Add cross_lab — the crossings, not more axes; bump pin to v0.32.0

[InauguralPhysicist]

Every existing lab stresses one axis: concurrent_lab is spawn, observer_lab is observe, io_lab is buffers, parser_lab is eval + listcomp. Eleven isolated axes is not a gauntlet — nothing here crossed anything.

The content also froze on 2026-07-06 while the pin kept advancing, so there was zero coverage of the task layer (v0.29.0), the tape, the buf_* kernels (v0.31.0), or lib/ui (v0.32.0).

The premise

Upstream's 15-finding review the same day produced three sharp bugs, and all three were interactions invisible to any single-feature test:

• comprehension var × a closure mentioning it (#633, BLOCKER)
• modal × focus navigation (#631)
• line numbers × generated scale (#630)

Single features pass. The interaction breaks.

Eight crossings

Using the surfaces that landed after the content froze:

crossing	what it pins
observer × task	an observed trajectory inside a task doesn't inherit or corrupt main's
task_self × self-detach	a task detaching itself reaps cleanly; join doesn't hang
lib/sync × comprehension	a lock held across a comprehension
buffer × JIT-hot × observer	three layers of assumption at once
eval × task	dynamic compilation on a task's copying stack
recursion × task	a task's stack is not main's C stack
module state × task	share-nothing — the task's mutation must not reach main
error shapes × task	thrown VALUE vs {kind,...} dict, both re-raised at join

It earned its keep on the first run

Two BLOCKERs. GAPS.md has open entries again for the first time since June.

EigenScript#642 (NEW) — a comprehension whose loop var shadows a local computes garbage:

define f() as:
    i is 0 - 1
    ys is [i * 2 for i in [1, 2, 3]]   # -> [-2, -2, -2], want [2, 4, 6]

The body reads the outer slot every iteration. rc=0, no diagnostic. Controls (one variable changed each): no outer i → [2,4,6]; module scope → [2,4,6]; outer named q → [2,4,6]. Identical under EIGS_JIT_OFF=1.

The crossing that found it was lib/sync × comprehension — the lock turned out incidental, and isolating it produced the controls above. It needs no lock and no closure, which is what makes it distinct from #633 and worse.

EigenScript#633 — re-derived independently (the read after such a comprehension is stale while eval sees the leaked value).

Per this repo's policy — no downstream workaround is a fix — the two affected assertions are parked with the exact lines to restore, not pinned to the buggy values. Everything else asserts the pinned runtime's real behavior, so a crossing that starts lying fails the gauntlet instead of going quiet.

Pin v0.30.0 → v0.32.0

Two behind; the crossings need the task layer and buf_*. Verified against the v0.32.0 tag built in a worktree, not local main — main carries unreleased APIs (file_dialog, element-wise log2) that would pass here and fail CI.

Full smoke green (12 workloads), lint clean.

Doc drift fixed while here

• CLAUDE.md said the pin was v0.30.0, listed 11 workloads, and named the retired T3200 as the hardware target.
• BASELINE.md marked stale: its 2007 Gateway went to EigenOS. The constrained regime is unchanged — an X540NA/N3350/4GB is not modern hardware either — so the baseline's purpose stands and only the capture is out of date. Needs a re-run on the X540NA.
• CLAUDE.md now records why this repo exists structurally: EigenScript CI is ~33 min of wall per PR across 17 checks (ASan alone 8+ min), so coverage added there is paid on every PR forever; the same coverage here costs ~30 s. That's as much the point as the workloads are.