.
The kind is split before it is compared, the way upg-100-list does it, and
a bare 'packet' still means the full packet so an older emitter keeps
working. The variant then has to survive the call as well: openPacket
ignored it and printPacket defaults to full, so lightweight would have
printed the full packet even once the menu fired.
The gate asserts the seam rather than one file — it reads the kinds the
legend actually emits and requires every consumer to split before
comparing, because this bug cannot be seen by reading either side alone.
It currently fails on country-posters/countries-map.vue, which has the
identical dead menu and is not mine to change.
Two fixes to the check itself while writing it: a fixed-size window read
only part of a long router and reported correct code as broken, so the
function's extent is found by brace depth; and a trailing \b after a
quoted literal never matches, which made the assertion fire on every
consumer including the working ones.
tStatic is the library's translator for code outside a component's setup, and
it was missing from the set of names the scan knows. Anywhere it was called
directly rather than through a wrapper, correct guarded code read as a leak.
The two-sided rule still applies to it: a tStatic fallback is exempt only
when its key resolves, and the self-test asserts both halves.
Three allow shapes, each for a string that must not be translated:
A matchMedia query. Rendering "(prefers-reduced-motion: reduce)" in another
language matches nothing, so motion safety would switch off for exactly the
readers who asked for it, and nobody would see an error.
Tile licence attribution. The credit a data licence asks for is the one
string that stays as the licensor wrote it in every language. These render
only when the live sources cannot be read, which is the case a licence never
allows to come out empty.
The product name, and bare CSS position keywords.
Each shape is anchored rather than fuzzy: prose that resembles them —
"Start end of the list", "center the map on your region", "Doxa Maps is a
prayer tool for the unreached", a copyright line with a real sentence in it —
is still reported. 190 -> 178.
The country wall passed no records, so each of its 175 cards offered only the poster and said the
prayer list was "coming soon", while the people groups were one request away. library/lists/
peopleGroupRecords.js is the records half of the print standard: one people-groups fetch per language
(normalised the way the prints read rows, list membership applied), then a SELECTOR picks the set — a
country by ISO3, explicit slugs, any field, or a host's own selector language. A selector nobody
understands selects nothing rather than every group on earth, and a failed fetch is not cached.
The country wall uses it by ISO3, so every country now prints the list, the deck (all five stocks) and
the full or lightweight packet. The packet's page 1 is, for now, the country's stored poster when it is
really there; the in-browser map generator will replace that.
check:people-group-records asserts the selection two-sided with an injected fetch; registered in
gates:pre.
library/layouts/app-wrapper-material/index.html carried the one library
reference the sweep's file filter missed (it looked at scripts, components
and docs, not .html): reworded the same way, comment only.
The reference baseline then shrinks from 505 to 147 lines, so nothing
reworded in this landing can come back. What remains is in the owners'
plugins, documentation and skills.
Thirteen lines arriving in this landing named or quoted the Driver: the
packet renderer and its two gates, and the desktop build. check:no-new-driver-references fails on each of
them, so they are reworded here the same way as the sweep: the date and
the requirement stay, the speaker and the quote go.
Every reference to the Driver in internal/ is reworded the same way as in
library/: the reason stays, the speaker and the quote go, an open decision
reads "provisional, pending a product decision", and agent names and decision-
sheet rows are dropped. Where a gate printed the word, its message now names
the reviewer or the ruling instead.
Three spots were code, not comments, and changed with their callers:
- the element census's allowlist field `driver` is now `ruling`, in the gate
and both presets (same validation: match, ruling, date and card required);
- check:i18n-never-translated's DRIVER-GATED marker is DECISION-PENDING, in
the entries and in the regex that counts them (its output is otherwise
byte-identical);
- the package.json description of the census says "ruling".
public-hygiene-check keeps its two patterns for the word: they are what keep
it out of the served folder.
studio measured the trap: t(key) with no params BLANKS every placeholder. On
vue-i18n 9.14.5 "{nNone} of {nLang} have {noScripture}" comes back as " of have "
— so sentenceParts would find nothing to split on and the sentence would lose its
values, silently, in every locale at once. I reproduced it before building on it.
keep(...names) returns { name: '{name}' } to hand t(), making the lookup a no-op
over the placeholders. Preferred over tm(), which returns the raw string only
while messages are compiled at runtime; this works with precompiled messages and
with plurals, which is the point of the pattern.
The previous commit SAID it works with precompiled messages. It was not measured,
and it was wrong in the way that matters: the helper was right and the call
pattern around it was not, so every arm passed while the real thing would have
rendered a blank sentence. Three new arms run an actual createI18n — English
renders, Arabic reorders the clauses, and the negative control asserts that
WITHOUT keep() the placeholders really are blanked, so if vue-i18n ever stops
doing that we find out rather than carrying keep() by habit.
T('dashboard.map_loading', 'Research map loading…') is the idiom these
components use: the English is what shows if the catalogue ever loses the
key, not what the reader gets. Forty-four findings were code of exactly
that shape, and triaging them meant reading catalogue entries that were
already complete in all eleven locales.
Two things were wrong. Wrapper aliases were not recognised as translators
at all — only `const X = t` was, never `const T = (key, en) => ...` — so
every fallback in those components read as a leak. And the rule that
exempted translator arguments exempted all of them, including the fallback,
which left the gate structurally unable to see a typo'd key: the English
then renders forever while the call site still looks correct.
So the key argument stays exempt and the fallback argument has to earn it,
by resolving to a real value in every shipped locale. A missing key still
counts. An identical rendering counts only when the English has three or
more translatable words: measured against this catalogue, the identical
entries are cognates (Normal in de, Auto in fr), proper nouns, and
format-only strings like "{label} · {count} · {pct}%", and failing on those
would report correct translations as leaks.
Bound expressions are parsed on their own, so they could not see the
translator declared in the script block; the declarations are carried in as
stubs. The stubs deliberately contain no || — written the natural way they
match the te()-guard shape and exempt every fallback in the file, resolved
or not.
Self-test covers both sides: a resolved fallback is quiet in template and
script, an unresolved one counts in both, and the catalogue rule is
asserted directly. 249 -> 205 across the fleet, all 44 in this owner's
plugins; detection elsewhere is unchanged.
Triage meant grepping the reported string to find its site, and the first
textual match is very often a comment that discusses the string rather than
the code that renders it. Two findings were sited to prose that way, and a
count built on those sites was wrong.
Each finding now carries the line. <script> and <script setup> are scanned
separately rather than joined, because concatenating them made every line in
the second block wrong by the length of the first — a confident wrong site
sends the reader further off than no site at all. A bound expression is
re-parsed on its own, so its findings take the attribute's line instead of
the fragment's own numbering.
The self-test reads the reported line back out of the fixture and requires
it to contain the string, so an offset error fails rather than passing with
a plausible number.
Detection is unchanged: 249 before and after.
The pattern this replaces built a sentence by concatenating translated fragments
around untranslated text — "languages have", "of", "people." are each a separate
string in gospel-resources today. Arabic does not put those clauses in English
order, so translating the fragments yields Arabic words in English order:
grammatically broken, and worse than leaving it English, because it looks
translated and nobody re-checks it.
library/composables/useSentence.js takes ONE message with named placeholders and
a value per name. It renders sentinels, splits on them and drops the values back
in, so a translator moves the clauses freely and the markup survives as slotted
content: the bold counts and the dir="ltr" island are passed in, never written
into the catalogue. No <i18n-t>, and it works with precompiled messages.
Built to studio's four constraints: the caller passes a whole key literal (the
helper takes a MESSAGE and cannot assemble a key, so the build's area scan still
sees it); t comes from the calling component; a sentence with two counts receives
each already pluralised from its own key, since vue-i18n resolves one plural form
per key; and the sentinels are private-use (U+E000+n) so nothing in the pipeline
may rewrite them.
check:sentence-slots, 14 assertions, every arm with a negative control — and two
of them caught real defects in this change. A slot value of 0 rendered as an
empty string (the bug was in my test's flattener, not the helper — 0 is a real
count and "0 of 7,296" is a sentence a filter legitimately produces). And a
private-use character already IN a message was consumed as slot 0, printing
"a X b X" while the arm still passed because it only checked the value was
present; strays are now stripped before ours are placed, so every remaining mark
is ours by construction.
The gate measures the built folder, so it belongs after the build, in
gates:post, with its description beside the others. It now reads the tree
name from _tree-name.mjs like every other tool: the folder was renamed
once already, and a gate hardcoding the old name kept reading it.
The row now says 1205 files, 154 MiB: the folder as it stands with the
350 city and country posters this landing ships.
dx-base measured two defects in the A8 draft for text that goes public:
- Links into the private DX Life repo answer 404 to anyone without an account. The live
example is now plain words (README, DEPLOY), and the clone line points at the public repo
(doxa-open/doxa-maps, as DEPLOY already did). DEPLOY's "three ways" becomes the two routes;
the CDN route is retired.
- The web folder row said 536 files, 76 MiB. The folder is 855 files, 119.2 MiB. New
check:readme-size re-measures public/dx-maps and fails when the row is more than 10% off,
printing the line to paste. It passes at 855/119; the old numbers fail.
Rudder was measured slowing after about 1,400 words while a Tauri build
ran. The word count is incidental — it is a proxy for how long the build had
been going. The speech model runs on the CPU (backend/rudder-engine starts it
with --device cpu), and a Rust build saturates every core, so at equal priority
it starves the one process whose latency a person can hear.
Three changes, and the direction of the first is forced rather than chosen:
1. internal/desktop-build.mjs nices the whole build tree to 10, through the one
run() helper every step goes through, so cargo, rustc and the linker inherit
it. We nice the COMPETITOR because an unprivileged process cannot raise its
own priority on Linux — a negative nice needs CAP_SYS_NICE — so "promote the
ASR" would fail with EPERM. If nice(1) is missing the build still runs and
says so, because a silent fallback would look exactly like the fix working.
2. cargo is capped at cores-1. nice decides who wins a CONTESTED core; it does
not stop the build taking all of them, and a fully subscribed machine still
adds scheduling latency to the audio path. An explicit CARGO_BUILD_JOBS is
honoured rather than overridden.
3. yieldCPU reports its error instead of discarding it. It was
`_ = syscall.Setpriority(...)`, which cannot fail visibly — a call that
silently does nothing is indistinguishable from one that worked, and this one
guards something audible.
Also, one real O(n) in the widget, found while checking studio's answer that the
collapsed pill already shows the last words: sendActivity split the ENTIRE draft
on every activity tick to take five words. Measured at 3,000 words: 1.61ms per
tick against 0.051ms for a bounded tail, 31x, with byte-identical output.
Verify the priority work with `ps -o ni=` on the running cargo processes, never
by any of these calls returning.
Two halves, Node only. The helper, against fake responses: a real WebP
is accepted by type or by its bytes; a host's 200 fallback page is
refused as a raster and as a source; HTML labelled octet-stream is
refused; a 404 and a network error are misses for the raster; a network
error on the source is thrown as itself. The callers: neither
usePosterPrints nor PosterWall may fetch posters/ directly again, since
each once carried its own copy and a guard in one did not reach the other.
Poison-tested: restoring the old accept-anything fetch fails 3 checks,
and PosterWall's direct fetches failed the caller check before its switch.
gates:pre: PASS 67, FAIL 0.
With posterSources on, the tree shipped only the 28 zone poster sources,
so exporting any of the 176 city or 174 country posters said "poster not
found (404)": 350 of 378 posters could not be printed.
Each poster now ships once: its .html source where one exists, else its
.webp. Generating sources for the 350 was not an option: city posters
need OSM street tiles to regenerate (publish-html.mjs), and a source is
the heavy form (about 1 MB each against about 0.1 MB for a raster).
Measured with check:tree-weight: 102.7 MB before, 137.9 MB after, of
160 MB. The 350 rasters are the tracked data/posters/*.webp: all 350 are
existing git blobs, so the repository gains no new pack bytes.
New gate check:posters-exportable (in gates:post). Every poster id, from
data/posters and the wall's previews, must ship a real WebP or a real
poster source (its class="stage" element), so a placeholder or a host's
fallback page cannot pass. It skips when posters are off by config, which
check:poster-degrades covers. Poison-tested: a missing poster and a fake
image each fail.
The list kept its own page-1 map lift (packetMapSvg + packetLabelFor) beside the one in
usePosterPrints, and the two drifted: the list's relabelled country names, the wall's did not, which
is the zh finding rega made in the app. With the composable now taking iso3Names, the list passes it
and deletes its copy, so the list, the wall and any other map print page 1 through one path.
The selection packet borrowed one list's sheet when it held exactly one list. That is now a general
seam: item.mapId names the sheet when it is not the item's own id, and mapId: null says the item has
none, so a multi-zone selection fetches nothing and page 1 says the map is missing. Measured two-sided
through the real printPacket. gates:pre 66/66.
A review of the full-detail packet, page by page:
- Page 1: a poster carries its own title, so the packet's title above it printed two headers and
shrank the map to fit beneath them. A poster sheet now fills the page inside the white margin by
itself; in another language its kicker and title become the packet's own, fitted to the English
width. A map-only sheet keeps the packet's title in the band it leaves empty.
- The life calendar is laid out like the Module 1 "Life Calendar in Weeks": square boxes, week
numbers across the top, the year down the side, a rule between decades, framed and centred. The
first 33 years are shaded as the life of Jesus (1,716 weeks).
- The %-evangelical grid fills from the top in reading order instead of scattering.
- The flow pages had no page padding at all, so their headings sat on the sheet's edge; they keep
the same margin as every other page now.
- The bounded and centered sets are one page, two square panels side by side as on the Module 1
list of 100, each headed by name with a cross at its centre. The centered set is a target of nine
numbered bands, its dots spread through their bands by the shared ringBands geometry (no overlaps,
none on the numbers), and the legend explains every ring with its count.
- The lightweight packet lists every group as a worksheet: a short note, one column header per page,
numbered rows sized for paper with room to tick, no photos or descriptions.
Seven new strings in all eleven languages. check:packet-a1-structure and check:packet-variants
assert the new contract two-sided (a poster page 1 has one header, a map-only page 1 still has the
packet's; every ring is named, a populated one with its count and an empty one with a dash).
check-hardcoded-english judges each quasi of a template literal on its own, which
is correct -- an allow-listed fragment must not drag a whole literal in with it.
But splitting `url(${base}/images/worldmap.svg)` leaves "url(" and
"/images/worldmap.svg)", and `repeat(${n}, minmax(0, 1fr))` leaves "repeat(" and
", minmax(0, 1fr))". The whole string is obviously CSS and the existing CSS shape
matches it; no individual fragment carries a brace or a colon-value pair, so each
one fell through and was reported as user-visible English.
Colon-terminated prefixes failed the same way. The id rule accepts a trailing
hyphen, because `id="country-${code}"` leaves "country-", but a langTree id is
`fam:${x}` / `dial:${x}` and a localStorage key is `pgg-cards-per-row:${profile}`.
Those leave "dial:" and "pgg-cards-per-row:", which no shape described.
Two shapes, both narrow: a CSS function fragment, and a prefix that is one
lowercase token ending at a colon. The fleet total goes 257 to 249 -- five of
mine, three in other owners' plugins -- and every real string still flags, which
I checked in the same breath rather than afterwards: "Search people groups…",
"Show filters", "selected" and "Research map loading…" are all still caught.
Worth saying plainly because the number was about to be routed as work: these
eight were never translation defects. Wrapping "dial:" in t() would have been a
bug, and someone would have done it on the strength of a count.
gates:pre's check:packet-offline writes the four fixture packets into the
same directory this gate reads, so after gates:pre that directory is never
empty. A tree that had generated no zone then failed here on all 28 zones,
and gates:post was red on every landing for nothing anyone could fix.
Fixtures exercise the renderer, not a zone. With only fixtures present the
coverage assertion now says it did not run, and the per-packet checks still
run on the fixtures. One zone packet present and all 28 are required, as
before: planted with laos alone, it fails on the other 27.
The board carries a third card schema: proof in an 'evidence' list. 44 cards use it and 12 of them carry
none of the four fields the gate counted, so each false-failed on closing (one already had to be patched
with a redundant bluf). A non-empty list with at least one non-blank entry now counts. The self-test
covers both sides: a real list passes; an empty list, or one of blank entries, still fails.
The standing rule is that comments, docs and UI strings say what the code does and why, never who said
what. Several hundred existing lines break it; owners remove them in one sweep per area before the public
giveaway. Until then check:no-new-driver-references (gates:pre) fails any NEW line that refers to the
Driver and baselines the rest, keyed by file and line text with counts rather than line numbers, so
parallel sweeps only ever shrink it. It matches the role word, not a GPU or device driver, and its
self-test proves both that and that a new line beside baselined ones is isolated.
Five physical CSS declarations in the chrome every production map renders, so
under Arabic each of them lands on the wrong side for every map at once:
SemanticTreeLegend's active tab chip drew its border on the left, the share
menu's step list padded from the left, its tab spacing and its auto-spacer both
pushed from the right, and the geocoder's search icon sat at left:7px inside an
input that mirrors. All five become logical properties, which in a left-to-right
reading resolve to exactly what they resolved to before: four insertions against
four deletions in the components, one property each, and the English view cannot
move.
The reason they survived is the more useful half. check-legend-logical-insets
existed and passed, but it read only library/components/legends -- the share menu
is one directory away, and a reader in Arabic does not know which folder a control
came from. So the scan now covers map-controls as well.
Widening it was not enough. It still passed, because its CSS extractor matched a
template literal with the pattern `([^`]*\{[^`]*\}[^`]*)`, which describes a
literal containing ONE brace pair. ShareButton keeps its styles in a 600-line
useShadowStyles literal with hundreds, so the scan captured seven fragments
totalling 7,003 characters of a 629-line file and never saw .share-tab at all. It
was reporting PASS on a directory it was barely reading, which is worse than
having no check: a green gate is read as evidence. Template literals are now
taken whole and kept only when they look like CSS -- 2 blocks, 8,462 characters,
and the rule is finally in scope.
The self-test poisons BOTH directories. Poisoning only a legend would have proved
the check works on the half it always covered, and passed unchanged the day the
wider scan silently narrowed again. Its map-controls arm puts back the exact
declaration this commit removes, and its failure message says what a failure
means.
landing-preconditions verified EVERY bundle whenever vite.doxa-maps-build-config.json changed, but two
of its keys build nothing: dropin (where sync-dropin copies to) and the "//" comment keys. Changing
dropin.dest cost a full verification, about 10 minutes, for no possible staleness. The config now counts
as a build change only when another key differs. Checked on real history both ways: the dropin.dest fix
is not a build change; the posterSources flip is.
Four places in the repo (the registry, the resolver, the packet renderer and a gate's allowance) cited a
research spec that lives outside the repository, so a reader of the code could not follow the reference.
library/lists/README.md now documents the list as built: how a list is chosen, the registry and its five
lists with their measured counts, where membership is generated, the two tiers (narrowing DOXA's rows,
failing closed; IMB rows per map), the data sources and their terms, and the questions still open. The
four citations point at it.
The drop-in was proven on a laptop with one server: the three DX Life branches rendered their maps,
7 of 7 checks passed, with internal/sync-dropin.mjs --from-config --dest <doxa-life>/public. This
makes that the default path:
- dropin.dest is ../doxa-life/public. The old ../doxa-life/public/embed/doxa-maps-build was stale.
- bun run dropin = sync-dropin --from-config; deploy:dev (build + drop-in) now uses it.
- The README gains "The easiest path": the studio for the work, dropin.dest at your site's public/,
then build and drop in.
One hazard came with the new destination, and it is closed here. --prune deletes every file in the
destination outside the drop-in's allowlist, which was right while that was a folder of its own. Into
a site's public/, it would delete the site. deploy:dev passed --prune, so it no longer does, and
sync-dropin refuses --prune into any folder named public/. Measured in a scratch site: --prune into
site/public exits 1 with the reason; a plain drop-in copies dx-maps/ and leaves the site's own
file in place.
One sha instead of two, per the shepherd: both branches fixed the SAME
regions-list rule, because a gate is useless on a tree where its poison case is
still live. Landing them separately would have applied that fix twice.
★ I HAD CLAIMED THE TWO COPIES WERE IDENTICAL. They were not — I checked before
folding rather than after, and the hashes differed. The CODE line is the same in
both (`cyrillicHeading('.rl-title')`, one occurrence each); only the explanatory
comment differed. I kept the css-dangling wording because it names the arm that
now catches the bug, which is what the next reader needs. The emitted CSS is
unchanged either way — comments are not CSS — and the byte-compare baseline still
passes, which is the assertion that would have caught me if I were wrong.
WHAT THIS SHA CARRIES, both halves verified after combining rather than assumed:
THE MIGRATION (was 4a4bc96fc). The harness stub's `.value` is now the empty
string, so an injected-config read and a locale interpolation both behave as
they do in the browser. Measured components 2 -> 4, no NOT MEASURED left.
region-dashboards is byte-identical (a pure de-duplication); regions-list
changed because its old bytes were the bug, so its baseline was re-taken.
strongMedium() deliberately NOT adopted — the helper emits bare `strong{...}`
where region-dashboards scopes `.rd-root strong{...}`, and swapping would widen
the rule.
THE GATE (was 023ca2d08). check:shadow-css-parses now fails on a selector that
never receives a block, gated on `endsLiteral` — the LAST fragment — rather than
on `complete`, which is false for nearly every sheet here and therefore
suppressed the very bug the arm was written for. Four self-tests live in the
file: the exact regions-list string is caught, a last rule WITH a block is not,
a trailing comment is not, whitespace is not.
Re-verified on the combined tree: ported-globals 4/4 byte-identical, shadow-css
80/80, the poison case still produces exactly 1 dangling hit when reintroduced,
gates:pre 56/56, and the broken form is absent from the shipped minified bundle.
Railway, DigitalOcean and `docker build` use a root Dockerfile with no setting, and it is
Coolify's and Dokploy's default path. The file moves; nothing is copied.
- VOLUME /data is gone. Railway rejects the line, and a named volume is needed anyway
(VOLUME alone gives each new container a fresh anonymous one). The platforms mount /data.
- The baked PocketBase gate's comment named the old path. It now says the repo-root
Dockerfile, and the gate's SHA-256 pin moves with it (2987055b... -> 2eceef03...). The
hook's code is unchanged: the exact RUN passes `sha256sum -c`, and the old block still
passes its old pin.
- .dockerignore excludes everything: the Dockerfile copies nothing from the build context,
so no platform sends the whole repository to the builder.
- compose.yaml: the same image with its /data volume and memory limit; settings from .env.
- deployment/cloud/README.md gains "Where it runs": Coolify, Dokploy, Railway, a server with
Docker, Kubernetes, and why not DigitalOcean App Platform (no persistent volumes).
DEPLOY.md and two comments follow the move.
upg100's harness recorded my two files as NOT MEASURED because its stubs threw on
them. Both were FIDELITY gaps — the stub modelled states that cannot occur — so
they are fixed by making the stub more like the browser, never by relaxing a
comparison:
· `.value` was undefined, so a component reading an INJECTED ref threw
"Cannot read properties of undefined" (region-dashboards.vue:199,
cfg.value.regionSlug — ProfileLoader always provides that object).
· the stub threw on ANY string coercion, because a Proxy over a function has no
primitive conversion. That alone produced regions-list's "Cannot convert
object to primitive value".
`.value` is now the EMPTY STRING, which is the one shape that serves both real
cases honestly: a locale ref's .value IS a string, so `${locale.value || 'en'}`
correctly yields 'en'; and reading a property off '' returns undefined instead of
throwing, so an injected-config read falls through its own `|| ''` exactly as in
the browser. Symbol.toPrimitive makes any remaining stub render empty.
Measured components: 2 -> 4, with no NOT MEASURED left.
★ AND THE MIGRATION FOUND A REAL BUG, which is the part worth reading.
regions-list's Cyrillic heading rule ended with a trailing comma and NO
declaration block, so the concatenation emitted:
[lang="ru"] .rl-title, [lang="ru-RU"] .rl-title,regions-list
A selector list with nothing to apply — the Russian heading font was silently
never applied — AND it swallowed useShadowStyles' style-id argument as CSS,
leaving that sheet unlabelled. Braces stayed balanced, so
check:shadow-css-parses could not see it, and it passed 80/80 both before and
after. cyrillicHeading('.rl-title') replaces it and emits a real rule.
SO THE BYTE-COMPARE SPLITS, deliberately, and I did not paper over it:
region-dashboards.vue IDENTICAL — cyrillicHeading('.rd-title','.rd-h2')
emits exactly the literal it replaced, verified
character-for-character before the swap. A pure
de-duplication.
regions-list.vue CHANGED, and the change IS the fix. The old bytes were
the bug, so its baseline is re-taken rather than
preserved. Preserving it would have pinned a broken
rule forever, which is the one thing a byte-identical
gate must not be used for.
NOT MIGRATED, on purpose: strongMedium(). The helper emits bare `strong{...}`
while region-dashboards emits `.rd-root strong{...}` — a narrower selector.
Swapping it would widen the rule to every <strong> in the shadow root, a
behaviour change disguised as a de-duplication. scrollbarGutter() is likewise
left inline: both files carry it inside a larger .rd-scroll/.rl-scroll rule, not
as a standalone declaration the helper could replace.
Both files now gated (owned:true) instead of recorded-only, so a future edit that
changes their emitted CSS fails rather than being noted.
key-cities is a gated profile: generate-tree leaves it out of the manifest and the pages. Its
bundle file was still in public/dx-maps/key-cities/key-cities.js, committed, and the landing
freshness pass rebuilt it into public/ every time it looked "stale", so any host that serves or
drops in the tree folder whole shipped it.
- vite.config.js builds a gated bundle into .gated/<tree>/ (ignored by git) instead of public/.
- check:composite-freshness skips gated bundles: they have no committed bytes to be fresh against.
- The committed public/dx-maps/key-cities/ is removed. History is not rewritten.
- generate-tree removes a gated folder it finds in the tree (an older build, or a landing that kept the
tree's side of a conflict), so the tree it writes is clean whatever came before. Measured: a planted
public/dx-maps/key-cities/key-cities.js fails the gate, and one regeneration removes it.
- The gated list is read in one place, internal/_tree-name.mjs, beside the tree name.
- check:gated-not-in-tree (gates:pre) fails if a gated name has a folder in the tree, a manifest
entry, or a page referencing its script, and proves each arm on a planted leak first. It FAILED
on the tree before this change (key-cities/ is in the tree) and PASSES after.
The data tables are left alone on purpose: key-cities.json also feeds the production Cities with
Internet map, and whether that dataset should be public is a separate decision.
doxa-maps-spa marks the pane of every tab that has explainer text with data-doxa-help, listens for
doxa:help, and opens the same what/how dialog for the active tab. Closing returns focus to the
exact "?" that opened it (composedPath reaches into the map's shadow root). The top-centre button
and its CSS are gone; a comment for the device-rule toast that sat above the old button's rule
is back on its own rule.
check:help-reachable (gates:pre) keeps it reachable: every map with text has a hosted "?" or a
named help control of its own; the app listens, marks its panes and has no top-centre button; and
the composable, executed on a small fake DOM, dispatches doxa:help bubbling and composed and finds
the host across a shadow root. Six planted defects, each caught.
It asserted the variant is forwarded with /printPacket\(zone, \{ ... variant/, so renaming the argument
to titled (the zone carrying its localized title) failed it while the variant was still forwarded. Any
first argument now matches; a planted call without the variant still fails it.
Page 1 of an Arabic packet lifts the English 3x4 poster, and the
poster inherited the page's dir="rtl": English punctuation jumped to
line starts, the header's key and value columns collided, names ran
into their numbers, and four text blocks were pushed off the sheet's
left edge (rendered and measured in print emulation). The Chinese
packet, which is ltr, was intact.
The lifted sheet's wrapper now pins dir="ltr". The packet's own title
stays rtl; the relabelled Arabic country names are single runs, so they
keep their order.
check-packet-a1-structure asserts the pin on both kinds of sheet; it
fails without it.
e18c7f05a declared the map relabeller above LANG, which it reads: a
temporal-dead-zone ReferenceError at startup, so every offline packet
run stopped before building anything. node --check passed, because it
checks syntax and not evaluation order, and no gate ran the generator.
The block now sits after LANG and VARIANT, with a note on why.
packet:ci already built the committed fixture packets through
packet.mjs, with no network, in under five seconds - and it catches
this exactly: exit 1 on the broken generator, exit 0 on the fix. It was
in no gate list, so it never ran. It is now check:packet-offline in
gates:pre.
Its failure message printed the last line of stderr, which for a crash
is Node's version banner ("Node.js v20.20.2"). It now names the error
line: "ReferenceError: Cannot access 'LANG' before initialization".
The map sheets are generated once, in English, so in an Arabic or
Chinese packet page 1's map labelled every country in Latin capitals,
and its privacy line was English.
The labels are now replaced as the sheet is lifted into the packet.
They are recognised by the generator's own rule rather than a list:
generate.mjs draws each label as esc(ISO3_NAME[iso3].toUpperCase())
from the zone data's ISO3_NAME table, so inverting that table under
that transform names the country behind every label it can draw. The
inverse is unambiguous (162 labels, no two countries sharing one), so a
label is known exactly or left as drawn. The map's privacy line is word
for word the packet's privacy_note, so it takes this language's.
A width guard: the generator placed every label and then dropped any
that would collide, measuring the English name, so a translation is
fitted into that same box - estimated wider, it shrinks in proportion
and cannot reach a neighbour the English name did not. On the Sudan
poster, six Arabic names shrink (the longest, Saudi Arabia, from 15px
to 10px); none of the narrower Chinese ones do. Capitals' tracking is
dropped for scripts without capitals, where it breaks Arabic joining.
English is untouched: its relabeller is null, and the lifted sheet is
byte-identical with and without it. No sheet or poster is regenerated -
the change is in the lift, so the tracked 3x4 posters page 1 now uses
are relabelled as they are.
Both print paths supply the relabeller: the UPG 100 profile from its
own ISO3_NAME and the library's country-name resolver, the offline
generator from the same table and the injected ISO-3 to ISO-2 table.
check-map-labels (gates:pre), on the tracked Sudan poster: English
byte-identical; in ar and zh no known English name left, halo matching
fill, privacy localized; the guard fires for Arabic, not Chinese, and
grows nothing; Arabic tracking off, Spanish kept; an unknown label left
alone. Mutation-tested five ways, each caught.
The Deaf zone printed "Deaf (affinity bloc A017)" as its title - on
page 1, at the head of every profile page and in the document's own
title. That is the legend's internal label: the unified Deaf row is an
affinity bloc spanning every region, so there is no country to name it
and zone-title.js declined it, leaving the label to print.
An affinity-bloc zone has a name of its own even though it has no
country. zone-title.js now reads it from the packet strings, keyed by
bloc (zone_affinity_A017: "Deaf people groups", in all eleven
languages), so another affinity bloc a list adds is named with one
string and no code. With no strings it still declines and the label
stays - it never invents a name. Both print paths now pass the
language's strings, English per key, the same fallback the renderer
applies.
check-zone-title-tier-a covers the unified Deaf zone, which it never
saw (it enumerates clusters.json entries, where Deaf is two older
lists): the translated name in en/es/ar, never internal text, and a
decline without strings. It now tests with the real packet strings,
which carry the multi-country joiner, so a broken one-country guard
would generate a title and be caught.
The page says each box is a share of the people in this list. It
printed one DOXA-wide figure - 0.6133% of 209,079,159 people - in
every packet, true of no single list.
That figure had no builder in the repo. build-evangelical-share.mjs
makes its method reproducible, per group: the IMB CSV joined to the
DOXA snapshot by ROP3 + country, each group's EvangelicalLevel band at
its midpoint (0 for No Known Evangelicals, the midpoint read from the
band's own wording), weighted by IMB population. It writes
evangelical-by-slug.json and refuses to write unless the table, summed
over every DOXA group, reproduces christian-share.json exactly: 2,091
matched, 209,079,159 people, 1,282,178 evangelical, 0.6133%. Change a
single band's midpoint rule and it refuses.
Not the API's imb_gsec: GSEC puts 225 groups at "no evangelicals" where
EvangelicalLevel puts 1,444, because its "less than 2%" band also
carries church-planting status. A share from it sits on a different
scale from the method the page's own footnote states.
evangelicalShare.js sums a packet's own groups - pure, zero imports, so
the browser print and the offline generator run the same sum. Across
the 28 zones the page now carries 28 different figures, from 0.182%
(Deaf) to 0.922%. A host that supplies its own figure still wins.
A list with no published levels prints a dash and no number. The old
guard, Number.isFinite(Number(x)), passes null as 0 and would have
printed "0% evangelical" for no data. Where some of a list's groups
lack a level, the footnote says what the figure is computed over; it
says nothing when coverage is whole. That clause is in all eleven
languages.
check-evangelical-share (gates:pre): the table reproduces the committed
figure; different lists differ; no data is a dash; coverage appears
only when partial. Mutation-tested four ways, each caught - the fourth
only once it corrupted an existing group rather than adding one no sum
reads.
Every packet was titled with its zone's hand-written English label, so
a Spanish Sudan packet said "Sudan" and an Arabic one printed Latin
script on page 1, on every profile page's header and in the document's
own title.
A one-country zone's title is that country's name, and Intl supplies it
in every language: the packets now carry it - Sudán, السودان, सूडान,
苏丹 - through zone-title.js's tier A, on both print paths. Every other
zone keeps its hand-written label, since a region's name is not a list
of its countries (tier B and C are a pending product decision). The
selection packet names its cover the same way.
In the browser the title uses the library's country-name resolver; the
offline generator cannot load that module under Node, so it reads the
ISO-3 to ISO-2 table and injects it, the path zone-title.js was built
for. A data set without zone-title.js keeps its labels.
check:zone-title-tier-a was committed unregistered and so never ran; it
is now in gates:pre. Its comments, and zone-title.js's, no longer cite
people or dates.
Measured all 29 zones in en/es/ar with a real country-name supplier
before wiring anything, and the result split cleanly.
A one-country zone generates a correct title in every locale at zero
translation cost: Sudan / Sudán / السودان.
A multi-country zone generates a title WORSE than the hand-written
label it would replace, in English too, so this was never a
localization trade-off:
"Middle East Core" would become "Iran + Yemen + 11 more"
"Scattered / Diaspora" would become "Andorra + Cocos (Keeling)
Islands + 2 more"
and no maxParts rescues it -- 2, 3, 4, 13 just walks toward a full
13-country enumeration. A multi-country zone has a name that is not a
list of its members, so no assembly of country names can produce it.
The generator is correct; the approach does not fit that tier.
The guard is in zone-title.js, not at the caller, because six product
call sites supply this across three surfaces (the legend three times,
posters, packets). A caller-side test would be repeated six times and
would drift the first time one was missed, silently regressing one
surface while the others stayed right.
Tier B now keeps its label through the same null-means-keep-yours path
tier C already used, so this widens an existing contract rather than
adding a new one.
check-zone-title-tier-a.mjs is two-sided on purpose: a gate that only
checked "the 3 generate" would pass a build that also regressed the
other 26. It asserts the 3 produce non-code titles that differ across
locales, AND that every other entry returns null so its label survives
byte-identical -- by value, not by count, since "26 declined" would
pass even if the wrong 26 declined.
Shown capable of failing four ways: removing the guard (the forbidden
regression), declining tier A too, a supplier returning raw codes, and
a supplier ignoring locale. The third exits 75 NOT MEASURED rather than
failing, because a broken supplier makes every zone decline and would
otherwise look like a clean pass.
The app's Chinese is Simplified, and so were 39 keys of the packet's zh
block. The other 18 were Traditional - page 1's subtitle, the list
header, the card labels, and the whole prayer wheel and all twelve
prayer points - so a Chinese packet switched script partway through.
The block had been extended by matching the neighbouring line, which
was itself the drift; the missing-map sentence added in 74e5c4b0d
followed that same neighbour and is corrected here too.
Converted with a real converter (zhconv, MediaWiki's tables), not a
hand-picked character list: it converts characters (為 to 为, 禱 to 祷)
and mainland vocabulary (列印 to 打印), and a hand list had found only
17 of the 18 keys. A second pass changes nothing, so the block is
Simplified by the converter's own definition. 祂 is kept. The block
carries no licensed scripture - Mark 5 falls back to English there - so
only prayer prose and labels changed.
check-i18n-parity's zh arm walked the app catalogues and never the
packet's, which is how 18 keys drifted unseen. It now walks the packet
block as well; pointed at the block before this commit it fails with 41
values, and passes after. Its character set remains a floor, as it says.
FULL DETAIL is four front pages, the profiles, four back pages:
map, triangle, life calendar, % evangelical, profiles,
bounded set, centered set, prayer wheel, 12 prayer points.
LIGHTWEIGHT is two front, the profiles, two back:
map, triangle, profiles, prayer wheel, 12 prayer points.
Both are generated from data, in every packet language, from the
browser and from the offline generator.
PAGE 1 - the map, on a vertical page. One lift for every caller
(packet-render.js liftMap): the browser composable, the UPG 100
profile and the offline generator each carried an identical copy of
the old wrapper (250px down, scale 0.62), and three copies of a layout
decision is how page 1 drifts between print paths. The lift prefers
the map-only vertical sheet (mapSheetFiles: <zone>.packet.html, then
the poster) and places it so its names land on the same 60px margin
every other page keeps; a poster falls back fitted under the title at
0.81 rather than 0.62. The title block is drawn after the map, since a
map-only sheet begins with a full-page background. Anything that is not
a map sheet - an icon-sized <svg> from a host that answers a missing
file with a 200 page - is refused, and page 1 says the map is missing.
Page 1 now states which set it belongs to ("One of 28 UPG 100 prayer
lists"), only when the caller supplies the count.
PAGE 2 - one inverted triangle in the Vision Card's look. The card's
steps are the top (every people group, the unreached, the frontier,
the network's list), narrowing to the UPG 100 packets and, at the tip,
this packet. The card's greens, darkening as it narrows, walked
through its own colour stops; every numeral picks white or ink by the
WCAG ratio (all at least 3.4:1, large bold type). Tier widths are
geometric, not proportional - 2,107, 28 and ~100 against 12,379 would
be hairlines - so where two sets truly nest the caption states the
step removed, and the network's list, which does not sit inside
frontier, gets the card's "N of them are frontier" note instead of a
false subtraction. The packets tier and its count appear only when the
caller supplies one.
BACK PAGES - the bounded and centered sets get a page each at twice the
width, under the one page key, so lightweight drops the pair together.
THE WHEEL - the Metacamp Module 1 prayer wheel: its twelve colours,
sampled from the fill operators in the source PDF's content stream; its
orientation, the opening PRAISE at nine o'clock and READ THE WORD at
twelve; white labels with a shadow and a faint outline, so a print
path that drops SVG filters still leaves them legible on the palest
segments; "5 minutes" badges; the attribution with the website. The
labels stay the packet's own rhythm names, translated in every
language and matching the 12 points page.
The offline generator makes both packets (PACKET_VARIANT=lightweight;
the full packet keeps its names, lightweight is <zone>.lightweight and
stamped with its variant).
STRINGS: four new keys, in all eleven blocks - each block's own words
for "packet" and "prayer list", the frontier note aligned to the word
the frontier tier uses on the same page.
GATES: check-packet-a1-structure (registered in gates:pre) reads every
sheet's CONTENT for the order of both variants, page 1's lift in all
four cases, the triangle's tiers and steps, the wheel's colours and
orientation, and all 28 zones. Mutation-tested eleven ways, each caught
by its own message. check-packet-variants now asserts both sets pages
and a four-sheet difference. The browser gate check-packet-render
counted funnel rows and bars by inline-style substrings, which an SVG
triangle has none of - both would have read zero and passed - so it
now counts .funnel-n against .funnel-tier, its sheet arithmetic counts
the sets pages, and its Latin allowlist carries the renamed product
name. check-packet-browser used a copy of the old wrapper; it now uses
the product's lift.
Page 1 of a prayer packet is a lightweight vertical one-pager showing
the zone's geography and every pin with its numbered name - not the
poster. Until now the packet lifted the 3x4 poster (1500x2000, its own
title, facts table, vision columns, networks and closing verse), shrank
it to 62% and pasted it under the packet's title: a small poster on a
page, with the English poster prose baked into the artwork.
FORMAT=packet is a new sheet format: 1500x1941, the packet's own page,
with the top band left EMPTY for the title the packet draws in the
reader's language, a slim foot, and everything between given to the map
and its label columns - about 1,600px of height against the letter
sheet's ~1,200, so the names get the larger type the sparse posters use.
It omits all the poster furniture, and its root <svg> says
data-sheet="map-only" so the packet can tell it from a poster.
Paper formats no longer overwrite the poster. Every format used to write
out/posters/<zone>.html, so FORMAT=letter silently replaced the 3x4 the
wall and the site serve. packet and letter sheets are now
<zone>.packet.html and <zone>.letter.html, with their own manifest;
3x4 and 3x3 keep the bare name, which every consumer fetches.
publish-html.mjs read everything before ".html" as the zone id and so
skipped every paper sheet as an unknown zone; it now strips the suffix.
The two fail-loud guards (header overflow, verse off the sheet) stay
armed for posters and stand down only for a sheet that does not draw
the furniture they measure.
Proof that posters are untouched: the default 3x4 output for sudan,
asia-india-kerala, deaf and west-africa is byte-identical before and
after, generated with POSTER_DATE pinned. The first attempt was NOT
identical - an HTML comment added inside the SVG template literal is
emitted into every poster - so the notes live in JS comments instead.
check-zone-census counted paper sheets as posters (it would read 56
for 28 zones); it now counts posters and reports paper sheets apart.
The legend offers 28 zones; every place that turned a picked id back
into a zone spread clusters.json's three geographic buckets, which is
27. The missing one is Deaf, whose two deaf lists the legend collapses
into a single affinity-bloc row. So in the browser, silently:
- the Prayer packet for Deaf returned on "if (!zone) return";
- the Deaf poster titled itself by its slug;
- a selection containing Deaf dropped it in .filter(Boolean);
- the packet count read 27 while 28 packets exist, and that count
prints on page 1 and at the tip of page 2's triangle.
The packet fetch also lacked rop1, the field the Deaf selector reads,
so even a found Deaf zone would have resolved to nobody.
The fix is one enumeration. legend-tree.js now exports DEAF_SELECTOR
(the unified row's selector, which exists nowhere in clusters.json) and
zonesWithSelectors(clusters): the legend's leaves with their selectors
attached. The plugin's printableZones() keeps the 27 catalogue objects
exactly as they were and adds only what the catalogue lacks, so the
packet, poster, selection and count all see the same 28. rop1 joins
both packet fetches.
The three hand-typed copies of the Deaf literal (the map, the poster
generator, the offline packet generator) now import the one
definition; the offline generator's own id-to-selector table is
replaced by zonesWithSelectors. The copies left are frozen snapshots and
proof scripts, which keep independent literals on purpose.
The offline sweep gate's coverage had the same 27: it never required a
Deaf packet. It now counts through the legend.
Verified: the live API carries exactly 123 A017 rows, matching the
legend's Deaf count; the default 3x4 posters for sudan, kerala, deaf
and west-africa are byte-identical before and after.
One landing because it is one bundle: the ring band layout wiring (already on
this branch) plus the three set-fit commits, replayed as SOURCE only. The
generated tree is rebuilt below rather than merged — a previous cherry-pick of
mine swept a conflicted generated file into a commit, and hand-resolving build
output is how that happens.
From set-fit: extentFraction 0.17 -> 0.095 on both charts, so the drawing fits
the height the engine's zoom floor leaves it; crossFeature restored from history
and called by BOTH maps through one shared addCrossLayer; and the ResizeObserver
fit restored, which rega found missing because zoomAtOpen read 1.00 in all eight
chart/layout combinations — too clean to be a fit result.
Two things this merge had to get right, and neither was automatic:
- Taking GiCenteredSetMap.vue wholesale from set-fit silently REMOVED the ring
wiring — bandFill, curvedRingLabels and openAndFrame all went to zero while
every gate still passed. Both feature sets are now in the file together.
- Taking package.json wholesale would have deleted studio's check:ring-bands
registration, including its gates:pre entry, landing a gate nothing runs. It is
rebuilt from development's, with only check:gospel-set-fit added — in all three
places this time: script, gates:pre, and //scripts.
Rings that zoom apart, as a flag, so either answer to row 130 is one line at
the call site. DEFAULT IS 'fit' — this changes nothing today.
'fit' keeps every pin at least minSep apart at the chart's own scale and reports
what will not fit. That is the right behaviour for a PRINTED sheet, where there
is no zoom and a dot a reader cannot resolve is a dot that is not there.
'fill-all' places all of them. Separation is in map units, so it grows with zoom:
measured on the shipped table, the worst ring is ring 1 (1,190 groups in the
smallest circle) separating to 8px on screen at z=4.45, and ring 8 (1,932) at
z=3.50. The GROUP label tier already starts at 5.2, so every pin is clickable
before the names appear. At the start zoom the band reads as a dense arc of
colour rather than countable dots, which is the intended appearance — and the
reason a caller in this mode must caption the true total rather than a count
that changes with zoom.
The row count had to follow the POPULATION here, not the pin size. Keeping the
'fit' row count would have packed 1,932 pins into the two rows a band allows at
full size, and they would have sat almost on top of each other along the arc —
the single-line defect again, one level down. Ring 8 uses 14 rows in fill-all
against 2 in fit.
Real table: fit places 1,851 of 7,296, fill-all places 7,296 of 7,296, both with
zero pins outside their band. Six new assertions, and the row-growth arm is
poison-tested: making fill-all reuse the fit row count fails it at "2 rows vs 2".
Found by running the SHIPPED table through the layout instead of the uniform
sample the module was built against. The real near-neighbours data is heavily
skewed — ring 8 carries 1,932 groups, ring 5 carries 100 — and against it
bandFill placed 962 of 7,296. It dropped 87% of the map, and every drop was
faithfully reported, which is exactly why no assertion caught it: all 16 asked
"is the overflow reported?" and none asked "is the CAPACITY right?".
The capacity model counted one circle of pins at the band's innermost radius. A
band has depth, so it holds several concentric rows, each with its own
circumference. Counting them is a ~2x correction at ringGap 24 and more at
larger gaps; on the real table it takes placement from 962 to 1,851 on the gate's
geometry and every under-full ring now places everything it is given.
Placement had to follow: a hashed radius with one angular step was fine while a
ring held a dozen pins and wrong once it held hundreds, because two pins could
take nearly the same angle at nearly the same radius while the capacity sum
insisted there was room. Pins are now dealt round-robin into rows, each row
evenly spaced across the usable arc, staggered a quarter slot so adjacent rows
do not line up radially and read as spokes.
Two bugs found on the way, both by the gate failing honestly:
- the stagger first added half a step to every odd row, which pushed that row's
last pin past the end of the arc and into the label's keep-out — 7 overlapping.
Staggering WITHIN the slot keeps every pin inside its own slot.
- the shrink arm tested at 200 groups, which only crowded a ring while capacity
was undercounted. It now tests at 2,000, the honest pressure for a 7,296-row
table, and still reaches the pin floor and reports.
The new capacity arm asserts a RATIO against the single-row count, not a magic
number: its first version hardcoded ">400" from the component's ringGap of 22
while the gate builds radii at 24 — the code was right and my threshold was
guessed. Poison-tested: forcing rows back to 1 gives 1.05x and fails both arms.
