feat(tasks): #408 increment 4 — seeded scheduling strategy (task_sched_seed)

[InauguralPhysicist]

What

The pluggable scheduling-strategy hook for the deterministic-simulation-tester consumer — the last of the #408 primitives. Builds on the virtual-time increment (#482).

task_sched_seed of n switches the cooperative scheduler from FIFO round-robin to picking the next ready task from a seeded, platform-independent PRNG (splitmix64, pure integer arithmetic).

• Same seed ⇒ same interleaving on every run and under EIGS_REPLAY, recording zero tape N records (the pick is a pure function of the seed + program order — no wall clock, no nondeterminism).
• A different seed ⇒ a different ordering. This is the lever a DST (liferaft) uses to search the interleaving space while keeping every failure it finds byte-for-byte reproducible.

Why

#408's differential-proof step is migrating liferaft onto the upstream task layer. A DST's whole job is exploring interleavings; without a seam to vary scheduling from a seed it would keep hand-rolling one. This is that seam.

Safety / scope

• The FIFO fast path is untouched. With no seed installed, sched_ready_pop is the same O(1) head pop, so every existing program behaves identically. The seeded pick (O(ready-count), tiny, only in DST mode) removes the chosen task and compacts the queue order-preservingly.
• task_sched_seed ensures the scheduler exists so the seed sticks even when set before the first task_spawn.
• Follows eigenscript-extend-vm: one deterministic builtin (no TRACE_NONDET wrapping), two POD TaskScheduler fields, no opcode/AST/CallFrame/JIT-helper sites.

Gates

• Release suite 2620/2620; ASan (detect_leaks=1) 2620/2620, leak tally 0, no sanitizer errors.
• New determinism gate: the seeded schedule is identical across two fresh processes, replays byte-identically, records zero tape N records, and differs from the FIFO order (proving the seed actually reorders).
• docs/SPEC.md "Seeded scheduling" section (byte-for-byte doc-example test), docs/BUILTINS.md, examples/task_seeded_schedule.eigs, and 5 new assertions in tests/test_tasks.eigs (seed 42 and seed 7 pinned byte-exact).

Refs #408.

🤖 Generated with Claude Code