ci(soak): run go test -race ./... by default — the only shape that has reproduced the crash

[deadpoets]

The soak was measuring the wrong thing, and its clean results were not the
evidence they looked like.

What happened

The workflow looped a prebuilt single-package test binary. Across 1250
iterations it found nothing — while two ordinary CI runs reproduced the
crash:

#	Run	Error
1	31982627382	fatal error: stack not a power of 2 (goroutine free list)
2	32201035150	fatal error: acquireSudog: found s.elem != nil in cache (sudog cache)

ci.yml runs go test -race ./..., which compiles and runs the core and
redact test binaries concurrently. For a bug whose signature is a stray
write landing in runtime free-list caches, what the allocator hands back next
is the entire question — and concurrent processes change that answer. The
prebuilt-binary loop is several times faster per iteration, but it is not the
same experiment.

mode now selects between them and defaults to gotest, because a default
that is fast and has never reproduced is worse than a slow one that has.

The signature is now specific

Both sightings are per-P runtime free-list caches found in an impossible
state:

Sighting	Cache	Impossible state
1	gfree	stack size not a power of two
2	sudog	released sudog with elem != nil

In both, the crashing frame is innocuous — TestCapabilities_String, then
t.Parallel(). The goroutine merely trips over damage done earlier. That is a
stray write into memory the Go runtime owns.

One useful discriminator fell out of sighting 2. Making a field non-nil
cannot be done by a zero-wipe — that would set it to nil. So secureWipe is a
weaker suspect than the canary fill (which writes a random pattern) and ordinary
buffer writes. Recorded in the workflow header.

What this does not claim

It does not claim #56 fixes this. That PR's tree is not what crashed, and its
defect wipes another live, mapped secmem buffer — destroying that buffer's
secret without obviously reaching Go-owned memory. A soak against the #56 branch
is running to see whether the rate moves.

Validated: bash syntax checked over the extracted 54-line script, YAML parsed.

[deadpoets]

Landed on main as eb1bbfd via the integration train in #61.

The soak now defaults to go test -race ./..., matching what ci.yml actually runs — the only shape that has ever reproduced the corruption. The prebuilt-binary loop remains available via mode=binary.

Rebased locally so the commit carries your SSH signature; different SHA, hence the manual close.