Compiler deep gate — nightly meta/native/compiled semantic parity

[juv4uk]

Goal

Extend the new fast/deep CI split so compiler optimizations can never quietly drift from Lisp semantics.

Fast lane

On ordinary PRs, run a bounded compiler parity corpus:

meta-eval == native == compiled

with exact values/errors for the admitted scope.

Deep lane

On the scheduled/manual deep workflow, run the full compiler oracle corpus plus expensive cases after the existing meta Advice Taker witness. The purpose is architectural falsification, not just build health.

Required invariants

• Deep compiler failures block claims of semantic equivalence even if fast CI is green.
• Optimized and unoptimized compiled paths are both compared where applicable.
• Exact-number identity and proof/provenance structures remain byte/structure comparable where contractual.
• No test is weakened merely to meet runtime targets; expensive witnesses may move to nightly instead.

Acceptance evidence

• Manual deep workflow can run the complete compiler parity gate on demand.
• A deliberately perturbed compiled observation is caught by both the appropriate fast/deep gate.
• Runtime and test scope are reported separately so speed gains cannot hide evidence loss.

Dependencies

• Compiler oracle corpus — freeze observable semantics before lowering #67
• Compiler first witness — execute compiled code with exact meta/native parity #69
• Compiler optimizer v0 — semantics-preserving transformations with witnesses #70
• PR ci: split deep meta witness into fast PR and nightly paths #65 fast/deep CI split