Token-exact parity gate — local vs delegated, zero-mismatch

[lineape]

What to build

The maintainer's actual ask: a token-exact worker-vs-local gate. A Python test runs glm_tiny in teacher-forcing prefill twice — local vs cluster-delegated — and asserts identical token signatures AND zero [ORACLE] mismatch lines. Determinism is pinned so the comparison is meaningful: same ARCH build on both sides, CPU-only backends, and a fixed numeric-env block (IDOT=0 I4S=1 COLI_NO_FUSED_PAIR=1 SPEC_PIN=0 XEXP=0 I4_ACC512=0 I3_AVX512=0 DRAFT=0 plus COLI_CUDA=0 COLI_METAL=0 COLI_VULKAN=0). Mirror JustVugg#550's test_dense_sharding.py, but be stronger: JustVugg#550 only asserts baseline≡delegated equality of mismatch lines; we assert both equality AND zero mismatch. Wire it into CI so the worker path is exercised, not dormant.

Acceptance criteria

• Runs local and delegated; asserts zero [ORACLE] mismatch lines AND identical token signatures.
• Fails on the pre-fix worker (fmt=6 bug) and passes post-fix — demonstrates the gate catches the bug.
• Determinism pinned on both sides (matching ARCH, CPU-only, numeric-env block above).
• Skips cleanly when the gitignored glm_tiny fixture is absent (no CI hard-fail).
• Compares routed experts only (shared expert stays coordinator-side).
• Runs at real dimensions, not toy; wired into CI so it's not dormant.

Environment: codegraph-indexed — run codegraph explore first; refs in /home/eli/dev/_reference-docs-and-repos/colibri.