ci: cut the nightly deep verification to about an hour

[P4suta]

Summary

Closes #40.

The nightly deep verification spent about twelve hours of runner time each night, ten of them fuzzing.
It now spends about an hour.

• Each fuzz target runs for fifteen minutes instead of five hours.
• The stress test runs 20000 spawns on x64, where it ran 100000, and 5000 on i686, where it ran 20000.
• The random cases run at ten times their size instead of fifty.
• The job timeouts follow the new budget, and ADR 0017 and CONTRIBUTING state it.

The PR checks don't change.

The PR also removes eight of the eleven registered lint exceptions.

• The spawn tests no longer silence as, expect, and unwrap_in_result across the whole file, and their crate doc keeps missing_docs satisfied off Windows.
  Handles pass between Win32 calls as pointers, a probe's handle travels as {:p} text, and a malformed probe or an open gate returns an error.
• A value becomes a handle through a byte offset from null, and the DWORD conversions move to domain::conversion_rules and convert through bytes.
  Three of the four as exceptions in sys and its truncation exception go, and Kani proves the width conversions for every input.
• Reading a handle's value, the sys module's unsafe, and a probe's exit keep their exceptions, because Rust 1.75 has neither addr nor #[expect].

Validation

• actionlint passed, and zizmor reports nothing for deep-verification.yml.
• On Windows 11, 1300 stress spawns took 7.7 seconds on x64 and 10.5 seconds on i686, so the nightly stress runs take minutes.
• On macOS, each fuzz target ran 2 million or more cases in one minute, so fifteen minutes still exercise them well.
• On Windows 11, cargo xtask ci passed through the Linux-empty step.
  There cargo deny couldn't refresh its advisory database, and on macOS it passed with the cached one.
• cargo xtask proofs kani verified every proof, including the three new ones, and every negative control failed on its marker.
• cargo miri test passed the conversion unit tests.
• Clippy passed for x64, i686, and ARM64 Windows, and Rust 1.75 checked x64 and i686.

Checklist

• The relevant local tests ran.
• Deterministic tests cover each behavior change.
• The docs state each new or changed ownership, safety, or cleanup contract.
• The public API doesn't change.
• The PR holds no secrets or public vulnerability details.
• The PR stays focused and updates the relevant documentation.

@coderabbitai ignore