Follow-up to the doc landing in 49, which was written before the guard existed.
Adds the guard: every browser gate reaches Playwright through
internal/lib/playwright-guarded.mjs and refuses unless browser-slot.sh exported
DOXA_BROWSER_SLOT; an unslotted run exits 75 with the command to use, quoted
verbatim. And says the slot is a COUNTING SEMAPHORE, not a mutex, with the reason
from its own header -- a single lock queued 10-19 deep for hours because the real
constraint is memory, ~300-500MB per chrome, not exclusivity.
Replaces the grep advice with `bun run browser-gates`. The grep still works and
is fine for a quick look, but it is a heuristic over file text: my own count from
grepping internal/check-*.mjs said 87 browser gates; the repo's own detector says
97, because browser gates also live in plugin-designs/*/scripts/. Corrected. The
authoritative list cannot drift from what is enforced, because
check:browser-gates-guarded uses the same detector.
One line the guard's header earns and this doc had not stated: being able to
launch is not the same as it being yours to launch. The guard stops the accident;
it does not hand a newcomer the seat. Testing whether the guard refuses is itself
a browser launch.
Does NOT claim path-into-node_modules detection: studio has that in landing 50,
and it is not in this tree, so the doc does not describe it yet.
238 checks; 148 run anywhere, 87 need a real browser, 3 shell out. A newcomer
cannot tell from the name -- check:packet-render sounds like a file check and
drives a browser; check:composites-mount sounds like it needs one and does not.
The test is one grep for playwright/chromium, and the doc says so.
The part worth writing down is exit 75. run-parallel.mjs:262 defines three
outcomes, not two: 75 (EX_TEMPFAIL) is NOT MEASURED -- the gate could not look --
and a gate that exits 0 while announcing SKIP is also NOT MEASURED. Only another
non-zero is a FAIL. Two traps follow: exit 0 alone does not mean a gate looked at
anything, and your own `timeout` produces 124, which says you stopped it, not
that the code failed.
Also states the rule the split exists for -- one machine, one browser seat, so
one person runs them -- and why it is not just logistics: a source read and a
render are different measurements. Four catalogue gates can pass while a page
still paints English; and of 16 cards parked as "needs a browser", 7 were about
markup, a path, a manifest field or a gate, and settled headlessly.
Every number measured on main d8baaefd2 rather than estimated, including
correcting gates:pre from 13 to 15 after re-running it.
