On 2026-09-29, 41 serve-strict processes had outlived the gates that started them: 279 MB resident,
the oldest 5.3 days, spread across eight worktrees. A gate that crashes, is killed, or is refused by the
browser-slot guard's process.exit cannot run its own srv.kill(), and _serve-reaped's handlers cover
signals and normal exit but not a SIGKILLed parent.
serve-strict takes --exit-with-parent: it records the pid that started it and exits within 2 s once
that process is gone. Opt-in, so a server someone deliberately left running is never taken down.
Every internal/ gate and tool that spawns it now passes the flag (14 files, plus _serve-reaped, the
shared helper). Proven: a parent that spawns the server with the flag and exits abruptly (as the
guard does) leaves it alive for a moment and gone within 3 s.
The orphans already running belong to other worktrees' runs and are left to their owners; plugin-local
scripts that spawn serve-strict should pass the flag too.
Run on a real token against seven production leaves, the gate reported "no
sized canvas appeared within 30s" for all six of its targets, and its own
negative arm failed too. Three separate causes, all in the harness.
IT LOADED 404 PAGES. The gate asked for `${base}/${id}/` -- a bare directory
URL -- and internal/serve-strict.mjs answers 404 for those deliberately,
because the live layer does and a gate passing on a URL that 404s in production
is the divergence that server exists to prevent. Its header even records that
only ad-hoc probes and one gate ever used the bare-directory form; this was that
gate. Measured here against a real serve-strict: the directory URL gives 404
with and without the query, index.html gives 200 both ways. Both call sites now
ask for index.html.
IT JUDGED BEFORE THE TILES ARRIVED. waitForFunction resolved as soon as it
could return {engine, features}, and that object is truthy with features:0 --
so it stopped at the first sized canvas and read an empty frame. It now holds
out for a rendered feature and lets the timeout report a genuinely empty map.
IT COULD ONLY SEE TWO OF THE SEVEN MAPS. Features were read from
window.__rcmMap || window.__rdashMap, the two instances this repo happens to
publish, so the other five leaves read -1 and would have FAILED as "0 rendered
features" even having drawn correctly. An init script now wraps
maplibregl.Map/mapboxgl.Map as soon as the CDN script defines it and collects
every instance the page builds, whatever it names them.
The WebGL flags are NOT among the causes, though they were the first
diagnosis. A three-way A/B -- these flags, ANGLE's software backend, none at all
-- gets WebGL and WebGL2 in every case. They stay as they were.
The negative arm is not a bug either; it is why this was caught. It failed
because MapLibre genuinely did not draw, which correctly refused to certify six
"no sized canvas" results as broken production maps when the fault was this
gate's own URL. Its message now says to suspect the harness first when both
engines report nothing.
Still exits 75 here: no token on this box.
This gate reported a 403 as "tileset entitlement, not the embed" and printed
only the first blocked URL. Both halves were wrong.
The claim is disproved. Measured on this gate's own token: /v4/ streets-v8 and
country-boundaries both return 200 -- with no Referer, and from
research.nuxtbase.app, maps-dev.nuxtbase.app, doxa.life, 127.0.0.1 and
example.com. The token is unrestricted and entitled. The old conclusion rested
on three no-Referer probes, which cannot tell a URL-restricted token from a
missing entitlement; it named a cause those probes could not support. Whatever
403s is a different request -- a style, a sprite, glyphs, a composite -- or a
different token than the one measured.
Printing one URL is what hid that. A leaf blocked on forty tiles showed
whichever tile lost the race, so the tileset was never legible and a second
blocked endpoint was invisible behind the first.
NOT MEASURED now names every distinct blocked path, with tile coordinates
folded to {z}/{x}/{y} so one blocked tileset is one fact rather than forty, and
styles, sprites and glyphs stay separate -- which is exactly the class the
evidence now points at. It also names where the token came from (env var or the
leaf's own baked one), never its value: the query string is stripped before any
path is printed, asserted in the exercise below. The real-failure arm gets the
same treatment.
The gate asserts no cause any more. It reports what was blocked and lets the
reader judge.
Exercised on six URLs spanning two tilesets, a sprite and a glyph range: 6 ->
4 distinct facts, and a planted token value does not appear in the output.
RT8 goal (b). Two gates already touch the engine and neither asserts this:
check-leaf-engine-per-provider proves MAPLIBRE purity (zero *.mapbox.com
requests under ?provider=maplibre) and that a map still draws, but never forces
Mapbox; check-provider-switch does load ?provider=mapbox — with a FAKE token, so
the engine LOADING was measured and a tile RENDERING never was, on one leaf.
So "Mapbox works" was unmeasured across the seven production tabs. This measures
it, and refuses to pretend when it cannot.
WHAT IT READS, and why each is the stronger property:
· THE ENGINE off the canvas that drew — .mapboxgl-canvas vs .maplibregl-canvas
— never from which globals exist. A leaf asked for Mapbox that quietly draws
with MapLibre looks identical to success in a screenshot and passes every
paint check; a peer script that merely loaded Mapbox proves nothing.
· THAT IT DREW via queryRenderedFeatures() > 0. A canvas at the right size is
not a map: this repo already had a country map render fully blank while "a
real GL canvas exists at 1200x780" passed.
THREE EXIT STATES, BECAUSE TWO WOULD FORCE A LIE. On this sandbox the Mapbox
/v4/ composite raster endpoint and country-boundaries both 403 — proven outside
the browser with no Referer, with 127.0.0.1 and with localhost, so it is tileset
ENTITLEMENT, not the embed (PRODUCTION-MATRIX:235). Failing on that would
manufacture work for every owner; passing on it would be false. It exits 75 NOT
MEASURED and names the endpoint. No token, or a non-pk. token, is also 75 — a
fake token is exactly the false green this gate exists to refuse.
THE NEGATIVE ARM is in the same file: ?provider=maplibre must draw with MapLibre.
If it does not, the engine check cannot tell the engines apart and every pass
above it is unfalsifiable.
★ TARGETS COME FROM THE MANIFEST, AND SELECTING ON isMap ALONE LOST A TAB.
Measured: it found 6 of the 7 curated tabs. research-map-plus-grid (Simple
Research) is a nested-elements-mfe with isMap:false whose CHILD holds the map, so
it would have been skipped in silence — the exact class of miss this gate is
meant to catch. A composed tab now counts if it is isMap ITSELF or composes a
profile that is: 7 of 7.
Registered in package.json (script + description) and in verify-browser-all's
GATES array — deliberately NOT gates:pre, because it needs the browser seat.
Verified by re-reading the parsed package.json, and check:no-orphan-gates names
only the pre-existing check-gospel-whiteboard.mjs, not this.
