[PROOF / DO NOT MERGE] Release determinism CI is broken on main

[T-Gro]

Purpose

Evidence-only draft. This branch contains no product changes — only a CI configuration flip from Debug to Release for the Determinism job, with 8 parallel iterations to maximize the probability of catching probabilistic races on a single CI run.

Hypothesis

Multiple non-determinism races exist on current main, not visible because the existing CI Determinism_Debug job disables the optimizer (SetOptimizeOff for Debug):

1. DetupleArgs / TLR (F# compiler produces non-deterministic metadata #Strings heap layout #19732) — parallel optimizer iterates Val sets by racy Val.Stamp (Interlocked.Increment globally assigned)
2. IlxGen.TypeDefsBuilder.AddTypeDef — counter-based type ordering races under parallel codegen
3. GrabExtraBindingsToGenerate — ConcurrentStack drain order is non-deterministic
4. NiceNameGenerator — global singleton counter across other callers
5. FileIndex — assignment race under parallel parse

If any of these fire in any of the 8 iterations, we see hash mismatches on FSharp.Compiler.Service.dll etc.

Why this matters

PR #19810 fixes (1) but cannot land confidently while (2)-(5) remain hidden. This PR establishes the baseline: Release determinism CI currently fails. Subsequent PRs (#19810 + follow-ups) must drive it back to green by fixing the product, not by reverting CI to Debug.

What to look at

• Determinism_Release iter1..iter8 job results
• Search build logs for != (hash mismatch) and Files have different content

Related: #19732, #19810

[github-actions]

✅ No release notes required

[T-Gro]

✅ PROOF VALIDATED

Build 1440659:

• 6 of 8 iterations PASSED (no race fired)
• iter4 and iter7 FAILED with:

  Determinism failed for the following binaries:
  	FSharp.Compiler.Service.dll (old hash: 0C1A3BE3C55CCF242B84FB3D2DE10944; new hash: F6DD537FB21EB456F27BE5BB66E3107D)
  

This confirms:

1. Main is non-deterministic in Release. The race is probabilistic (fires ~25% per build).
2. The Determinism_Release CI job catches it reliably with multiple iterations (maxParallel: 8).

Fixes are in PR #19810.