feat(tasks): #408 increment 3 — virtual time (task_sleep/task_now)

[InauguralPhysicist]

What

Adds a logical virtual clock to the cooperative task scheduler — increment 3 of #408.

• task_sleep of ticks suspends the current task until the clock reaches now + ticks.
• task_now of null reads the clock.

The clock starts at 0 and only ever jumps forward, to the earliest sleeper, when nothing else is runnable (discrete-event simulation). A program that sleeps therefore runs in zero real time and stays deterministic by construction — no wall clock, no tape records — exactly like the rest of the task layer.

Why

This is the election-timeout / heartbeat primitive the deterministic Raft simulator (liferaft) needs to migrate onto the upstream task layer (the #408 differential-proof step). Timeouts and periodic intervals are inherently time-based; without a virtual clock a DST would have to hand-roll one (which liferaft currently does).

Behavior details

• Sleepers waiting on the clock are no longer mistaken for a deadlock — the trampoline advances the clock (sched_wake_sleepers) before deciding anything is stuck. A genuine deadlock (no sleepers) still errors loudly.
• Ties at the same wake time break by ascending task id (main = 0 first), keeping the interleaving deterministic.
• A negative sleep clamps to 0; task_sleep of 0 is a same-tick yield; with no task ever spawned it is a no-op, like task_yield.
• The initial vm_execute suspension guard is aligned with the trampoline re-enqueue guard, so a main that sleeps (or recvs) first is not wrongly re-readied.

Follows the eigenscript-extend-vm checklist: new Task fields (sleeping/wake_at, POD), two deterministic builtins registered, no opcode/AST/CallFrame/JIT-helper sites. task_sleep/task_now are not nondeterministic — they read a logical counter, so no TRACE_NONDET wrapping.

Gates

• Release suite 2617/2617.
• ASan (detect_leaks=1) 2617/2617, leak tally 0.
• New gate: virtual time is deterministic across two fresh processes, replays byte-identically under EIGS_TRACE/EIGS_REPLAY, and records zero tape `N` records (a positive assertion that the clock leaks no nondeterminism).
• docs/SPEC.md "Virtual time" section (byte-for-byte doc-example test), docs/BUILTINS.md, examples/task_virtual_time.eigs (a watchdog/timeout race), and 12 new assertions in tests/test_tasks.eigs.

Noted (pre-existing, not in this diff)

While adversarially probing, I confirmed that a program which deadlocks (or otherwise dies of an uncaught error with live tasks) leaks the TaskScheduler + suspended-task slices at its fatal exit — a process-exit leak, not steady-state, on a path no test reaches (so the leak tally is unaffected). This is pre-existing: the deadlock path allocates nothing new under this change. Filing separately rather than scope-creeping this increment.

Refs #408.

🤖 Generated with Claude Code