liferaft differential proof: migrate onto the upstream #408 task layer

[InauguralPhysicist]

Split out of #408 (closed — the cooperative task layer shipped in v0.28.0). The one acceptance item from #408 not yet done: the differential proof that migrates liferaft off its hand-rolled src/scheduler.eigs onto the upstream task_spawn/task_yield/task_send/task_recv + task_sched_seed layer and confirms its byte-for-byte replay-under-chaos results still hold.

Why it's worth doing (not blocking)

liferaft's own src/scheduler.eigs works today, so this is validation, not a fix. But it's the strongest evidence the upstream layer is a faithful superset of a real hand-rolled deterministic scheduler — and it deletes a duplicated primitive from a consumer (the forcing-function payoff). If the migration surfaces a gap (a scheduling decision liferaft needs that task_sched_seed can't express), that gap is the real deliverable.

Acceptance

• liferaft drops src/scheduler.eigs in favor of the upstream task builtins.
• Its existing seeded chaos-replay suite stays byte-for-byte green (same-seed two-run diff).
• Any expressiveness gap found is filed upstream rather than worked around.

[InauguralPhysicist]

Done — liferaft PR InauguralSystems/liferaft#11 (merged, CI green against v0.29.0).

Acceptance, item by item:

• src/scheduler.eigs deleted; the sim runs on task_spawn/task_recv/task_send/task_sleep/task_kill/task_detach + task_sched_seed (seeded from liferaft's own PRNG, so the interleaving stays a pure function of the run's seed).
• Seeded chaos-replay suite byte-for-byte green: two-process determinism (plain + adversary), byte-identical EIGS_TRACE/EIGS_REPLAY (tape carries exactly the one argv N-record, trace: args builtin is an untaped nondeterminism source (closed-world hole for replay) #471), 50-seed invariant sweep, 40-seed M4 chaos sweep clean, fault injection caught with a verified=1 minimal repro.
• Gaps surfaced upstream instead of worked around — four of them, all fixed and shipped in v0.29.0: dict_remove exponential hash inflation (fix(dict): dict_remove no longer inflates the hash table exponentially #531), no task reaping / 255-spawn lifetime cap → task_detach + ready-queue poisoning (task layer: finished tasks are never reaped — 255-task LIFETIME cap per process (task_detach) #530→feat(tasks): task_detach — fire-and-forget tasks reap their handle slot (#530) #532), JIT compiling task code mid-task via OSR with an unchecked suspend in jit_helper_call (JIT: OSR compiles task code mid-task — jit_helper_call has no suspend check (state corruption after ~threshold iterations) #533→fix(jit): task code stays interpreted at every JIT entry point (#533) #534), and handle-id-dependent sleeper wake order breaking same-process seed reproducibility (task layer: same-instant sleeper wake order depends on handle IDs — interleaving isn't a pure function of the program+seed once slots recycle #535→fix(tasks): same-instant sleeper wake order is spawn-order, not id-order (#535) #536 — caught by this suite's own fault-verify gate).

The strongest kind of differential proof: the migration didn't just validate the layer, it broke it four ways first.