feat(engine): Phase 0 RZF engine with gating oracle suite + spec v2.2

[justin06lee]

Build the zeroforce RZF engine first (per spec §15), since a subtly wrong EPT is silent and catastrophic. Math-correctness gates CI before anything is built on top.

Engine (packages/zeroforce/):

• engine_py: pure-Python reference Markov-chain DP (normalized weights so denominator = 1; self-loop fixed via next_mask != mask).
• engine_numba: nopython JIT mirror of the reference DP.
• reachability: R(v) closure + induced subgraph G[R(v)]; EPT scoped to R(v).
• ranking: impact_score = n_reachable / EPT (excludes self), sink separation.
• simulator: independent Monte Carlo RZF (second definition for differential).
• api.compute(graph): asserts the §7.2 row-stochastic invariant on the FULL graph, runs per-seed DP on induced subgraphs, ranks.

Gating tests (41 passed, pristine under -W error):

• hand-computed 3-node oracle (EPT derived from first principles, trusts paper 0)
• Monte Carlo differential (sim vs DP, n<=6, <2e-2)
• engine parity (py vs numba, n<=10, <1e-10)
• properties (directed-path depth, seed monotonicity) + invariant guards
• §4.6 closed forms kept xfail(strict=False) until paper-body verified

Two clarifications over the spec pseudocode, documented in the engine README:

• the row-stochastic invariant is asserted on the full graph, not subgraphs (unreachable suppliers leave subgraph rows summing < 1; this is correct)
• the recurrence uses the unconditional P(next) divided once by (1 - p_stay) (the prose's "P(next | flip)" plus a second division would double-divide)

Spec bumped to v2.2: rename Foresight -> zeroforce; widen numpy pin to

=1.26,<3.0 with an import numba CI smoke test; explicit async polling
contract + 5-min job timeout; Gemini Pass B retry/fallback; oracle strategy promoted to a standalone callout.